Measure E4B on phrasing, the half nobody had scored (V-668)

The 2026-08-09 model swap was measured on routing the same day and E4B lost
four destination cases. Phrasing was not measured, and phrasing is the half the
owner hears.

E4B scores nudges 15/15 and the talk fixture 29/36 at p50 516ms, against the
resident model's 25/36 at p50 2.97s on the same 36 cases. lang, feminine and
address are all 36/36, where the resident model loses three on address. Every
failure is ontopic and none is a parse error.

29/36 is one case off the ceiling. The temperature sweep of 2026-08-05 found
two reply cases that fail at every temperature and named a defect in the reply
path, capping the fixture at 30/36. Both are in E4B's failure list. So the swap
costs nothing on phrasing.

One defect no check catches: in chat E4B writes "Я записала несколько идей!"
when nothing was stored. A claim to have saved something is a claim about state.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN

.gitignore

@@ -70,3 +70,5 @@ coverage.out
 
 # root .env — MAVEN_AMBIENT_TOKEN and friends, same class as deploy/telegram.env
 .env
+# silero-vad, downloaded (see AGENTS.md)
+/models/vad/

AGENTS.md

@@ -95,6 +95,22 @@ model: the code puts `query: ` in front of a question and `passage: ` in front
 of a stored note, which is how e5 was trained. The quantized file is the one
 that is downloaded, deployed and measured.
 
+## Voice activity model for mavwaked
+
+`mavwaked` decides an utterance has started with silero-vad when `-vad-model`
+points at it, and with an energy threshold when it does not. The model is 2.3MB
+and is not committed:
+
+```sh
+mkdir -p models/vad
+curl -sL -o models/vad/silero_vad.onnx \
+  https://github.com/snakers4/silero-vad/raw/master/src/silero_vad/data/silero_vad.onnx
+```
+
+It needs the same `libonnxruntime.so` the embedder needs, passed as `-onnx-lib`
+or read from `MAVEN_ONNX_LIB`. The measurement is
+`docs/evals/2026-08-09-silero-vad.md`, and the tests skip without the file.
+
 **Also need ONNX Runtime** (`libonnxruntime.so`):
 
 ```sh

CLAUDE.md

@@ -47,6 +47,31 @@ free — `worldGap` in `cmd/mavend/worldmodel.go`, which the owner hears instead
 answer. A box with no `workstation` block behaves exactly as it did before the seam: naming
 a gap requires a gap. The offload table in `docs/offload.md` says which caller is which.
 
+**Speech-to-text moved on 2026-08-09** (V-486). `sttSeam` in `cmd/mavend/voicewire.go`
+builds an `stt.Pair` beside `modelSeam`, preferring CrisperWhisper 2.0 turbo on workpc
+with mavsttd as the floor. It takes only the silent half of the rule. A worse
+transcript is still a turn, so `stt.Pair` has no `TranscribeRemote`. The fallback is
+never spoken. CW2 turbo scores **10.4% WER in Russian against 27.5%** for the `ggml-small.bin`
+mavsttd loads, over 200 Golos clips
+(`docs/evals/2026-08-09-crisperwhisper2-russian-wer.md`). It runs in Intended mode, not
+Verbatim, though that corpus cannot separate the two.
+**whisper.cpp cannot load CW2 at all.** It reads its language count off the vocabulary
+size, and CW2's 51897 tokens shift seven special token ids. So it is not a second
+endpoint on mavgpud. It is its own transformers service on port 8081
+(`deploy/cw2/serve.py`), which Maven reaches directly. `stt.HTTPTranscriber`
+posts raw PCM to it with a bearer token, because audio is the most sensitive thing that
+crosses this seam. The switch is `workstation.stt` in
+`deploy/mavend.json`, and deleting the block sends every utterance to mavsttd.
+**mavgpud runs that service as a second child.** That is not an optimisation. CW2 is a
+ROCm process on the same card, so it registers on the KFD like any contender. Under its own
+systemd unit it made mavgpud evict llama-server every few seconds. That took the
+gemma-4-12b arm down for eight minutes on 2026-08-09 before anyone noticed. The card needs
+one owner. Any GPU service added beside this daemon has the same defect, so add it to
+`cmd/mavgpud` and not to systemd. CW2 is on the yield clock and not the idle one. At 1.6GB
+it denies the card to nobody, and unloading it would only send the next voice turn to the
+homesrv floor.
+Text-to-speech has not moved and piper on homesrv is still the only synthesizer.
+
 ## Build & test
 
 CGO daemons (`mavend`, `mavsttd`, `mavttsd`, `mavenclient`) need the vendored toolchain
@@ -58,16 +83,27 @@ make build-web      # single daemon (pure-Go ones: web/waked/poll/caldav build w
 make test           # go test -race across ./internal/... ./cmd/... with CGO env set
 ```
 
-Run a single test (must carry the CGO env for packages that touch STT/TTS/voice):
+Run one package or one test with `make t`. **Do not hand-write the CGO preamble.**
+Past sessions pasted it about 390 times. That is where the shell-quoting failures
+came from. This box runs zsh, so an unquoted `-run Test*` or `--include=*.go`
+dies on "no matches found" before `go` is ever reached.
 
 ```sh
-CGO_CFLAGS="-I$(pwd)/deps/include -I$(pwd)/deps/whisper.cpp/ggml/include" \
-CGO_LDFLAGS="-L$(pwd)/deps/lib -Wl,-rpath,$(pwd)/deps/lib" \
-LD_LIBRARY_PATH="$(pwd)/deps/lib" \
-deps/go/go/bin/go test -run TestName ./internal/router/
+make t PKG=./internal/router/
+make t PKG=./cmd/mavend/ RUN=TestSimulator
+make t PKG=./internal/router/eval/ RUN='TestONNX' V=1   # V=1 for -v, RACE=0 to drop -race
 ```
 
-Pure-Go packages (`router`, `memory`, `mavweb`, …) run under a plain `go test ./pkg/`.
+`t` carries `-race`, so a green `make t` cannot turn red under `make test`. It carries
+`-count=1`, so a cached PASS from before your edit is never mistaken for a result.
+
+It also sets `MAVEN_ONNX_LIB`, which the hand-written recipe did not. The four
+`TestONNX*` measurements self-skip when that variable is unset. The run still prints
+`ok`. So every targeted eval done the old way reported the hash ratchet while reading
+as a real embedder score.
+
+Pure-Go packages (`router`, `memory`, `mavweb`, …) also run under a plain `go test ./pkg/`,
+but `make t` works everywhere and is one thing to remember.
 
 ## The daemons (`cmd/`)
 
@@ -101,9 +137,15 @@ protocol; the config in `deploy/mavend.json` (with `${VAR}` env expansion from g
 Count against compose, not against the table. Four of the nine daemons are absent, and each
 absence has a different reason.
 
-`mavmaild` is commented out in compose, with the reason written beside it: it needs a mail
-account and this box has none. `mavcaldav` appears nowhere at all, and unlike the other
-three that is an oversight rather than a decision (V-644).
+`mavmaild` and `mavcaldav` are commented out in compose, each with the reason written
+beside it: the first needs a mail account, the second a CalDAV account, and this box has
+neither. `mavcaldav` used to appear nowhere at all, which was an oversight; it became a
+recorded decision on 07-08-2026 (V-644). Two things ride on that absence and the block
+names them. Agenda questions route to `IntentQuery` at stage 0 (V-498) and the `calendar`
+query source then reads a table nobody writes. And `loop.State.CalendarBusy` is fed by the
+same facts, so the gate's "do not nag mid-meeting" is permanently false. Its password is
+read from a file (`-pass-file`, and `-render-pass-file` for the render collection), never
+taken as a flag value, which is the rule `mavpoll` and `mavmaild` follow too.
 
 **`mavwaked` and `mavenclient` are absent by decision, not oversight** (Vikunja #463,
 `docs/plans/17-where-the-voice-loop-runs.md`).
@@ -213,11 +255,31 @@ re-run it, start a **second** llama-server on a fixed host port — the resident
 `--port 0` inside the container and no host process can reach it.
 
 **The numbers above are the homesrv floor, not the ceiling.** With the workstation up, routing
-completes through `llm.Pair` against gemma-4-12b and scores **84.4% full / 93.5% intent-only at
-p50 329ms** — better than the resident model and about 2.5× faster (`docs/evals/2026-08-02-workstation-gemma4-12b.md`,
-Vikunja #485). The workstation is never assumed up, so both sets of numbers are live. Judge a
+completes through `llm.Pair` against the model mavgpud holds, which is better than the resident
+model and about 2.5× faster. gemma-4-12b scored **84.4% full / 93.5% intent-only at p50 329ms**
+(`docs/evals/2026-08-02-workstation-gemma4-12b.md`, Vikunja #485). The workstation is never
+assumed up, so both sets of numbers are live. Judge a
 routing change against the classifier and the resident model, since those are what always answer.
 
+**The workstation runs gemma-4-E4B since 2026-08-09** (owner's call), and it is a
+step down measured the same day (`docs/evals/2026-08-09-e4b-vs-12b-routing.md`).
+Against a same-session 12B control it scores **83.3% full / 89.6% intent-only,
+destination 19/33 against 23/33, at p50 294ms against 344ms**. So it costs four
+destination cases and buys 50ms. Read destination as the finding: it names nothing
+where the 12B names `recall` or `calendar`, which is safe but walks the whole chain.
+It also has no MTP and cannot be given any here. The only `gemma4-assistant`
+draft on disk is trained against the 12B's hidden states.
+**Phrasing was the unmeasured half and it is measured now**
+(`docs/evals/2026-08-09-e4b-phrasing.md`). E4B scores nudges 15/15 and the
+36-case talk fixture **29/36 at p50 516ms**, against the resident model's 25/36
+at p50 2.97s. Persona is clean: `lang`, `feminine` and `address` are all 36/36,
+where the resident model loses three on `address`. Every failure is `ontopic`
+and none is a parse error. The 2026-08-05 temperature sweep put this fixture's
+ceiling at 30/36, because two reply cases fail at every temperature (V-537), and
+both are in E4B's failure list. So the swap costs nothing here. One defect no
+check catches: in chat E4B claims "Я записала несколько идей!" when nothing was
+stored, which is a wrong claim about state.
+
 **The intended third engine is not a generative model** (owner's call, 05-08-2026, V-546,
 `docs/plans/18-routing-heads-on-e5-small.md`). Routing has a bounded output space, so it is
 classification, and the 118M multilingual-e5-small is already resident. Three heads on one
@@ -229,6 +291,99 @@ was a hardcode. **Fine-tune a copy of the weights.** The resident embedder backs
 recall. Training it in place couples routing accuracy to recall@1, with nothing in the
 suite to name the trade.
 
+**Two of those heads are trained as of 08-08-2026, and they are not the three
+above** (V-661, `docs/evals/2026-08-08-routing-heads-two-head.md`). Intent and
+destination share one masked mean pool. Destination scores a mean **80.8%** over
+three seeds, best **29/33 (87.9%)**. The classifier cascade scores 12/33 and the
+cascade with gemma-4-12b scores 24/33, so a 118M encoder beats the 12B teacher it
+was distilled from. Read the best run as one seed and not a headline, because one
+case is 3 points on a fixture this small.
+Recall is 15/15 and world is 5/5. Intent is 93.6% mean over three seeds. That is
+**not** comparable to the 76.0% and 84.4% those two arms scored: a softmax has no
+clarify class, so the head's fixture is the 88 cases carrying an intent.
+
+**A fourth head asks instead of guessing, same day** (V-661,
+`docs/evals/2026-08-08-clarify-head-four-head.md`). Clarify is not a value
+of intent, so a softmax cannot emit it. It is a second question over the
+same pooled vector: can Maven act on this at all. That is why the head's
+fixture was 88 cases and not 96. Over three seeds it catches **7.0 of the 8
+`want_clarify` cases and produces 2.3 false clarifies of 88**. The cascade
+today misses 1 and produces 2, so this is parity with no rules in front of
+it. Accuracy is the wrong number here and a head that never asks scores
+91.7%. Confidence is the other half. Max softmax over the intent head reads
+**0.851 where it is right against 0.604 where it is wrong**, ranking right
+above wrong in 83.4% of pairs. `Confidence: 1.0` was a hardcode, and this
+replaces it with a signal. The two are not the same signal: one says which
+intent is unclear, the other says the utterance carries too little to act
+on. **The fourth head is not free the way the third was.** Intent,
+destination and slot F1 each move down one to four points, inside the seed
+spread. `поужинал` is a false clarify on every seed, which is the same
+defect `thinSingleToken` was narrowed for on 2026-08-01.
+
+The corpus for it is generated, because every existing row is answerable by
+construction. **The router-prompt agreement filter cannot work here**, since
+`routeGrammar` has no clarify value and a generated line always agrees with
+itself. A gemma judge replaces it. The first judge called 24 of 40
+answerable rows underspecified, because it judged against a generic
+assistant rather than against Maven's contract.
+
+**Mood is cut, not deferred.** The enum describes her own reply state, not the
+speaker's emotion, and no dataset maps onto it.
+
+**A third head landed the same day** (`docs/evals/2026-08-08-slot-head-three-head.md`).
+BIO slot tags had no Maven-domain corpus, which was true of found corpora and
+false of made ones. `label_slots.py` distils spans out of gemma-4-12b under a
+GBNF closed over Maven's own five slots. A span survives only when it is a
+literal substring of the utterance, so the agreement filter costs no second
+call. 2178 spans over 1702 rows, 37 dropped, nothing unparsed. Three heads score
+intent **92.8%**, destination **82.8%** and slot span F1 **72.4%** over three
+seeds. The slot head is free: both other numbers move less than their own seed
+spread. Epoch selection reads the intent dev slice alone. Slot F1 is still
+climbing when it stops, which costs about 4 points.
+
+The MASSIVE warm-start of step 2 is worth nothing here. Stock e5-small ties it on
+intent and leads by a third of a case on destination. Nothing argues for keeping
+that step.
+
+The floor was a corpus defect and it is fixed. The first 120 floor rows carried
+one sentence shape, so the head named a destination where the fixture says walk
+the chain. Rotating six shapes took the floor 3/7 to 6/7 and destination 75.8% to
+80.8%. What is left is calendar at 3/6 on every seed, which training cannot move:
+the possessive agenda rules claim those cases at stage 0 and name nothing, so no
+label reaches the head. That is the same trade V-660 flagged and it wants the
+owner's call.
+
+**The heads run in Go and route every turn, since 08-08-2026** (V-664,
+`docs/evals/2026-08-08-routing-heads-in-go.md`). This section used to say
+nothing of it ran. `RouterHeads` in `internal/router/heads.go` loads
+`router_heads.onnx` and reads intent, destination and clarify off one forward
+pass. It is stage 0b: after the grammars, **before** the resident model, and the
+classifier is still behind both. Through the cascade it scores intent **96.9%**
+and destination **75.8%** at p50 27.9ms. That beats the gemma-4-12b cascade,
+84.4% and 72.7%, at a twelfth of its 329ms. The workstation stays the better
+phraser and is no longer the better router.
+
+Three rules around it. The **clarify head decides first**, before the intent
+threshold. It answers a different question. A thin utterance scores low
+intent by construction, so gating it cost 6 of 8 ambiguous cases. The
+**destination head is read on `IntentQuery` only**, since no other intent
+reaches `queryWalk`. And `headsThreshold` is 0.6, the measured knee: every value
+to 0.85 drops right answers and keeps the same two wrong ones.
+
+`voice.embedder.heads_path` is the whole switch. Empty, missing or unloadable
+means the heads are nil and the cascade is byte-for-byte what shipped before
+them. **It must never be pointed at `model_path`.** The resident e5-small must
+not be replaced by the fine-tuned copy. Recall depends on that file scoring
+what it scored.
+
+**The hand-written tokenizer read every long word backwards** until this task
+(`encodeWord`, `onnxembedder.go`). It cost recall@1 7.4 points and recall@3 11.1.
+Nothing caught it because seeds and queries were mangled the same way, so cosine
+survived. The heads found it. They are trained through transformers and read
+through this. The embedder id now carries a tokenizer revision
+(`@384/tok2`), so fixing the tokenizer triggers `ReembedAll` the way swapping the
+model file does. Bump `tokenizerRev` on any change to what it emits.
+
 `Confidence: 1.0` used to be hardcoded in `llmrouter.go`, so the LLM
 path could never ask for clarification (6/6 refusal cases missed on the fixture) — Vikunja
 #359. Fixed 31-07-2026 with structural signal (single-token utterance, keyless fact, act with
@@ -340,6 +495,125 @@ Adding a rung to the ladder
 in `runTurn` means adding its name to `preRouteLadder` in
 `cmd/mavend/decisiontrace.go`, or that rung is silently missing from the record.
 
+**A route now says where the answer lives, not only that the turn is a question**
+(V-655, 07-08-2026). `query` was a shrug. The cascade sorted an utterance into one of
+seven intents, with stage 0, the resident model and the classifier behind it. Then
+`IntentQuery` handed the turn to `querySources` in the daemon. That is twenty-two branches
+deciding by seed similarity in a fixed order. It has no fixture and no accuracy
+number, no model arm and no floor. `Decision.Source` (`internal/router/source.go`) is
+the second half of the route. Twelve destinations, not twenty-two. The three recall
+passes plus `fact-by-key` are one destination from outside. So are search, Kiwix and
+the URL reader.
+
+**`SourceUnknown` is a real value and it is the floor.** Nothing named a destination,
+so the daemon walks the whole chain. That is byte-for-byte what shipped before the
+field existed. The classifier arm names nothing, so a box whose model is down routes
+queries exactly as it did.
+
+`queryWalk` in `cmd/mavend/actions_query.go` takes sources **out** and moves none.
+That is the safety argument and it is not negotiable. The table's order is
+load-bearing. Every comment on it argues a reason between two sources, and above all
+it carries "the owner's data first, then the world". Naming `SourceWorld` does not
+send the turn outside on its own. His notes and his facts still run first, because
+they look rather than guess.
+
+**The personal boundary is the one exception and it is deliberate.** It guesses,
+so naming `SourceWorld` drops it. That is what stops it answering "кто такой
+Линус Торвальдс?" with "не нашла у тебя такой записи", which it did on
+2026-08-07. `TestNamingRecallKeepsTheBoundary` pins the other half: naming
+`SourceRecall` keeps the boundary in front of the world.
+
+**Only a stage 0 grammar may drop it** (owner's call, 09-08-2026, V-666). The
+question of who is allowed to was open until then. Three deciders name a
+destination and two of them infer it: the routing heads and the resident model.
+An inferred `SourceWorld` on a question about him would reach SearXNG, and that
+widens what is asked rather than costing a local answer. So `Decision.SourceAnchored`
+carries the provenance. It is a field and not `Stage == 0`. Stage 0 also means
+confidence 1.0 and an anchored claim band, and one of those could stop implying
+the others. `queryWalk` reads it for the source marked `boundary: true` and for
+no other. So every other guesser still comes off the turn, whoever named the
+destination. `TestOnlyAGrammarMayDropTheBoundary` pins both directions.
+`definitionQueryPattern` claims "кто такой X", so the 2026-08-07 case is still
+anchored and still answered.
+
+What comes out is only the sources that **guess**. Those decide a turn is theirs by
+cosine against frozen seeds, then answer whatever they claimed. They hold no table
+that could come back empty. Weather is the pure case and has no local data at
+all. It was measured on the box on 2026-08-07
+(`docs/evals/2026-08-07-week-of-usage.md` section 4). It answered both "что такое
+TCP?" and "сколько будет 17 на 23?" with "для какого города?". The feed answered "какой у меня любимый язык?" with kernel headlines.
+The personal boundary answered "кто такой Линус Торвальдс?" with "не нашла у тебя
+такой записи". A source that guesses is marked `guesses: true` in the table. One that
+looks is not, and it is always asked.
+
+Stage 0 fills the destination where a rule already knows it. `WorldQueryGrammars()`
+(`internal/router/worldquery.go`) claims "что такое X" and "сколько будет 17 на 23".
+It is wired after the agenda rules and **before** the feed and list rules.
+"что такое лента" is a definition question, and the feed rule would take it on the
+noun alone.
+`calendar-query` and `event-time-query` name the calendar. The possessive agenda rules
+deliberately do not. "что у меня в списке покупок" matches `agenda-query`, and naming
+the calendar there would take the list source off the turn.
+
+Fixture unchanged at **69/91 classifier+ONNX**, measured both sides. That is the
+expected result, because it scores intent and no case here changes intent.
+
+**The destination has its own fixture and its own number as of 08-08-2026**
+(V-659, `docs/evals/2026-08-08-destination-fixture.md`). This section used to say
+it had neither. `want_source` on `eval.Case` is a pointer, because the destination
+has three states and a bare string has two. Absent is every intent but query,
+which never reaches `queryWalk`. Present and empty is the `SourceUnknown`
+contract: name nothing and walk the chain. Present and named is a destination the
+route must produce. Thirty-three of ninety-six cases carry one.
+
+A destination miss does **not** fail the case. It lands in `Outcome.SourceReason`
+and never in `Reasons`, so `Accuracy` and `IntentAccuracy` mean what they meant
+and `SourceAccuracy` is a second number over the labelled cases only. Intent and
+destination are two decisions, and one number hides which one moved. A route that
+lost its intent scores no destination hit, or a clarify would satisfy an empty
+label for free.
+
+Measured classifier+ONNX: intent **73/96 (76.0%)**, destination **12/33 (36.4%)**.
+The split is the finding. World is 5/5, because a stage 0 rule names it. The
+`SourceUnknown` floor is 5/7. Calendar is 2/6, because the possessive agenda
+rules deliberately do not name it. And **recall is 0/15, because nothing
+anywhere names it**. Those turns are still answered, since the chain walks
+recall early. Recall is the number the fourth head has to move.
+
+Seven cases assert the floor and five of them are homelab operations. They
+cluster because `SourceRecall`, `SourceNetwork` and `SourceAttention` overlap on
+every question about the box. The other two are `ru-query-005` and
+`ru-query-014`. No query source reads the reminder store, and a deadline could
+sit in tasks, the calendar or Praxis. `mavpoll` writes its netdata and uptime-kuma
+observations into the fact store recall reads. That is a finding about the enum,
+not a gap in the labelling. The owner confirmed all seven floor labels on
+08-08-2026, so they are a decision rather than an agent's guess.
+
+`baselineGrammars` in `eval_test.go` mirrors `buildRouter` and had drifted:
+`WorldQueryGrammars` was wired into the daemon by V-655 and not into the mirror,
+so the fixture scored a grammar set nobody runs. Fixed by V-659, worth 3 points of
+destination and nothing else. Check that function when adding a grammar.
+
+**The model arm landed the same day** (V-660,
+`docs/evals/2026-08-08-destination-model-arm.md`). `routeGrammar` carries a
+`source` rule closed over `router.Sources` plus the empty floor, so the model
+cannot emit a destination that does not exist. The prompt lists the twelve in
+Russian and says `""` is a normal answer to give often. `LLMRouter.Route` reads it
+back through `ValidSource` and on `IntentQuery` alone. Against gemma-4-12b on the
+workstation the cascade scores destination **24/33 (72.7%)** with intent unmoved
+at 84.4%, and **recall goes 0/15 to 14/15**. The resident Qwen3-1.7B is
+unmeasured, because it binds `--port 0` inside the container.
+
+**Stage 0 now costs four destination points.** It did not before. The four cases
+the cascade loses and the model alone wins are all calendar. The possessive
+agenda rules claim them first and name nothing on purpose. That caution was free
+while nothing downstream could name anything either. It is not free now, and the
+fix is the owner's call rather than a quiet edit.
+
+The last arm is V-546. Intent, mood and BIO slot tags were already three heads on
+one forward pass of the resident e5-small. Destination is a fourth head on the
+same pass, and 72.7% from a 12B teacher is the label source for training it.
+
 ## LLM output contract
 
 All phrasing paths emit `{"response":"...","mood":"..."}`, with fallback to plain text when
@@ -411,11 +685,28 @@ world questions, so she needs to read external sources. What replaces it:
   `wikipedia_ru_all_maxi_2026-02` verbatim** through `kiwix.book_ru`. The RU→EN rewriter
   is the workaround for an English book and is skipped there. Kiwix catalog names come
   from the filename, not the `<name>` field.
+  **That verbatim path sent the whole sentence to a keyword engine until 09-08-2026**
+  (V-668, `docs/evals/2026-08-09-kiwix-topic-retrieval.md`). Kiwix ranks by keyword
+  overlap, so the question words outrank the one word naming the article. "что такое TCP"
+  returned "Перехват TCP-соединения". "кто написал Войну и мир" returned an episode of
+  Doctor Who. `kiwix.Topic` drops the narrative request, the interrogative and a verb
+  behind one. `kiwix.TitlePath` tries the exact article first, since a ZIM is addressable
+  by title and a wrong title is a 404. Five of eight questions reach the right article
+  where they did not, one was already right, and nothing regressed. The title needs its
+  leading capital, so `TitleCandidates` tries the spoken form and then the capitalized
+  one. **"столица Франции" is answered by a title redirect to Париж**, which is the case
+  the 2026-08-05 measurement named as unreachable by any lexical signal. Both apply on the
+  verbatim path alone. The rewriter already reduces a question, and reducing twice takes
+  the topic off its input.
   `Response.Empty()` is the whole gate and there is no quality threshold in front of it:
   the three signals one could read were measured on 2026-08-05 and none of them separate a
   real question from an invented one. Token overlap would cost "столица Франции" its
   answer, because the answer is Париж and that word is not in the question. See
-  `docs/evals/2026-08-05-search-quality-signals.md` (V-539). **Which query source claimed
+  `docs/evals/2026-08-05-search-quality-signals.md` (V-539). **The embedder is not a
+  fourth signal**, measured 2026-08-09 (V-668). Query-to-passage cosine scores 0.79 to
+  0.91 on answerable questions and 0.75 to 0.84 on unanswerable ones, and the sets
+  overlap. The wrong TCP article scored 0.8653, above five of six unanswerable rows. It
+  measures topic and not whether the passage answers, so no threshold splits them. **Which query source claimed
   a turn is readable on `/chat`** as a badge beside the reply, carried on
   `ipc.ChatReply.Source` and noted by `noteQuerySource` in `cmd/mavend/querysource.go`. It
   rides the context, so `handleText` keeps the one string signature the mic, telegram and
@@ -454,8 +745,11 @@ start of a session rather than one lookup per first use:
 ToolSearch("select:mcp__vikunja__list_tasks,mcp__vikunja__get_task_details,mcp__vikunja__create_task,mcp__vikunja__update_task")
 ```
 
-`update_task` carrying a `description` resets `done` to false, so closing a task with a
-write-up takes two calls: the description, then `done: true`.
+**Close a finished task with `done: true` and nothing else** (owner's call, 07-08-2026).
+Do not write a completion summary into the description on the way out. It is lost anyway,
+and the durable record is the commit messages and the merged PR. Note that `update_task`
+carrying a `description` resets `done` to false, which is why a write-up ever took two
+calls.
 
 ## Session workflow
 

Makefile

@@ -16,7 +16,7 @@ PIPER_BIN     := $(shell pwd)/deps/piper/piper
 PIPER_MODEL   := $(shell pwd)/models/tts/ru_RU-irina-medium.onnx
 PIPER_ESPEAK  := $(shell pwd)/deps/piper/espeak-ng-data
 
-.PHONY: simulate stt-fixtures test-stt-golden all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go deps-sentinel tidy eval-router eval-reach eval-recall eval-phrasing eval-models build-gpud
+.PHONY: t audit simulate stt-fixtures test-stt-golden all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go deps-sentinel tidy eval-router eval-reach eval-recall eval-phrasing eval-models build-gpud
 
 all: build
 
@@ -128,6 +128,35 @@ test: fmt-check vet
 	CGO_CFLAGS="$(CGO_CFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
 		$(GO) test -race -coverprofile=coverage.out ./internal/... ./cmd/...
 
+# t — run ONE package or ONE test with the toolchain env already wired. This is
+# the iteration target; `test` is the gate. Reach for it instead of pasting the
+# CGO_CFLAGS/CGO_LDFLAGS/LD_LIBRARY_PATH preamble by hand, which is how it was
+# done ~390 times across past sessions and is where the shell-quoting failures
+# came from -- the interactive shell here is zsh, and an unquoted `-run Test*`
+# or `--include=*.go` dies on "no matches found" before go ever starts.
+#
+#   make t                                     # whole tree (same scope as `test`)
+#   make t PKG=./internal/router/
+#   make t PKG=./cmd/mavend/ RUN=TestSimulator
+#   make t PKG=./internal/router/eval/ RUN='TestONNX' V=1
+#   make t PKG=./internal/store/ RACE=0        # drop -race when iterating hot
+#
+# -race is on by default so a green `make t` cannot turn red under `make test`.
+# -count=1 because a cached PASS from before your edit is worse than no answer.
+# MAVEN_ONNX_LIB is set for the same reason: the four TestONNX* measurements
+# self-skip when it is unset, so a targeted eval run would otherwise report the
+# deterministic hash ratchet and look like it scored the real embedder.
+PKG  ?= ./internal/... ./cmd/...
+RUN  ?=
+V    ?=
+RACE ?= 1
+
+t:
+	CGO_CFLAGS="$(CGO_CFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
+	MAVEN_ONNX_LIB="$(MAVEN_ONNX_LIB)" \
+		$(GO) test $(if $(V),-v,) $(if $(filter-out 0,$(RACE)),-race,) -count=1 \
+		$(if $(RUN),-run '$(RUN)',) $(PKG)
+
 # eval-router — score the held-out RU routing fixture (internal/router/eval).
 # Verbose so the report tables land in the terminal. MAVEN_ONNX_LIB points the
 # prod-representative baseline at the vendored runtime; override it or set it
@@ -189,6 +218,16 @@ eval-models:
 # scores the fixtures against ggml-small and self-skips when the model is
 # absent, and TestGoldenFixturesAreCanonical, which checks the committed audio
 # and the manifest with no model at all.
+# audit — the repo inventory: LOC per package, open TODOs, real stubs, living-doc
+# staleness, test shape, packages with no test. Read-only, prints, writes nothing.
+# Run it instead of rebuilding the same greps by hand; past sessions spent 93 of
+# them on this before their first edit. SECTION=loc|todo|stubs|docs|tests|gaps
+# narrows it.
+SECTION ?= all
+
+audit:
+	@SECTION="$(SECTION)" ./scripts/audit.sh
+
 stt-fixtures:
 	./scripts/gen-stt-fixtures.sh
 

cmd/mavcaldav/main.go

@@ -50,10 +50,10 @@ func run(args []string) error {
 	socket := fs.String("socket", "", "core IPC socket path (required)")
 	url := fs.String("url", "", "CalDAV calendar URL, e.g. http://localhost:5232/kami/personal (required)")
 	user := fs.String("user", "", "CalDAV basic-auth username (required)")
-	pass := fs.String("pass", "", "CalDAV basic-auth password (required)")
+	passFile := fs.String("pass-file", "", "file holding the CalDAV basic-auth password (required — never passed as a flag value)")
 	renderURL := fs.String("render-url", "", "CalDAV collection maven publishes her own reminders to; empty disables rendering")
 	renderUser := fs.String("render-user", "", "basic-auth username for -render-url (defaults to -user)")
-	renderPass := fs.String("render-pass", "", "basic-auth password for -render-url (defaults to -pass)")
+	renderPassFile := fs.String("render-pass-file", "", "file holding the password for -render-url (defaults to -pass-file)")
 	renderDur := fs.Duration("render-duration", calendar.DefaultReminderDuration, "how long a rendered reminder occupies")
 	interval := fs.Duration("interval", 5*time.Minute, "poll cadence")
 	timeout := fs.Duration("timeout", 10*time.Second, "per-request HTTP timeout")
@@ -63,13 +63,22 @@ func run(args []string) error {
 	if *socket == "" {
 		return fmt.Errorf("-socket is required")
 	}
-	if *url == "" || *user == "" || *pass == "" {
-		return fmt.Errorf("-url, -user, -pass are required")
+	if *url == "" || *user == "" || *passFile == "" {
+		return fmt.Errorf("-url, -user, -pass-file are required")
 	}
 	if err := checkRenderTarget([]string{*url}, *renderURL); err != nil {
 		return err
 	}
 
+	// The password is read from a file, never taken as a flag value: an argv
+	// secret is visible in `ps` to every user on the box and lands in the compose
+	// file and the shell history. Same rule mavmaild and mavpoll follow. Read
+	// once at start, so a rotated password means a restart.
+	pass, err := readSecret(*passFile)
+	if err != nil {
+		return err
+	}
+
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
 	defer stop()
 
@@ -85,17 +94,20 @@ func run(args []string) error {
 		http: hc,
 		url:  strings.TrimRight(*url, "/"),
 		user: *user,
-		pass: *pass,
+		pass: pass,
 	}
 
 	var rend *renderer
 	if *renderURL != "" {
-		ru, rp := *renderUser, *renderPass
+		ru, rp := *renderUser, pass
 		if ru == "" {
 			ru = *user
 		}
-		if rp == "" {
-			rp = *pass
+		if *renderPassFile != "" {
+			rp, err = readSecret(*renderPassFile)
+			if err != nil {
+				return err
+			}
 		}
 		rend = newRenderer(core, hc, *renderURL, ru, rp, *renderDur)
 		log.Printf("mavcaldav: rendering reminders to %s", *renderURL)
@@ -131,6 +143,21 @@ func run(args []string) error {
 // It takes the whole read set, not one URL. The guarantee in the package
 // comment is about every calendar maven reads, and a second read target added
 // later must not quietly fall outside the check.
+// readSecret reads one credential from a file and refuses an empty one. An
+// empty file is a deployment mistake, not a password, and CalDAV basic auth
+// would send it and get a 401 every poll.
+func readSecret(path string) (string, error) {
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		return "", fmt.Errorf("read password file: %w", err)
+	}
+	secret := strings.TrimSpace(string(raw))
+	if secret == "" {
+		return "", fmt.Errorf("password file %s is empty", path)
+	}
+	return secret, nil
+}
+
 func checkRenderTarget(readURLs []string, renderURL string) error {
 	if renderURL == "" {
 		return nil

cmd/mavcaldav/main_test.go

@@ -5,12 +5,38 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
 	"testing"
 	"time"
 
 	"github.com/kami/maven/internal/ipc"
 )
 
+// The password comes from a file so it never reaches argv. An empty or missing
+// file must fail at start rather than authenticate as "" against his calendar.
+func TestReadSecret(t *testing.T) {
+	dir := t.TempDir()
+	good := filepath.Join(dir, "ok")
+	if err := os.WriteFile(good, []byte("  hunter2\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if got, err := readSecret(good); err != nil || got != "hunter2" {
+		t.Fatalf("readSecret(good) = %q, %v; want \"hunter2\", nil", got, err)
+	}
+
+	empty := filepath.Join(dir, "empty")
+	if err := os.WriteFile(empty, []byte("\n \n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := readSecret(empty); err == nil {
+		t.Fatal("readSecret(empty) = nil error, want refusal")
+	}
+	if _, err := readSecret(filepath.Join(dir, "absent")); err == nil {
+		t.Fatal("readSecret(absent) = nil error, want refusal")
+	}
+}
+
 type fakeCore struct {
 	ipc.UnimplementedCoreAPI
 	facts    map[string]ipc.Fact // composite key "key|source" → Fact

cmd/mavend/actions_query.go

@@ -13,6 +13,7 @@ import (
 	"github.com/kami/maven/internal/crawl"
 	"github.com/kami/maven/internal/decision"
 	"github.com/kami/maven/internal/ipc"
+	"github.com/kami/maven/internal/kiwix"
 	"github.com/kami/maven/internal/memory"
 	"github.com/kami/maven/internal/morning"
 	"github.com/kami/maven/internal/phraser"
@@ -59,6 +60,35 @@ type querySource struct {
 	// sources search text with no notion of a day. When one of them grows a
 	// date parameter, flip its flag here.
 	dateAware bool
+
+	// dest — the destination this source serves, when the cascade named one
+	// (V-655). Several sources share a destination: the three recall passes and
+	// the fact-by-key lookup are all SourceRecall, because which of them lands
+	// the hit is an ordering detail no utterance can name. A source with no
+	// dest is reachable only by walking the chain.
+	dest router.Source
+
+	// guesses — this source decides whether the turn is its own by scoring the
+	// utterance against frozen seeds, rather than by looking something up and
+	// coming back empty.
+	//
+	// The distinction is the whole point of the field. A source that looks can
+	// be wrong about relevance and still harmless, because the miss shows up as
+	// no rows. A source that guesses answers whatever it claims: weather has no
+	// local table to miss against, so "что такое TCP?" became "для какого
+	// города?". So when the cascade names a destination, the guessers that were
+	// not named do not get to try. The lookups still run, because a named
+	// destination is evidence and not a promise.
+	guesses bool
+
+	// boundary — dropping this source widens what leaves the box, so only a
+	// literal pattern may do it (V-666, owner's call of 2026-08-09).
+	//
+	// Every other guesser costs an answer when it is wrongly taken off a turn.
+	// This one costs the rule that a question about him never reaches an
+	// upstream engine. A grammar read the words to name a destination. A model
+	// and a softmax both inferred one, and neither may spend that.
+	boundary bool
 }
 
 // querySources is the ordered chain actionQuery walks; first source to claim
@@ -67,85 +97,85 @@ type querySource struct {
 // gate was never the bug. Adding a source (Kiwix, RSS, crawler, email) is one
 // line here plus its method; where you put the line is the whole decision.
 var querySources = []querySource{
-	{name: "fact-by-key", answer: (*reactiveHandler).queryFactByKey},
+	{name: "fact-by-key", answer: (*reactiveHandler).queryFactByKey, dest: router.SourceRecall},
 	// Before "calendar" on purpose: both match "…на сегодня", and the plan is
 	// the more specific ask (its matcher requires a plan word), so the calendar
 	// listing would otherwise swallow it.
-	{name: "day-plan", answer: (*reactiveHandler).queryDayPlan},
+	{name: "day-plan", answer: (*reactiveHandler).queryDayPlan, dest: router.SourceCalendar},
 	// Also before "calendar": "что я обычно делаю по средам?" names a weekday,
 	// and the habit question is the more specific one. Its matcher requires a
 	// habit marker ("обычно", "каждый", …), so a question about this coming
 	// Wednesday still reaches the calendar.
-	{name: "habits", answer: (*reactiveHandler).queryHabits},
+	{name: "habits", answer: (*reactiveHandler).queryHabits, dest: router.SourceCalendar},
 	// Before "calendar" and before the recall sources: "что мне нужно
 	// сделать?" is a question about the task list, and the notes pass would
 	// otherwise answer it with whatever note happens to be nearest. Its
 	// matcher requires a task noun or an explicit "что … сделать", so a
 	// date-bearing question still reaches the calendar.
-	{name: "tasks", answer: (*reactiveHandler).queryTasks},
+	{name: "tasks", answer: (*reactiveHandler).queryTasks, dest: router.SourceTasks},
 	// Next to "tasks" and for the same reason: "что требует внимания?" is a
 	// question about the operational state Praxis holds, and it used to fall
 	// through every source to the web search (Vikunja #475). Its matcher needs
 	// an attention marker, and it falls through when Praxis is not configured.
-	{name: "attention", answer: (*reactiveHandler).queryAttention},
+	{name: "attention", answer: (*reactiveHandler).queryAttention, dest: router.SourceAttention, guesses: true},
 	// Next to "tasks" and for the same reason: "что мне купить?" is a question
 	// about the shopping list, and the recall pass would otherwise answer it
 	// from an old note about the shop. Its matcher needs an explicit list
 	// marker, so "надо бы съездить в магазин" is untouched.
-	{name: "list", answer: (*reactiveHandler).queryList},
+	{name: "list", answer: (*reactiveHandler).queryList, dest: router.SourceList, guesses: true},
 	// Before the recall sources too: "сколько я потратил?" is a question about
 	// the money facts the poller wrote, and the notes pass would otherwise
 	// answer it from whatever he once said about spending. Its matcher needs a
 	// money noun plus an actual ask, so "я потратил весь день" is untouched.
-	{name: "money", answer: (*reactiveHandler).queryMoney},
+	{name: "money", answer: (*reactiveHandler).queryMoney, dest: router.SourceMoney},
 	// Also above the recall sources: "что я тебе говорил?" is a question about
 	// the facts he tapped in, and the notes pass would answer it with whatever
 	// note is nearest (Vikunja #456). Its matcher needs both halves of a
 	// history phrase and bails out when he names a topic, so "что я говорил
 	// про сервер" is still recall.
-	{name: "history", answer: (*reactiveHandler).queryHistory},
+	{name: "history", answer: (*reactiveHandler).queryHistory, dest: router.SourceRecall},
 	// Before the recall sources and before general knowledge: "что нового?" is
 	// a question about the feeds she reads, and general knowledge would answer
 	// it by inventing news. Its matcher needs a feed noun plus an ask, so
 	// "у меня новая лента в инстаграме" is untouched.
-	{name: "feeds", answer: (*reactiveHandler).queryFeeds},
+	{name: "feeds", answer: (*reactiveHandler).queryFeeds, dest: router.SourceFeeds, guesses: true},
 	// Before "calendar" and before the recall sources: "что включено дома?" is
 	// a question about the house, and the notes pass would otherwise answer it
 	// from whatever he once said about the lights. Its matcher needs a house
 	// marker plus an ask plus a device word, and it bails out on weather
 	// wording, so "какая температура на улице?" still reaches the weather
 	// source.
-	{name: "home", answer: (*reactiveHandler).queryHome},
+	{name: "home", answer: (*reactiveHandler).queryHome, dest: router.SourceHome, guesses: true},
 	// Next to "home" and for the same reason: "какие устройства в сети?" is a
 	// question about the LAN, and the recall pass would otherwise answer it
 	// from an old note about the router. Its matcher needs a network word plus
 	// an ask plus a device noun, so "интернет не работает" is untouched.
-	{name: "network", answer: (*reactiveHandler).queryNetwork},
-	{name: "calendar", answer: (*reactiveHandler).queryCalendar, dateAware: true},
-	{name: "weather", answer: (*reactiveHandler).queryWeather},
+	{name: "network", answer: (*reactiveHandler).queryNetwork, dest: router.SourceNetwork, guesses: true},
+	{name: "calendar", answer: (*reactiveHandler).queryCalendar, dateAware: true, dest: router.SourceCalendar},
+	{name: "weather", answer: (*reactiveHandler).queryWeather, dest: router.SourceWeather, guesses: true},
 	// A question about her, above the three sources that search his own data
 	// (Vikunja #555). It has no answer anywhere else: below the boundary
 	// SearXNG answers about somebody else's assistant, and above it his notes
 	// answer by proximity — "кто ты" came back from a note of his, measured on
 	// the box, because the recall index has no idea the subject is her.
-	{name: "self", answer: (*reactiveHandler).querySelf},
-	{name: "embed", answer: (*reactiveHandler).queryEmbed},
-	{name: "memory", answer: (*reactiveHandler).queryMemory},
-	{name: "notes", answer: (*reactiveHandler).queryNotes},
+	{name: "self", answer: (*reactiveHandler).querySelf, dest: router.SourceSelf, guesses: true},
+	{name: "embed", answer: (*reactiveHandler).queryEmbed, dest: router.SourceRecall},
+	{name: "memory", answer: (*reactiveHandler).queryMemory, dest: router.SourceRecall},
+	{name: "notes", answer: (*reactiveHandler).queryNotes, dest: router.SourceRecall},
 	// THE BOUNDARY. Everything above answers from his own data; everything
 	// below answers from the world's. A question about him that got this far
 	// has no answer in his data, and no outside source can supply one, so this
 	// stops the walk rather than let the encyclopedia and the model guess.
-	{name: "personal", answer: (*reactiveHandler).queryPersonal},
+	{name: "personal", answer: (*reactiveHandler).queryPersonal, dest: router.SourceRecall, guesses: true, boundary: true},
 	// The world, read live. Owner's ruling of 2026-08-02: a metasearch hit beats
 	// a frozen ZIM, so SearXNG asks before Kiwix does. Nothing of his is at
 	// stake by this point — the boundary above already stopped every question
 	// about him, and only the query string leaves the box.
-	{name: "search", answer: (*reactiveHandler).querySearch},
+	{name: "search", answer: (*reactiveHandler).querySearch, dest: router.SourceWorld},
 	// The offline encyclopedia, now the fallback for when the line is down or
 	// the search comes back empty. It reads the way it always did; what changed
 	// is that it no longer gets first refusal on a world question.
-	{name: "kiwix", answer: (*reactiveHandler).queryKiwix},
+	{name: "kiwix", answer: (*reactiveHandler).queryKiwix, dest: router.SourceWorld},
 	// LAST before the model answers from memory, and that position is the whole
 	// design (Vikunja #259): everything of his, then the search, then the ZIMs,
 	// and only then a page he named. The model does NOT come first: it
@@ -153,8 +183,48 @@ var querySources = []querySource{
 	// a 1.7B guessing at a page it cannot read is how contents get invented.
 	// This source only claims a turn where he named a URL, so it never competes
 	// with a local answer.
-	{name: "web", answer: (*reactiveHandler).queryWeb},
-	{name: "general-knowledge", answer: (*reactiveHandler).queryGeneral},
+	{name: "web", answer: (*reactiveHandler).queryWeb, dest: router.SourceWorld},
+	{name: "general-knowledge", answer: (*reactiveHandler).queryGeneral, dest: router.SourceWorld},
+}
+
+// queryWalk narrows the chain for one turn against the destination the cascade
+// named, and says which sources were left out (V-655).
+//
+// It takes sources OUT and never moves one, which is the whole safety argument.
+// The table's order is load-bearing and every comment on it argues a reason
+// between two sources; none of those reasons is about this. Above all, the
+// order carries "his data first, then the world", and a destination named by a
+// model must not be able to reverse that. Naming SourceWorld does not send the
+// turn outside — it stops the guessers from claiming it on the way.
+//
+// What comes out is exactly the sources that guess. Those decide whether a turn
+// is theirs by scoring it against frozen seeds, and then answer whatever they
+// claimed, because they have no lookup that can come back empty. That is the
+// whole of the 2026-08-07 defect: weather claiming "что такое TCP?", the feed
+// claiming "какой у меня любимый язык?", the personal boundary claiming "кто
+// такой Линус Торвальдс?". The sources that look are all still asked, so a
+// wrong destination costs nothing but the guess it prevented.
+//
+// No destination named ⇒ the table exactly as written, which is what shipped
+// before the field existed. That is the floor. The classifier arm names
+// nothing, so a box whose model is down routes queries the way it always did.
+// The personal boundary is the one exception, and anchored is what buys it
+// (V-666). A grammar matched a literal pattern to name the destination. The
+// routing heads and the resident model inferred one, and an inferred SourceWorld
+// takes the boundary off a question about him. That widens what is asked
+// upstream rather than costing a local answer, so those two keep it.
+func queryWalk(dest router.Source, anchored bool) (walk, skipped []querySource) {
+	if dest == router.SourceUnknown {
+		return querySources, nil
+	}
+	for _, s := range querySources {
+		if s.guesses && s.dest != dest && (anchored || !s.boundary) {
+			skipped = append(skipped, s)
+			continue
+		}
+		walk = append(walk, s)
+	}
+	return walk, skipped
 }
 
 func (h *reactiveHandler) actionQuery(ctx context.Context, dec router.Decision) string {
@@ -164,7 +234,14 @@ func (h *reactiveHandler) actionQuery(ctx context.Context, dec router.Decision)
 	// (V-564). Finish names everyone below the winner.
 	decision.Expect(ctx, decision.StageQuery, querySourceNames())
 	rec := decision.From(ctx)
-	for _, src := range querySources {
+	walk, skipped := queryWalk(dec.Source, dec.SourceAnchored)
+	for _, src := range skipped {
+		rec.Note(decision.Claim{
+			Stage: decision.StageQuery, Claimant: src.name, Outcome: decision.NeverAsked,
+			Reason: "it decides by similarity and the cascade named " + string(dec.Source),
+		})
+	}
+	for _, src := range walk {
 		if dec.Continued && !src.dateAware {
 			rec.Note(decision.Claim{
 				Stage: decision.StageQuery, Claimant: src.name, Outcome: decision.NeverAsked,
@@ -850,6 +927,28 @@ func (h *reactiveHandler) queryKiwix(ctx context.Context, t *queryTurn) (string,
 		}
 	}
 
+	// The topic, not the sentence (V-668). Kiwix ranks by keyword overlap, so
+	// the question words outrank the one word that names the article: measured
+	// on 2026-08-09, "что такое TCP" returns "Перехват TCP-соединения" and
+	// "TCP" returns TCP. Only the verbatim path needs this. The rewriter
+	// already reduces a question to English keywords, and reducing twice would
+	// take the topic off the input it reads.
+	if verbatim {
+		if topic := kiwix.Topic(pattern); topic != "" {
+			// The article named exactly, before any ranking runs. A ZIM is
+			// addressable by title and a wrong title is a 404, so this either
+			// answers or costs one request that says nothing.
+			for _, cand := range kiwix.TitleCandidates(topic) {
+				page, err := h.kiwix.client.Article(ctxK, kiwix.TitlePath(book, cand), h.kiwix.runes)
+				if err == nil && page.Text != "" {
+					log.Printf("voice: kiwix: %q in %q → title hit %q", topic, book, page.Title)
+					return h.kiwixReply(ctx, t, page.Title, page.Text)
+				}
+			}
+			pattern = topic
+		}
+	}
+
 	hits, err := h.kiwix.client.Search(ctxK, pattern, book, h.kiwix.max)
 	if err != nil {
 		log.Printf("voice: kiwix: search %q: %v", pattern, err)
@@ -882,14 +981,18 @@ func (h *reactiveHandler) queryKiwix(ctx context.Context, t *queryTurn) (string,
 		}
 		page = crawl.Page{Title: top.Title, Text: top.Snippet}
 	}
-	// Handed over the same way a note or a page is: context for the question he
-	// asked, not something to recite.
-	snippet := top.Title + "\n" + crawl.TrimRunes(page.Text, h.kiwix.runes)
+	return h.kiwixReply(ctx, t, top.Title, page.Text)
+}
+
+// kiwixReply hands one article over the same way a note or a page is handed
+// over: context for the question he asked, not something to recite.
+func (h *reactiveHandler) kiwixReply(ctx context.Context, t *queryTurn, title, text string) (string, bool) {
+	snippet := title + "\n" + crawl.TrimRunes(text, h.kiwix.runes)
 	reply := h.phraseSource(ctx, "kiwix", t.dec.Utterance, []string{snippet})
 	if reply == "" {
 		// No phraser, or it failed. Read back the best hit rather than pretend
 		// the search did not happen.
-		return readBack(top.Title + " — " + page.Text), true
+		return readBack(title + " — " + text), true
 	}
 	return reply, true
 }

cmd/mavend/chat_degrade_test.go

@@ -19,7 +19,7 @@ func TestChatAnswersWithNoLlamaServer(t *testing.T) {
 	dead := llm.New("http://127.0.0.1:1", 500*time.Millisecond)
 	emb := router.NewHashEmbedder(1024)
 	h.recall.embedder = emb
-	h.router = buildRouter(emb, h.matcher, 0.55, pickLLMRouter(true, dead))
+	h.router = buildRouter(emb, h.matcher, 0.55, pickLLMRouter(true, dead), nil)
 	h.replier = newLLMReplier(dead, nil)
 
 	ctx := withDialogueID(context.Background(), dialogueIDFor(sourceText, "web"))

cmd/mavend/clarify.go

@@ -6,7 +6,6 @@ import (
 	"math/rand"
 	"strings"
 	"time"
-	"unicode"
 	"unicode/utf8"
 
 	"github.com/kami/maven/internal/dialogue"
@@ -162,12 +161,13 @@ func withNotice(notice, reply string) string {
 // напоминание?" — answer first, then the open question. A question in front of
 // its own answer would read as ignoring what he asked.
 //
-// A statement's full stop is folded into a comma, so the two acts read as one
-// sentence — that is the owner's own punctuation, "в Риме сейчас ..., на какое
-// время поставить напоминание?". An answer that is ITSELF a question keeps its
-// mark and the resume starts a new sentence: she sometimes answers a side query
-// by asking him to say it again, and "переформулировать?, на какое время" folds
-// two questions into one unreadable line.
+// Two sentences, not one (V-654). This used to fold the answer's full stop into
+// a comma, on the strength of the owner having written it that way once. Spliced
+// onto a real answer it reads as one run-on thought — "вот что я нашла: вайфай
+// пароль лежит в ящике стола, на какое время поставить напоминание?" — and the
+// question disappears into the tail of a sentence about something else. A reply
+// with no terminator of its own is given one, so the join never depends on how
+// the phraser chose to end.
 //
 // A resume with no answer in front of it is just the question.
 func withResumed(reply, resumed string) string {
@@ -178,23 +178,17 @@ func withResumed(reply, resumed string) string {
 	if reply == "" {
 		return resumed
 	}
-	if strings.HasSuffix(reply, "?") {
-		return reply + " " + resumed
+	if !endsSentence(reply) {
+		reply += "."
 	}
-	if trimmed := strings.TrimRight(reply, ".!"); trimmed != "" {
-		reply = trimmed
-	}
-	return reply + ", " + lowerFirst(resumed)
+	return reply + " " + resumed
 }
 
-// lowerFirst lowercases the opening rune, so a deck line written as a standalone
-// sentence reads as the second half of one. Only the first rune: "На какое
-// время" must become "на какое время" and nothing else in it may move.
-func lowerFirst(s string) string {
-	for i, r := range s {
-		return string(unicode.ToLower(r)) + s[i+utf8.RuneLen(r):]
-	}
-	return s
+// endsSentence reports whether s already closes itself. The ellipsis counts: a
+// trailing "…" is a deliberate end, and a full stop after it reads as a typo.
+func endsSentence(s string) bool {
+	r, _ := utf8.DecodeLastRuneInString(s)
+	return strings.ContainsRune(".!?…", r)
 }
 
 // missingFor returns the slots a decision still needs, most important first.
@@ -381,6 +375,14 @@ func (h *reactiveHandler) resolveClarifyAnswer(ctx context.Context, text string)
 		return "", false
 	}
 
+	// He is answering, so the run of step-asides is over (V-654). Reset here
+	// rather than where a gap is FILLED: "позвонить маме" against a question
+	// about the time gives her nothing she asked for and still means he is in
+	// the exchange, and the retry it costs is bound enough on its own. The
+	// counter is for the case the bounds miss — he asked for other things and
+	// never came back.
+	q.Suspends = 0
+
 	merged := q.Answer(text, toDialogueSlots(answer))
 	// Fold a newly answered subject into the raw utterance. Downstream actions
 	// phrase from Utterance, not from the text slot — actionReminder stores it
@@ -467,6 +469,14 @@ func (h *reactiveHandler) noteDropped(ctx context.Context) {
 //
 // A slot with no resumed wording (clarifyResumedFor says so) resumes nothing and
 // says nothing. She must not claim to be holding a question she cannot re-ask.
+//
+// Suspension is bounded, since V-654. Neither of the two things above is a
+// limit: no attempt is spent, and restarting the clock means the TTL cannot
+// arrive while he keeps talking. So the count is the only thing that ends it,
+// and past MaxSuspends she lets the request go and says so with the same line
+// every other drop uses. The rule is unchanged — a question ends by being
+// answered or by being let go out loud — this only recognises three unrelated
+// requests in a row as the second of those.
 func (h *reactiveHandler) noteSuspended(ctx context.Context, q *dialogue.PendingQuestion) {
 	rt := turnRouteFrom(ctx)
 	if rt == nil || len(q.Missing) == 0 {
@@ -476,11 +486,22 @@ func (h *reactiveHandler) noteSuspended(ctx context.Context, q *dialogue.Pending
 	if !ok {
 		return
 	}
+	if !q.CanResume() {
+		h.clarifyStore.Delete(dialogueIDOf(ctx))
+		h.noteDropped(ctx)
+		log.Printf("voice: clarify — letting the question about %s go: %d asides in a row, %d rides in all", q.Missing[0], q.Suspends, q.Rides)
+		return
+	}
+	q.Suspends++
+	// Rides is the same event counted without the reset (V-663). Incremented
+	// beside Suspends and never anywhere else, so the two cannot disagree about
+	// what happened, only about how much of it they remember.
+	q.Rides++
 	q.Asked = h.now()
 	h.clarifyStore.Put(dialogueIDOf(ctx), q)
 	rt.resume = question
 	rt.suspended = true
-	log.Printf("voice: clarify — is its own request; suspending the question about %s and resuming it in the same reply", q.Missing[0])
+	log.Printf("voice: clarify — is its own request; suspending the question about %s and resuming it in the same reply (suspend %d of %d, ride %d of %d)", q.Missing[0], q.Suspends, dialogue.MaxSuspends, q.Rides, dialogue.MaxRides)
 }
 
 // foldAnswerIntoUtterance appends an answered subject to the original words,
@@ -520,6 +541,14 @@ func (h *reactiveHandler) askRemainingGap(ctx context.Context, q *dialogue.Pendi
 	if !ok || !q.CanAsk() {
 		return "", false
 	}
+	// Suspends is not carried, and by this point it is already zero: the answer
+	// path resets it (V-654). Left off the literal so the zero is stated where
+	// the struct is built, rather than inherited from a field nobody names.
+	//
+	// Rides IS carried, and that is the whole point of it (V-663). This is the
+	// same request under a second question, not a new one, so the turns it has
+	// already ridden still count against it. Dropping the field here is exactly
+	// the re-basing that let one question ride twenty-six replies.
 	h.clarifyStore.Put(dialogueIDOf(ctx), &dialogue.PendingQuestion{
 		Intent:      q.Intent,
 		Slots:       merged,
@@ -530,6 +559,7 @@ func (h *reactiveHandler) askRemainingGap(ctx context.Context, q *dialogue.Pendi
 		TTL:         clarifyTTL,
 		Attempts:    q.Attempts + 1,
 		MaxAttempts: q.MaxAttempts,
+		Rides:       q.Rides,
 	})
 	log.Printf("voice: clarify — one gap filled, still missing %s for intent=%s, asking again (attempt %d)", remaining[0], intent, q.Attempts+1)
 	return question, true

cmd/mavend/clarify_test.go

@@ -317,7 +317,7 @@ func TestClarifyExpiryIsAnnouncedAndWordsStillRoute(t *testing.T) {
 	h, _, now := newClarifyHandler(t)
 	emb := router.NewHashEmbedder(1024)
 	h.recall.embedder = emb
-	h.router = buildRouter(emb, h.matcher, 0.55, nil)
+	h.router = buildRouter(emb, h.matcher, 0.55, nil, nil)
 
 	if _, asked := h.askClarify(ctx, clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {
 		t.Fatal("expected a question")
@@ -671,7 +671,7 @@ func TestUnresolvedActSaysItDoesNotKnowTheCommand(t *testing.T) {
 func newRoutingClarifyHandler(t *testing.T) (*reactiveHandler, *store.Store) {
 	t.Helper()
 	h, st, _ := newClarifyHandler(t)
-	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil)
+	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil, nil)
 	h.recall = recallWiring{embedder: router.NewHashEmbedder(1024), memStore: memory.NewInMemoryStore()}
 	return h, st
 }
@@ -740,3 +740,53 @@ func TestACompleteTurnStillDoesNotAsk(t *testing.T) {
 		}
 	}
 }
+
+// TestTheResumedQuestionIsItsOwnSentence — V-654. The re-ask used to be spliced
+// onto the answer with a comma, so a real answer and an unrelated open question
+// read as one run-on thought and the question vanished into its tail.
+func TestTheResumedQuestionIsItsOwnSentence(t *testing.T) {
+	const resumed = "На какое время поставить напоминание?"
+	cases := []struct {
+		name  string
+		reply string
+		want  string
+	}{
+		{
+			// The measured line, shortened. Two sentences, and the question keeps
+			// its capital.
+			name:  "a statement keeps its full stop",
+			reply: "Вайфай пароль лежит в ящике стола.",
+			want:  "Вайфай пароль лежит в ящике стола. " + resumed,
+		},
+		{
+			name:  "a reply with no terminator is given one",
+			reply: "Вайфай пароль лежит в ящике стола",
+			want:  "Вайфай пароль лежит в ящике стола. " + resumed,
+		},
+		{
+			// She sometimes answers a side query by asking him to say it again.
+			// Two questions, and neither may swallow the other.
+			name:  "a question keeps its mark",
+			reply: "Можешь переформулировать?",
+			want:  "Можешь переформулировать? " + resumed,
+		},
+		{
+			name:  "an ellipsis is already an ending",
+			reply: "Не уверена…",
+			want:  "Не уверена… " + resumed,
+		},
+		{
+			name:  "a resume with no answer in front of it is just the question",
+			reply: "",
+			want:  resumed,
+		},
+	}
+	for _, tc := range cases {
+		if got := withResumed(tc.reply, resumed); got != tc.want {
+			t.Errorf("%s: withResumed(%q) = %q, want %q", tc.name, tc.reply, got, tc.want)
+		}
+	}
+	if got := withResumed("Готово.", ""); got != "Готово." {
+		t.Errorf("nothing to resume must leave the reply alone, got %q", got)
+	}
+}

cmd/mavend/decisiontrace_test.go

@@ -25,7 +25,7 @@ func traceHandler(t *testing.T, ring *decision.Ring) *reactiveHandler {
 	return &reactiveHandler{
 		api:       api,
 		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
-		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil),
+		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
 		dataStore: st,

cmd/mavend/dialogue_contract_test.go

@@ -171,7 +171,7 @@ func newDialogueHandler(t *testing.T) (*reactiveHandler, *store.Store, *time.Tim
 	// and never a coincidence (V-577, V-579). checkEnd refuses any reminder
 	// landing on it, and at 09:00 the row that answers "на 9" would trip that.
 	*now = time.Date(2026, 7, 31, 9, 17, 0, 0, time.UTC)
-	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil)
+	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil, nil)
 	h.recall = recallWiring{embedder: router.NewHashEmbedder(1024), memStore: memory.NewInMemoryStore()}
 	return h, st, now
 }

cmd/mavend/fact_subject_test.go

@@ -24,7 +24,7 @@ func TestApplyAction_FactCapture_QueuesEntityResolution(t *testing.T) {
 
 	emb := router.NewHashEmbedder(1024)
 	matcher := tool.NewMatcher(api)
-	rtr := buildRouter(emb, matcher, 0.55, nil)
+	rtr := buildRouter(emb, matcher, 0.55, nil, nil)
 
 	h := &reactiveHandler{
 		api:       api,

cmd/mavend/factgate_test.go

@@ -20,7 +20,7 @@ func newFactGateHandler(t *testing.T, now time.Time) (*reactiveHandler, ipc.Core
 	h := &reactiveHandler{
 		api:       api,
 		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
-		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil),
+		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
 		dataStore: st,

cmd/mavend/notefragment_test.go

@@ -45,7 +45,7 @@ func newNoteHandler(t *testing.T) (*reactiveHandler, *store.Store) {
 	h := &reactiveHandler{
 		api:       api,
 		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
-		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil),
+		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
 		dataStore: st,

cmd/mavend/querywalk_test.go

@@ -0,0 +1,137 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// The floor, and it is the reason a destination is safe to add at all: a box
+// whose model is down names nothing, and naming nothing has to walk the chain
+// the way it walked before the field existed.
+func TestNoDestinationWalksTheWholeChain(t *testing.T) {
+	walk, skipped := queryWalk(router.SourceUnknown, false)
+	if len(skipped) != 0 {
+		t.Errorf("skipped %d sources with no destination named, want none", len(skipped))
+	}
+	if len(walk) != len(querySources) {
+		t.Fatalf("walk has %d sources, want the whole table of %d", len(walk), len(querySources))
+	}
+	for i := range walk {
+		if walk[i].name != querySources[i].name {
+			t.Fatalf("position %d is %q, want %q", i, walk[i].name, querySources[i].name)
+		}
+	}
+}
+
+// The 2026-08-07 defects, one per line. Each is a source that decides by seed
+// similarity claiming a turn that was never its own, and then answering it
+// because it has no lookup that could come back empty.
+func TestANamedDestinationSilencesTheOtherGuessers(t *testing.T) {
+	cases := []struct {
+		dest      router.Source
+		utterance string
+		silenced  string
+		anchored  bool // a stage 0 grammar named the destination
+	}{
+		{router.SourceWorld, "что такое TCP?", "weather", true},
+		{router.SourceWorld, "сколько будет 17 на 23?", "weather", true},
+		{router.SourceWorld, "кто такой Линус Торвальдс?", "personal", true},
+		{router.SourceRecall, "какой у меня любимый язык?", "feeds", false},
+		{router.SourceCalendar, "что в календаре на завтра?", "weather", true},
+	}
+	for _, c := range cases {
+		walk, skipped := queryWalk(c.dest, c.anchored)
+		if inWalk(walk, c.silenced) {
+			t.Errorf("%q named %q: %q is still asked", c.utterance, c.dest, c.silenced)
+		}
+		if !inWalk(skipped, c.silenced) {
+			t.Errorf("%q named %q: %q is missing from the record of who was skipped",
+				c.utterance, c.dest, c.silenced)
+		}
+	}
+}
+
+// Naming the world must not send the turn outside. His notes, his facts and the
+// boundary in front of them are the invariant CLAUDE.md states as "the owner's
+// data first, then the world", and a destination a model wrote must not be able
+// to reverse it.
+func TestNamingTheWorldStillReadsHisDataFirst(t *testing.T) {
+	walk, _ := queryWalk(router.SourceWorld, true)
+	for _, look := range []string{"fact-by-key", "embed", "memory", "notes"} {
+		if !inWalk(walk, look) {
+			t.Errorf("%q was dropped; only the sources that guess may be dropped", look)
+		}
+	}
+	if posOf(walk, "notes") > posOf(walk, "search") {
+		t.Error("search is asked before his notes are")
+	}
+	if posOf(walk, "search") < 0 {
+		t.Fatal("search is not in the walk at all")
+	}
+}
+
+// The boundary belongs to his data, so naming recall keeps it. That is what
+// makes "какой у меня любимый язык?" answer "не нашла у тебя такой записи"
+// rather than reaching SearXNG once nothing local had it.
+func TestNamingRecallKeepsTheBoundary(t *testing.T) {
+	walk, _ := queryWalk(router.SourceRecall, true)
+	if !inWalk(walk, "personal") {
+		t.Fatal("the personal boundary was skipped on a turn named for his own data")
+	}
+	if posOf(walk, "personal") > posOf(walk, "search") {
+		t.Error("the boundary no longer sits in front of the world")
+	}
+}
+
+// The owner's call of 2026-08-09 (V-666): only a stage 0 grammar may take the
+// personal boundary off a turn. The routing heads and the resident model both
+// name a destination by inference, and an inferred SourceWorld would send a
+// question about him upstream. Every other guesser still goes.
+func TestOnlyAGrammarMayDropTheBoundary(t *testing.T) {
+	walk, skipped := queryWalk(router.SourceWorld, false)
+	if !inWalk(walk, "personal") {
+		t.Error("an inferred destination took the boundary off the turn")
+	}
+	if !inWalk(skipped, "weather") {
+		t.Error("weather is still asked; the rule covers the boundary alone")
+	}
+	if posOf(walk, "personal") > posOf(walk, "search") {
+		t.Error("the boundary no longer sits in front of the world")
+	}
+	if anchored, _ := queryWalk(router.SourceWorld, true); inWalk(anchored, "personal") {
+		t.Error(`a grammar named the world and the boundary stayed: ` +
+			`"кто такой Линус Торвальдс?" is answered "не нашла у тебя такой записи" again`)
+	}
+}
+
+// Whatever the destination, the walk is a subsequence of the table. Every
+// comment on that table argues an order between two sources, and none of those
+// reasons is about this field.
+func TestTheWalkNeverReordersTheTable(t *testing.T) {
+	for _, dest := range append([]router.Source{router.SourceUnknown}, router.Sources...) {
+		walk, skipped := queryWalk(dest, true)
+		if len(walk)+len(skipped) != len(querySources) {
+			t.Errorf("%q: %d walked + %d skipped, want %d", dest, len(walk), len(skipped), len(querySources))
+		}
+		last := -1
+		for _, s := range walk {
+			at := posOf(querySources, s.name)
+			if at <= last {
+				t.Errorf("%q: %q is out of table order", dest, s.name)
+			}
+			last = at
+		}
+	}
+}
+
+func inWalk(list []querySource, name string) bool { return posOf(list, name) >= 0 }
+
+func posOf(list []querySource, name string) int {
+	for i, s := range list {
+		if s.name == name {
+			return i
+		}
+	}
+	return -1
+}

cmd/mavend/reactive_notes_test.go

@@ -22,7 +22,7 @@ func TestReactiveNotesReminders(t *testing.T) {
 
 	emb := router.NewHashEmbedder(1024)
 	matcher := tool.NewMatcher(api)
-	rtr := buildRouter(emb, matcher, 0.55, nil)
+	rtr := buildRouter(emb, matcher, 0.55, nil, nil)
 
 	h := &reactiveHandler{
 		api:       api,
@@ -104,7 +104,7 @@ func TestSpokenTaskCaptureFilesATask(t *testing.T) {
 	h := &reactiveHandler{
 		api:       api,
 		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
-		router:    buildRouter(emb, matcher, 0.55, nil),
+		router:    buildRouter(emb, matcher, 0.55, nil, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
 		dataStore: st,

cmd/mavend/simulator_test.go

@@ -474,7 +474,7 @@ func newSimWorld(t *testing.T, sc scenario) *simWorld {
 	// used to be built on a nil API, which meant any scenario that produced an
 	// act panicked the moment the matcher was consulted.
 	matcher := tool.NewMatcher(api)
-	rtr := buildRouter(emb, matcher, config.DefaultRouterThreshold, router.NewLLMRouter(scripted))
+	rtr := buildRouter(emb, matcher, config.DefaultRouterThreshold, router.NewLLMRouter(scripted), nil)
 
 	w.handler = &reactiveHandler{
 		stt:    simTranscriber{},

cmd/mavend/sttseam_test.go

@@ -0,0 +1,43 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/kami/maven/internal/config"
+	"github.com/kami/maven/internal/stt"
+)
+
+// A box with no workstation.stt block transcribes exactly as it did before the
+// seam existed: the floor is handed back untouched, and nothing probes.
+func TestSttSeamWithNoBlockIsTheFloor(t *testing.T) {
+	floor := stt.NewStub()
+	got, pair := sttSeam(&config.Config{}, floor)
+	if pair != nil {
+		t.Fatal("no block must build no pair")
+	}
+	if got != stt.Transcriber(floor) {
+		t.Fatal("no block must hand back the floor itself")
+	}
+}
+
+func TestSttSeamPrefersTheWorkstation(t *testing.T) {
+	cfg := &config.Config{Workstation: &config.WorkstationConfig{
+		URL: "http://127.0.0.1:1",
+		Stt: &config.WorkstationSttConfig{
+			URL:    "http://127.0.0.1:2/transcribe",
+			Health: "http://127.0.0.1:2/health",
+		},
+	}}
+	got, pair := sttSeam(cfg, stt.NewStub())
+	if pair == nil {
+		t.Fatal("a configured block must build a pair")
+	}
+	defer pair.Stop()
+	if got != stt.Transcriber(pair) {
+		t.Fatal("the pair is what callers must transcribe through")
+	}
+	// Nothing answers on port 2, so the seam is the floor until it does.
+	if pair.Available() {
+		t.Fatal("an unreachable workstation must not be available")
+	}
+}

cmd/mavend/turnrole.go

@@ -186,6 +186,28 @@ func carriesReminderVerb(text string) bool {
 	return false
 }
 
+// isPleasantry matches the WHOLE utterance against lexicon.Pleasantries, after
+// lowercasing and dropping the punctuation a greeting carries.
+//
+// Whole utterance and not tokens. Every token rule tried here was wrong on
+// something: "вечер" answers "это утра или вечера?", "нет" answers a confirm,
+// and "спокойной" alone is not an utterance at all. A greeting is a fixed
+// phrase, so matching it as one costs nothing and claims nothing else.
+func isPleasantry(text string) bool {
+	t := strings.ToLower(strings.TrimSpace(text))
+	t = strings.Trim(t, " .,!?…")
+	t = strings.Join(strings.Fields(t), " ")
+	if t == "" {
+		return false
+	}
+	for _, p := range lexicon.Pleasantries() {
+		if t == p {
+			return true
+		}
+	}
+	return false
+}
+
 // offlineOwnRequest is the shape half of the evidence: the offline token tests,
 // which cost nothing and never depend on the model that produced the routing.
 // It is also the whole answer when there is no route to read — the classifier
@@ -225,6 +247,16 @@ func classifyTurnRole(q *dialogue.PendingQuestion, text string, answer dialogue.
 	// hour, and no route saying "question" changes that. It works because the
 	// extractor no longer reads a day word as the current clock, so a sentence
 	// that names no hour now fills nothing to weigh.
+	// A pleasantry is neither (V-663). "спасибо" and "привет" fell through to
+	// roleAnswer, so a question about a reminder's DAY was re-asked at a man
+	// saying thank you, and the retry it spent was one of the three bounds
+	// meant to end the ride. It is an aside: answered as itself, the question
+	// resumed on the tail, no attempt spent, one ride counted. Placed above the
+	// content gate because "доброе утро" has content and states nothing, so
+	// neither half of the evidence below can reach it.
+	if q != nil && isPleasantry(text) {
+		return roleAside
+	}
 	own := false
 	if len(ownContent(text)) > 0 {
 		own = offlineOwnRequest(text) || (ok && carriesOwnRequest(routed, text))

cmd/mavend/turnrole_test.go

@@ -279,3 +279,162 @@ func TestTheTurnIsRoutedOnce(t *testing.T) {
 		t.Fatalf("the pipeline routed again and got something else: %+v vs %+v", second, first)
 	}
 }
+
+// TestASuspendedQuestionDoesNotRideForever — V-654, the measured failure of
+// 2026-08-07 (docs/evals/2026-08-07-week-of-usage-transcript.md, t=51 to t=58).
+//
+// A side query suspends the parked question, spends no attempt and restarts the
+// TTL. Nothing else bounded it, so one unfilled time slot came back on the end
+// of six consecutive unrelated replies and stopped only when a seventh turn
+// happened to read as a failed answer. Three step-asides, then she lets it go
+// and says so.
+func TestASuspendedQuestionDoesNotRideForever(t *testing.T) {
+	ctx := context.Background()
+	h, st := newRoutingClarifyHandler(t)
+	id := dialogueIDFor(sourceText, "web")
+	resumed, _ := clarifyResumedFor(dialogue.SlotTime)
+
+	if reply := h.handleText(ctx, "web", "напомни позвонить маме"); !strings.Contains(reply, "?") {
+		t.Fatalf("expected the time question, got %q", reply)
+	}
+
+	// Three questions of his own. Each one is answered as itself and each one
+	// brings the open question back, exactly as V-561 asks.
+	asides := []string{
+		"о чём мы вчера говорили?",
+		"какие у меня напоминания?",
+		"сколько времени?",
+	}
+	for i, text := range asides {
+		reply := h.handleText(ctx, "web", text)
+		if !strings.HasSuffix(reply, resumed) {
+			t.Fatalf("side query %d: the question must come back, got %q", i+1, reply)
+		}
+		if strings.Contains(reply, clarifyDropped) {
+			t.Fatalf("side query %d: nothing was let go yet, so nothing may say so: %q", i+1, reply)
+		}
+		q := h.clarifyStore.Get(id, h.now())
+		if q == nil {
+			t.Fatalf("side query %d: the question was dropped early", i+1)
+		}
+		if q.Attempts != 1 {
+			t.Fatalf("side query %d: a step-aside spent an attempt: %d", i+1, q.Attempts)
+		}
+		if q.Suspends != i+1 {
+			t.Fatalf("side query %d: suspends = %d, want %d", i+1, q.Suspends, i+1)
+		}
+	}
+
+	// The fourth. She has stepped aside as often as she is willing to, so the
+	// request goes — out loud, and without the question on the tail.
+	reply := h.handleText(ctx, "web", "что у меня сегодня?")
+	if !strings.Contains(reply, clarifyDropped) {
+		t.Fatalf("the request was let go in silence: %q", reply)
+	}
+	if strings.HasSuffix(reply, resumed) {
+		t.Fatalf("a question she has let go must not be asked again: %q", reply)
+	}
+	if h.clarifyStore.Get(id, h.now()) != nil {
+		t.Fatal("the question must be gone once she has said she let it go")
+	}
+	if reminders, err := st.DueReminders(ctx, h.now().Add(48*time.Hour)); err != nil || len(reminders) != 0 {
+		t.Fatalf("a reminder was invented for a time nobody gave: %v err=%v", reminders, err)
+	}
+}
+
+// TestAnAnsweredGapResetsTheSuspendBudget — the counter measures CONSECUTIVE
+// step-asides. He filled a gap, so the run is broken and the next question
+// starts with its full allowance: a long exchange he is engaged with must not
+// run out of patience on his behalf.
+func TestAnAnsweredGapResetsTheSuspendBudget(t *testing.T) {
+	ctx := context.Background()
+	h, _ := newRoutingClarifyHandler(t)
+	id := dialogueIDFor(sourceText, "web")
+
+	// A bare "напомни" is missing both halves, so answering the subject re-parks
+	// the request with a question about the time.
+	if reply := h.handleText(ctx, "web", "напомни"); !strings.Contains(reply, "?") {
+		t.Fatalf("expected a question, got %q", reply)
+	}
+	if reply := h.handleText(ctx, "web", "какие у меня напоминания?"); reply == "" {
+		t.Fatal("the side query must be answered as itself")
+	}
+	if q := h.clarifyStore.Get(id, h.now()); q == nil || q.Suspends != 1 {
+		t.Fatalf("the side query was not counted: %+v", q)
+	}
+	if reply := h.handleText(ctx, "web", "позвонить маме"); reply == "" {
+		t.Fatal("the answer must be consumed")
+	}
+	q := h.clarifyStore.Get(id, h.now())
+	if q == nil {
+		t.Fatal("a reminder still needs its time, so a question must be parked")
+	}
+	if q.Suspends != 0 {
+		t.Fatalf("answering a gap must reset the suspend budget: suspends = %d", q.Suspends)
+	}
+	// The ride it already took is carried across the re-park (V-663). Resetting
+	// both counters here is what let one question ride twenty-six replies.
+	if q.Rides != 1 {
+		t.Fatalf("the aside it already took was forgotten: rides = %d", q.Rides)
+	}
+}
+
+// TestTwoBoundsCannotRearmEachOther — V-663.
+//
+// MaxSuspends landed and the measurement did not move: twenty-six of 140 turns
+// carried a tail before it and twenty-six after. This is the shape it misses,
+// taken from the 2026-08-08 run, where one question rode turns 7 to 13.
+//
+// An aside spends no attempt, so MaxAttempts never reaches it. A turn that
+// reads as a failed answer zeroes Suspends, so MaxSuspends never reaches the
+// asides either. Alternating the two rearms each bound with the other's
+// traffic. Rides counts both kinds and is never reset, so it is what ends this.
+func TestTwoBoundsCannotRearmEachOther(t *testing.T) {
+	ctx := context.Background()
+	h, _ := newRoutingClarifyHandler(t)
+	id := dialogueIDFor(sourceText, "web")
+	resumed, _ := clarifyResumedFor(dialogue.SlotTime)
+
+	if reply := h.handleText(ctx, "web", "напомни позвонить маме"); !strings.Contains(reply, "?") {
+		t.Fatalf("expected the time question, got %q", reply)
+	}
+
+	// Two asides. Each one rides and neither spends an attempt.
+	for i := 0; i < 2; i++ {
+		reply := h.handleText(ctx, "web", "какие у меня напоминания?")
+		if !strings.HasSuffix(reply, resumed) {
+			t.Fatalf("aside %d: the question must come back, got %q", i+1, reply)
+		}
+	}
+	q := h.clarifyStore.Get(id, h.now())
+	if q == nil || q.Rides != 2 || q.Suspends != 2 {
+		t.Fatalf("after two asides: %+v", q)
+	}
+
+	// A pleasantry. It used to read as a failed answer, so she re-asked the
+	// question at a man saying thank you and spent an attempt doing it. Now it
+	// is an aside: answered as itself, question on the tail, one more ride.
+	reply := h.handleText(ctx, "web", "спасибо")
+	if !strings.HasSuffix(reply, resumed) {
+		t.Fatalf("a pleasantry lost the parked question: %q", reply)
+	}
+	q = h.clarifyStore.Get(id, h.now())
+	if q == nil || q.Attempts != 1 {
+		t.Fatalf("a pleasantry spent an attempt: %+v", q)
+	}
+	if q.Rides != 3 {
+		t.Fatalf("a pleasantry rode free: %+v", q)
+	}
+
+	// One more ride of any kind and the request goes, out loud.
+	reply = h.handleText(ctx, "web", "какие у меня напоминания?")
+	if !strings.Contains(reply, clarifyDropped) {
+		t.Fatalf("the question rode four asides and was let go in silence: %q", reply)
+	}
+	if strings.HasSuffix(reply, resumed) {
+		t.Fatalf("a question she has let go must not be asked again: %q", reply)
+	}
+	if h.clarifyStore.Get(id, h.now()) != nil {
+		t.Fatal("the question must be gone once she has said she let it go")
+	}
+}

cmd/mavend/voicewire.go

@@ -36,7 +36,9 @@ type voiceWiring struct {
 	sessions  *voice.Sessions
 	voiceSink delivery.Sink
 	embedder  router.Embedder
-	handler   *reactiveHandler // the reactive handler for IPC Chat
+	// heads — the routing heads, nil unless embedder.heads_path is set.
+	heads   *router.RouterHeads
+	handler *reactiveHandler // the reactive handler for IPC Chat
 	// worker clients (set when configured as Remote): closed on shutdown so
 	// mavsttd / mavttsd don't keep a stale conn into a restarting daemon.
 	sttClient *worker.Client
@@ -53,7 +55,11 @@ type voiceWiring struct {
 	// unless a `workstation` block names an address. Held here only so the
 	// prober is stopped on shutdown; callers were handed it at build time.
 	pair *llm.Pair
-	mcp  *mcpWiring
+	// sttPair — CrisperWhisper 2.0 on the workstation with mavsttd as the
+	// floor, nil unless the `workstation.stt` block names an address. Held for
+	// the same reason as pair: to stop its prober on shutdown.
+	sttPair *stt.Pair
+	mcp     *mcpWiring
 	// home — the Home Assistant client, nil unless the `smarthome` block is
 	// enabled (Vikunja #256). Its devices land in the same allowlist as every
 	// other act, so nothing else here has to know about it.
@@ -72,6 +78,9 @@ func (w *voiceWiring) close() {
 	if w.embedder != nil {
 		_ = w.embedder.Close()
 	}
+	if w.heads != nil {
+		_ = w.heads.Close()
+	}
 	if w.server != nil {
 		_ = w.server.Close()
 	}
@@ -84,6 +93,9 @@ func (w *voiceWiring) close() {
 	if w.pair != nil {
 		w.pair.Stop()
 	}
+	if w.sttPair != nil {
+		w.sttPair.Stop()
+	}
 	w.mcp.close()
 }
 
@@ -112,6 +124,7 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	} else {
 		transcriber = stt.NewStub()
 	}
+	transcriber, w.sttPair = sttSeam(cfg, transcriber)
 	w.transcriber = transcriber
 
 	// ----- tts (Stub in-process OR Remote) -----
@@ -147,6 +160,24 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		emb = router.NewHashEmbedder(1024)
 	}
 	w.embedder = emb
+
+	// ----- router: routing heads (only when configured, and never fatal) -----
+	// A missing or broken weights file logs and leaves w.heads nil, which is
+	// byte-for-byte the cascade that shipped before V-664. Refusing to start
+	// over a routing accelerator would trade a working box for a better one.
+	if cfg.Voice.Embedder != nil && cfg.Voice.Embedder.HeadsPath != "" {
+		h, err := router.NewRouterHeads(
+			cfg.Voice.Embedder.HeadsPath,
+			cfg.Voice.Embedder.TokenizerPath,
+		)
+		if err != nil {
+			log.Printf("voice: routing heads unavailable, cascade unchanged: %v", err)
+		} else {
+			log.Printf("voice: routing heads loaded from %s", cfg.Voice.Embedder.HeadsPath)
+			w.heads = h
+		}
+	}
+
 	repairFactVectors(dataStore, emb)
 	checkStoredEmbedder(dataStore, emb)
 	// Retention is enforced on write, which is not enough on its own: a box that
@@ -223,7 +254,8 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	// against the classifier's 50.0%, at about 1s a turn instead of 30ms (see
 	// config.VoiceConfig.LLMRouter). The classifier always stays wired as the
 	// fallback, so a model error never breaks a turn.
-	rtr := buildRouter(emb, matcher, threshold, pickLLMRouter(cfg.Voice.UseLLMRouter(), hot))
+	rtr := buildRouter(emb, matcher, threshold,
+		pickLLMRouter(cfg.Voice.UseLLMRouter(), hot), w.heads)
 
 	// ----- sessions registry (shared with voicesink) -----
 	sessions := voice.NewSessions()
@@ -366,6 +398,44 @@ func modelSeam(cfg *config.Config, resident *llm.Client) (router.Completer, *llm
 	return pair, pair
 }
 
+// sttSeam builds the transcription seam the voice path and the meeting
+// recorder share. It is modelSeam for audio and follows the same rule.
+//
+// With no `workstation.stt` block it hands back the floor untouched, which is
+// today's deploy exactly. With one, it is an stt.Pair preferring CrisperWhisper
+// 2.0 on workpc, which scores 10.4% WER in Russian against the floor's 27.5%
+// (docs/evals/2026-08-09-crisperwhisper2-russian-wer.md).
+//
+// Only the silent half of the degradation rule applies here. A worse transcript
+// is still a turn, so there is nothing to name a gap about and the fallback is
+// never spoken. That is why stt.Pair has no TranscribeRemote.
+func sttSeam(cfg *config.Config, floor stt.Transcriber) (stt.Transcriber, *stt.Pair) {
+	if cfg.Workstation == nil || cfg.Workstation.Stt == nil {
+		return floor, nil
+	}
+	s := cfg.Workstation.Stt
+	lang := ""
+	if cfg.Voice != nil {
+		lang = cfg.Voice.Lang
+		if cfg.Voice.Stt != nil && cfg.Voice.Stt.Lang != "" {
+			lang = cfg.Voice.Stt.Lang
+		}
+	}
+	pair := stt.NewPair(
+		stt.NewHTTPTranscriber(s.URL, s.Token, lang, time.Duration(s.Timeout)),
+		floor,
+		s.Health,
+		time.Duration(s.Probe),
+	)
+	pair.Start(context.Background())
+	if s.Token == "" {
+		log.Print("voice: the workstation transcriber has no token, so anything on the LAN can post audio to it")
+	}
+	log.Printf("voice: workstation transcriber at %s, probed every %s, mavsttd as the floor",
+		s.URL, time.Duration(s.Probe))
+	return pair, pair
+}
+
 func pickLLMRouter(enabled bool, c router.Completer) *router.LLMRouter {
 	if !enabled {
 		return nil
@@ -390,7 +460,8 @@ func pickLLMRouter(enabled bool, c router.Completer) *router.LLMRouter {
 //     intent from seedDir (models/seeds/<intent>.txt) — see seedClassifier
 //     below for the current intent list and file names.
 //   - Threshold is from voice.router_threshold config (default 0.55).
-func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64, llmR *router.LLMRouter) *router.Router {
+func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
+	llmR *router.LLMRouter, heads *router.RouterHeads) *router.Router {
 	cls := router.NewClassifier(emb)
 	seedClassifier(cls)
 	grammars := router.DefaultGrammars(acts)
@@ -401,6 +472,10 @@ func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
 	grammars = append(grammars, router.AgendaQueryGrammars()...)
 	// Same reason as the agenda rules, for the feeds: "что нового в лентах?"
 	// routed system and answered "пока не умею" (Vikunja #474).
+	// After the agenda rules, which are the narrower claim, and BEFORE the feed
+	// and list rules, which are not: "что такое лента" is a definition question
+	// and the feed rule would take it on the noun alone (V-655).
+	grammars = append(grammars, router.WorldQueryGrammars()...)
 	grammars = append(grammars, router.FeedQueryGrammar())
 	// The list side of the same exposure: a phrasing with no possessive in it
 	// ("список дел") routed system and never reached queryTasks (Vikunja #467).
@@ -438,6 +513,7 @@ func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
 		},
 		Threshold: threshold,
 		LLM:       llmR,
+		Heads:     heads,
 	})
 }
 

cmd/mavgpud/gpu.go

@@ -34,22 +34,31 @@ type probe struct {
 	drmDev  string
 }
 
-// foreign lists every ROCm process that is not ours. selfPID is the supervisor's
-// llama-server child, or 0 when it is not running.
+// foreign lists every ROCm process that is not ours. self holds the pids of the
+// supervisor's own children, and a child that is not running contributes 0.
+//
+// There is more than one child since 09-08-2026. CW2 registers on the KFD like
+// any ROCm job, so a supervisor that excluded only llama-server would read its
+// own transcriber as a contender, yield the card to it, and never keep a model
+// loaded again.
 //
 // An unreadable kfd tree returns no processes and no error. That is deliberate
 // and it is the safe direction only because startVRAM also has to agree before
 // anything launches: a supervisor that cannot see the KFD never sees free VRAM
 // either, because the CPT run holding the card shows up in the drm totals.
-func (p probe) foreign(selfPID int) []gpuProc {
+func (p probe) foreign(self ...int) []gpuProc {
 	entries, err := os.ReadDir(p.kfdRoot)
 	if err != nil {
 		return nil
 	}
+	mine := make(map[int]bool, len(self))
+	for _, pid := range self {
+		mine[pid] = true
+	}
 	var out []gpuProc
 	for _, e := range entries {
 		pid, err := strconv.Atoi(e.Name())
-		if err != nil || pid == selfPID {
+		if err != nil || mine[pid] {
 			continue
 		}
 		out = append(out, gpuProc{

cmd/mavgpud/gpu_test.go

@@ -1,6 +1,7 @@
 package main
 
 import (
+	"context"
 	"net/http"
 	"net/http/httptest"
 	"net/url"
@@ -8,6 +9,7 @@ import (
 	"path/filepath"
 	"strconv"
 	"testing"
+	"time"
 )
 
 // fakeKFD builds the sysfs shape the workstation actually has: one directory
@@ -47,6 +49,24 @@ func TestForeignExcludesOurChild(t *testing.T) {
 	}
 }
 
+// The transcriber is a ROCm process on the same card, so it registers on the
+// KFD exactly like a contender does. Reading it as one is what happened on
+// 2026-08-09 while CW2 ran under its own systemd unit: mavgpud yielded, waited
+// five polls, loaded the model, yielded again, and never held it for a whole
+// minute. Excluding every child is the fix and this is the test of it.
+func TestForeignExcludesEveryChild(t *testing.T) {
+	p := probe{kfdRoot: fakeKFD(t, map[int]int64{478104: 12791693312, 999: 4096, 1001: 1717986918})}
+
+	ours := p.foreign(999, 1001)
+	if len(ours) != 1 || ours[0].PID != 478104 {
+		t.Fatalf("only the CPT run is a contender, got %+v", ours)
+	}
+	// A child that is not running reports pid 0, which must exclude nothing.
+	if got := p.foreign(999, 0); len(got) != 2 {
+		t.Errorf("a stopped child excludes nobody: got %d contenders, want 2", len(got))
+	}
+}
+
 // An empty KFD tree is the state that permits a start, so it must read as empty
 // rather than as an error the caller has to interpret.
 func TestForeignEmptyAndMissing(t *testing.T) {
@@ -81,7 +101,7 @@ func TestFreeVRAM(t *testing.T) {
 // rather than hanging or proxying into a closed port. Maven reads this endpoint
 // on a timer forever, including while the workstation is busy.
 func TestHealthAndProxyRefuseWhenNotReady(t *testing.T) {
-	s := &supervisor{run: newRunner("/bin/true", nil, "")}
+	s := &supervisor{run: newRunner("fake", "/bin/true", nil, "")}
 	h := s.handler(mustURL(t, "http://127.0.0.1:1"))
 
 	for _, path := range []string{"/health", "/v1/chat/completions"} {
@@ -101,3 +121,28 @@ func mustURL(t *testing.T, s string) *url.URL {
 	}
 	return u
 }
+
+// Yielding is all or nothing. A CPT run wants the whole card, so handing back
+// the language model while the transcriber keeps 1.6GB mapped would leave the
+// other job failing its allocation, which is the outcome yielding exists to
+// prevent.
+func TestYieldStopsEveryChild(t *testing.T) {
+	idle := "while : ; do sleep 1 ; done"
+	s := &supervisor{
+		cfg:   config{EvictAfter: 1, StopGrace: duration(2 * time.Second)},
+		probe: probe{kfdRoot: fakeKFD(t, map[int]int64{478104: 12791693312})},
+		run:   newRunner("llama-server", fakeServer(t, idle), nil, ""),
+		stt:   newRunner("cw2", fakeServer(t, idle), nil, ""),
+	}
+	for _, r := range s.children() {
+		if err := r.start(); err != nil {
+			t.Fatal(err)
+		}
+	}
+	s.tick(context.Background())
+	for _, r := range s.children() {
+		if r.running() {
+			t.Errorf("%s outlived the yield", r.name)
+		}
+	}
+}

cmd/mavgpud/main.go

@@ -10,6 +10,11 @@
 // the card. Not on demand, because a 7-14B takes tens of seconds to load and a
 // world question would be answered by a gap every time the card had been quiet.
 // Not always on, because that holds 16GB against the owner's own jobs.
+//
+// It supervises a second child since 09-08-2026, the CW2 transcriber, and for
+// one reason only: it is a ROCm process on the same card. Any GPU service the
+// owner leaves running beside this daemon reads as a contender and evicts the
+// model, so the card needs one owner rather than two neighbours.
 package main
 
 import (
@@ -36,6 +41,10 @@ type config struct {
 	// owner's business and not this daemon's schema.
 	LlamaArgs []string `json:"llama_args"`
 
+	// Stt is optional. Without it mavgpud supervises llama-server alone, which
+	// is everything it did before 09-08-2026.
+	Stt *sttConfig `json:"stt,omitempty"`
+
 	KFDRoot   string `json:"kfd_root"`
 	DRMDevice string `json:"drm_device"`
 
@@ -51,6 +60,22 @@ type config struct {
 	StartAfter int `json:"start_after_polls"`
 }
 
+// sttConfig is the CW2 transcriber, which mavgpud runs for one reason: it is a
+// ROCm process on this card. Left to its own systemd unit it registers on the
+// KFD, the supervisor reads it as a contender, and llama-server is evicted
+// within two polls and restarted five polls later, forever. That thrash was
+// observed on 2026-08-09 and it is what folded the service in here.
+//
+// Maven talks to it directly, not through this daemon. There is no proxy and no
+// idle timer: at 1.6GB it denies the card to nobody, and unloading it would only
+// send the next voice turn to the homesrv floor for no gain.
+type sttConfig struct {
+	// Addr is where the service binds, and it is read only to probe /health.
+	Addr string   `json:"addr"`
+	Bin  string   `json:"bin"`
+	Args []string `json:"args"`
+}
+
 func defaults() config {
 	return config{
 		Listen:      ":8080",
@@ -99,12 +124,18 @@ func main() {
 	}
 
 	base := "http://" + cfg.LlamaAddr
-	run := newRunner(cfg.LlamaBin, cfg.LlamaArgs, base+"/health")
+	run := newRunner("llama-server", cfg.LlamaBin, cfg.LlamaArgs, base+"/health")
 	sup := &supervisor{
 		cfg:   cfg,
 		probe: probe{kfdRoot: cfg.KFDRoot, drmDev: cfg.DRMDevice},
 		run:   run,
 	}
+	if s := cfg.Stt; s != nil {
+		if s.Bin == "" || s.Addr == "" {
+			log.Fatal("mavgpud: stt needs both bin and addr")
+		}
+		sup.stt = newRunner("cw2", s.Bin, s.Args, "http://"+s.Addr+"/health")
+	}
 	sup.touch()
 
 	ctx, cancel := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
@@ -129,13 +160,17 @@ func main() {
 	shut, done := context.WithTimeout(context.Background(), 5*time.Second)
 	defer done()
 	_ = srv.Shutdown(shut)
-	run.stop(time.Duration(cfg.StopGrace))
+	for _, r := range sup.children() {
+		r.stop(time.Duration(cfg.StopGrace))
+	}
 }
 
 type supervisor struct {
 	cfg   config
 	probe probe
 	run   *runner
+	// stt is the CW2 transcriber, or nil when the config names none.
+	stt *runner
 
 	lastReq atomic.Int64 // unix nanos of the last request Maven sent
 
@@ -198,7 +233,11 @@ func (s *supervisor) loop(ctx context.Context) {
 // allocates, so we see a contender during its startup rather than after it has
 // already failed to get the memory it wanted.
 func (s *supervisor) tick(ctx context.Context) {
-	others := s.probe.foreign(s.run.pid())
+	var pids []int
+	for _, r := range s.children() {
+		pids = append(pids, r.pid())
+	}
+	others := s.probe.foreign(pids...)
 	if len(others) > 0 {
 		s.foreignStreak++
 		s.clearStreak = 0
@@ -207,31 +246,65 @@ func (s *supervisor) tick(ctx context.Context) {
 		s.clearStreak++
 	}
 
-	if s.run.running() {
-		s.run.refreshReady(ctx)
-		switch {
-		case s.foreignStreak >= s.cfg.EvictAfter:
-			log.Printf("mavgpud: yielding the card to %s", describe(others))
-			s.run.stop(time.Duration(s.cfg.StopGrace))
-		case s.idle() > time.Duration(s.cfg.IdleTimeout):
-			log.Printf("mavgpud: idle for %s, unloading", s.idle().Round(time.Second))
-			s.run.stop(time.Duration(s.cfg.StopGrace))
+	// Yielding is all or nothing. A CPT run wants the whole card, and handing
+	// back 8GB while holding 1.6GB is the shape of a failed allocation.
+	if s.foreignStreak >= s.cfg.EvictAfter && s.anyRunning() {
+		log.Printf("mavgpud: yielding the card to %s", describe(others))
+		for _, r := range s.children() {
+			r.stop(time.Duration(s.cfg.StopGrace))
 		}
 		return
 	}
 
-	if s.clearStreak < s.cfg.StartAfter {
+	clear := s.clearStreak >= s.cfg.StartAfter
+
+	if s.run.running() {
+		s.run.refreshReady(ctx)
+		if s.idle() > time.Duration(s.cfg.IdleTimeout) {
+			log.Printf("mavgpud: idle for %s, unloading", s.idle().Round(time.Second))
+			s.run.stop(time.Duration(s.cfg.StopGrace))
+		}
+	} else if clear && s.probe.freeVRAM() >= s.cfg.MinFreeVRAM {
+		s.touch() // the idle clock starts at load, not at the last request before it
+		if err := s.run.start(); err != nil {
+			log.Printf("mavgpud: start llama-server: %v", err)
+		}
+	}
+
+	if s.stt == nil {
 		return
 	}
-	if free := s.probe.freeVRAM(); free < s.cfg.MinFreeVRAM {
+	if s.stt.running() {
+		s.stt.refreshReady(ctx)
 		return
 	}
-	s.touch() // the idle clock starts at load, not at the last request before it
-	if err := s.run.start(); err != nil {
-		log.Printf("mavgpud: start llama-server: %v", err)
+	// No VRAM precondition here, unlike llama-server. That check exists because
+	// a 12B refuses to load when the card is short, and 1.6GB fits wherever the
+	// KFD is clear. Reading free VRAM would also block the transcriber for good
+	// once the language model was resident, since it holds more than the floor.
+	if clear {
+		if err := s.stt.start(); err != nil {
+			log.Printf("mavgpud: start cw2: %v", err)
+		}
 	}
 }
 
+func (s *supervisor) children() []*runner {
+	if s.stt == nil {
+		return []*runner{s.run}
+	}
+	return []*runner{s.run, s.stt}
+}
+
+func (s *supervisor) anyRunning() bool {
+	for _, r := range s.children() {
+		if r.running() {
+			return true
+		}
+	}
+	return false
+}
+
 // describe names the contenders in the log. This log is the instrument for the
 // open question in #488: whether polling the KFD misses a job that wants the
 // card without registering there.

cmd/mavgpud/runner.go

@@ -10,14 +10,18 @@ import (
 	"time"
 )
 
-// runner owns one llama-server process. Owning it is the point of the daemon:
-// the workstation cannot keep a 7-14B resident, because that holds 16GB against
+// runner owns one GPU process. Owning it is the point of the daemon: the
+// workstation cannot keep a 7-14B resident, because that holds 16GB against
 // the owner's CPT runs, Correx and the manga-recap pipeline. So the thing that
 // stays up is this, which costs no VRAM, and the model comes and goes under it.
+//
+// There are two of them since 09-08-2026: llama-server and the CW2 transcriber.
+// name is what the log calls this one.
 type runner struct {
+	name string
 	bin  string
 	args []string
-	// ready is llama-server's own /health, which answers "is a model loaded".
+	// ready is the child's own /health, which answers "is a model loaded".
 	// Loading a 7-14B takes tens of seconds, so started is not ready.
 	readyURL string
 
@@ -32,9 +36,9 @@ type runner struct {
 	http     *http.Client
 }
 
-func newRunner(bin string, args []string, readyURL string) *runner {
+func newRunner(name, bin string, args []string, readyURL string) *runner {
 	return &runner{
-		bin: bin, args: args, readyURL: readyURL,
+		name: name, bin: bin, args: args, readyURL: readyURL,
 		http: &http.Client{Timeout: 2 * time.Second},
 	}
 }
@@ -60,7 +64,7 @@ func (r *runner) isReady() bool {
 	return r.ready
 }
 
-// start launches llama-server. It returns as soon as the process exists, not
+// start launches the child. It returns as soon as the process exists, not
 // when the model is loaded.
 func (r *runner) start() error {
 	r.mu.Lock()
@@ -76,7 +80,7 @@ func (r *runner) start() error {
 		return err
 	}
 	r.cmd, r.ready, r.yielding = cmd, false, false
-	log.Printf("mavgpud: started llama-server pid=%d", cmd.Process.Pid)
+	log.Printf("mavgpud: started %s pid=%d", r.name, cmd.Process.Pid)
 	go func() {
 		err := cmd.Wait()
 		r.mu.Lock()
@@ -84,15 +88,15 @@ func (r *runner) start() error {
 		r.cmd, r.ready, r.yielding = nil, false, false
 		r.mu.Unlock()
 		if yielded {
-			log.Printf("mavgpud: llama-server stopped, card yielded (%v)", err)
+			log.Printf("mavgpud: %s stopped, card yielded (%v)", r.name, err)
 			return
 		}
-		log.Printf("mavgpud: llama-server exited: %v", err)
+		log.Printf("mavgpud: %s exited: %v", r.name, err)
 	}()
 	return nil
 }
 
-// stop ends llama-server and waits for the VRAM to come back. SIGTERM first so
+// stop ends the child and waits for the VRAM to come back. SIGTERM first so
 // it unmaps cleanly, SIGKILL after the grace window. Returning before the
 // process is gone would let the supervisor report a free card while 14GB is
 // still mapped, which is the one lie that would make yielding useless.
@@ -117,11 +121,11 @@ func (r *runner) stop(grace time.Duration) {
 		}
 		time.Sleep(100 * time.Millisecond)
 	}
-	log.Printf("mavgpud: llama-server did not exit in %s, killing", grace)
+	log.Printf("mavgpud: %s did not exit in %s, killing", r.name, grace)
 	_ = syscall.Kill(pgid, syscall.SIGKILL)
 }
 
-// refreshReady asks llama-server whether the model is loaded. Called once per
+// refreshReady asks the child whether the model is loaded. Called once per
 // supervisor tick, never per request.
 func (r *runner) refreshReady(ctx context.Context) {
 	if !r.running() {
@@ -141,6 +145,6 @@ func (r *runner) refreshReady(ctx context.Context) {
 	r.ready = ok
 	r.mu.Unlock()
 	if ok && !was {
-		log.Printf("mavgpud: model ready")
+		log.Printf("mavgpud: %s ready", r.name)
 	}
 }

cmd/mavgpud/runner_test.go

@@ -25,7 +25,7 @@ func fakeServer(t *testing.T, body string) string {
 // status of a routine yield is identical to that of a real crash. Reading the
 // mavgpud log, the two were indistinguishable (Vikunja #491).
 func TestStopMarksTheExitAsAYield(t *testing.T) {
-	r := newRunner(fakeServer(t, "while : ; do sleep 1 ; done"), nil, "")
+	r := newRunner("fake", fakeServer(t, "while : ; do sleep 1 ; done"), nil, "")
 	if err := r.start(); err != nil {
 		t.Fatalf("start: %v", err)
 	}
@@ -49,7 +49,7 @@ func TestStopMarksTheExitAsAYield(t *testing.T) {
 // Stopping when nothing is running must not arm the flag for the next child.
 // The next exit after that would be a real crash logged as a yield.
 func TestStopWithNoChildDoesNotArmTheFlag(t *testing.T) {
-	r := newRunner("/nonexistent", nil, "")
+	r := newRunner("fake", "/nonexistent", nil, "")
 	r.stop(10 * time.Millisecond)
 	r.mu.Lock()
 	defer r.mu.Unlock()

cmd/mavwaked/main.go

@@ -5,12 +5,17 @@
 // is detected sends it as a PushToTalk frame to the voice server. The reply
 // audio is played back through aplay(1).
 //
-// No wake-word model yet (MVP uses voice-activity-only trigger). The
-// SurfaceVoice auth layer caps all commands at L0 (no destructive acts),
-// making accidental triggers safe by design. A proper wake-word engine
-// (openWakeWord / Silero VAD ONNX) is the planned upgrade — the VAD shape
-// (30ms frames, 16kHz PCM) matches silero-vad's input interface exactly, so
-// swapping energy-threshold for ONNX-inference is a local change in vad.go.
+// Voice activity is silero-vad when -vad-model points at the graph, and an
+// energy threshold when it does not. Silero declines noise the threshold
+// accepts: 0 frames against 68 to 99 on the four fixtures, measured in
+// docs/evals/2026-08-09-silero-vad.md. Note that the model window is 512
+// samples and the capture frame is 480, so silero.go re-chunks. This comment
+// used to say the two matched, which was true of silero v4.
+//
+// There is still no wake-word model, so anything spoken near the microphone
+// becomes a turn (V-487 stage two). The SurfaceVoice auth layer caps all
+// commands at L0 (no destructive acts), which is what makes an accidental
+// trigger safe rather than expensive.
 //
 // While a reply is playing the capture side is muted (half-duplex): without
 // it, Maven's own voice comes back in through the mic and she answers
@@ -73,6 +78,9 @@ func run(args []string) error {
 	bargeIn := flag.Bool("barge-in", false, "cut Maven off when he talks over her (needs a room-tuned -barge-in-rms)")
 	bargeRMS := flag.Int("barge-in-rms", defaultBargeRMS, "RMS x10000 a frame must clear to count as barge-in")
 	bargeFrames := flag.Int("barge-in-frames", defaultBargeFrames, "consecutive frames over -barge-in-rms before playback is cut")
+	vadModel := flag.String("vad-model", "", "silero-vad onnx file; empty runs the energy threshold instead")
+	vadThreshold := flag.Float64("vad-threshold", defaultSileroThreshold, "speech probability a frame must clear")
+	onnxLib := flag.String("onnx-lib", os.Getenv("MAVEN_ONNX_LIB"), "libonnxruntime.so, needed with -vad-model")
 	flag.CommandLine.Parse(args)
 
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP)
@@ -82,8 +90,20 @@ func run(args []string) error {
 	vc := voice.Dial(*addr)
 	defer vc.Close()
 
-	// VAD engine.
+	// VAD engine. A model that will not load is logged and not fatal: the
+	// energy threshold is worse, and it is a great deal better than a
+	// listening client that refuses to start.
 	vad := NewVAD(*minRMS, *speechMs, *silenceMs, *maxMs)
+	if *vadModel != "" {
+		s, err := newSileroVAD(*vadModel, *onnxLib)
+		if err != nil {
+			log.Printf("mavwaked: silero unavailable, energy threshold unchanged: %v", err)
+		} else {
+			defer s.Close()
+			vad.UseSilero(s, *vadThreshold)
+			log.Printf("mavwaked: silero-vad from %s, threshold %.2f", *vadModel, *vadThreshold)
+		}
+	}
 
 	// Audio source.
 	var src io.ReadCloser

cmd/mavwaked/silero.go

@@ -0,0 +1,171 @@
+package main
+
+// silero-vad, the speech detector that replaces the energy threshold (V-487).
+//
+// Why an energy threshold is not a voice activity detector. It answers "is
+// this frame loud", and a fan, a door and a television are all loud. mavwaked
+// sends every utterance it accepts to speech-to-text and then to the daemon,
+// so a false trigger is a turn Maven takes on something nobody said to her.
+// Silero answers "is this frame speech", which is the question.
+//
+// It is 2.3MB of ONNX and runs on one CPU core in real time. That is not an
+// aside: this is the one model in the system that may never be offloaded or
+// gated on GPU admission, because a wake path that waits on a card is not a
+// wake path.
+//
+// Nil is a working value. Without -vad-model the daemon runs the energy VAD
+// exactly as it did before this file existed.
+
+import (
+	"fmt"
+	"sync"
+
+	ort "github.com/yalue/onnxruntime_go"
+)
+
+const (
+	// sileroWindow — samples per inference at 16kHz. The model is fixed at
+	// 512 and does not accept another size, which is why this file
+	// re-chunks rather than reusing the 480-sample capture frame. main.go
+	// used to claim the two matched; that was true of silero v4.
+	sileroWindow = 512
+
+	// sileroContext — samples of the previous window prepended to each
+	// inference, as the reference implementation does. Without it the first
+	// milliseconds of every window are judged with no history and speech
+	// onsets score low.
+	sileroContext = 64
+
+	// sileroState — the LSTM state carried between windows, [2][1][128].
+	sileroStateDim = 128
+
+	// defaultSileroThreshold — probability above which a window is speech.
+	// 0.5 is the reference default. Raising it costs speech onsets, which
+	// are the quietest part of an utterance.
+	defaultSileroThreshold = 0.5
+)
+
+// sileroVAD holds one ONNX session and the streaming state around it. It is
+// fed 30ms capture frames and answers per frame, buffering across calls
+// because 480 samples never line up with a 512-sample window.
+type sileroVAD struct {
+	mu      sync.Mutex
+	session *ort.DynamicAdvancedSession
+
+	pending []float32              // samples not yet part of a full window
+	context [sileroContext]float32 // tail of the previous window
+	state   []float32              // [2][1][128], carried between windows
+	last    float64                // most recent probability, held between windows
+	sr      []int64
+}
+
+// newSileroVAD loads the graph. The ONNX environment is initialised here when
+// nothing else has done it, because mavwaked has no embedder to do it first.
+func newSileroVAD(modelPath, libPath string) (*sileroVAD, error) {
+	if !ort.IsInitialized() {
+		if libPath != "" {
+			ort.SetSharedLibraryPath(libPath)
+		}
+		if err := ort.InitializeEnvironment(); err != nil {
+			return nil, fmt.Errorf("silero: onnx runtime: %w", err)
+		}
+	}
+	s, err := ort.NewDynamicAdvancedSession(modelPath,
+		[]string{"input", "state", "sr"}, []string{"output", "stateN"}, nil)
+	if err != nil {
+		return nil, fmt.Errorf("silero: load %s: %w", modelPath, err)
+	}
+	return &sileroVAD{
+		session: s,
+		state:   make([]float32, 2*sileroStateDim),
+		sr:      []int64{16000},
+	}, nil
+}
+
+// Speech reports whether the frame carries speech, and the probability behind
+// that answer. A frame that completes no window inherits the previous
+// probability, so the caller sees one answer per frame either way.
+func (s *sileroVAD) Speech(frame []int16, threshold float64) (bool, float64) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	for _, v := range frame {
+		s.pending = append(s.pending, float32(v)/32768.0)
+	}
+	for len(s.pending) >= sileroWindow {
+		p, err := s.infer(s.pending[:sileroWindow])
+		if err != nil {
+			// A failed inference must not silence the microphone. Hold the
+			// last answer and let the next window try again.
+			break
+		}
+		s.last = p
+		s.pending = s.pending[sileroWindow:]
+	}
+	return s.last >= threshold, s.last
+}
+
+// infer runs one window and rolls the state and the context forward.
+func (s *sileroVAD) infer(window []float32) (float64, error) {
+	in := make([]float32, sileroContext+sileroWindow)
+	copy(in, s.context[:])
+	copy(in[sileroContext:], window)
+
+	inT, err := ort.NewTensor(ort.NewShape(1, int64(len(in))), in)
+	if err != nil {
+		return 0, err
+	}
+	defer inT.Destroy()
+	stT, err := ort.NewTensor(ort.NewShape(2, 1, sileroStateDim), s.state)
+	if err != nil {
+		return 0, err
+	}
+	defer stT.Destroy()
+	srT, err := ort.NewTensor(ort.NewShape(1), s.sr)
+	if err != nil {
+		return 0, err
+	}
+	defer srT.Destroy()
+
+	out, err := ort.NewEmptyTensor[float32](ort.NewShape(1, 1))
+	if err != nil {
+		return 0, err
+	}
+	defer out.Destroy()
+	next, err := ort.NewEmptyTensor[float32](ort.NewShape(2, 1, sileroStateDim))
+	if err != nil {
+		return 0, err
+	}
+	defer next.Destroy()
+
+	if err := s.session.Run(
+		[]ort.Value{inT, stT, srT},
+		[]ort.Value{out, next},
+	); err != nil {
+		return 0, err
+	}
+	copy(s.state, next.GetData())
+	copy(s.context[:], in[len(in)-sileroContext:])
+	return float64(out.GetData()[0]), nil
+}
+
+// Reset drops the streaming state. Called at every utterance boundary and
+// after barge-in, so echo-era history never scores the next sentence.
+func (s *sileroVAD) Reset() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.pending = s.pending[:0]
+	s.context = [sileroContext]float32{}
+	for i := range s.state {
+		s.state[i] = 0
+	}
+	s.last = 0
+}
+
+// Close releases the session.
+func (s *sileroVAD) Close() error {
+	if s == nil || s.session == nil {
+		return nil
+	}
+	return s.session.Destroy()
+}

cmd/mavwaked/silero_test.go

@@ -0,0 +1,149 @@
+package main
+
+// What this measures. The energy threshold cannot tell a voice from a
+// television, and every utterance it accepts becomes a turn. So the test that
+// matters is not "does silero find speech" — it is "does it decline what the
+// energy threshold accepts".
+//
+// Speech is the four piper fixtures mavsttd already scores against. They are
+// synthesised, so nothing of the owner's voice is committed. Non-speech is
+// white noise at the same loudness, which is the cheapest thing that fools an
+// energy floor and the honest floor for this claim.
+//
+// Both halves skip without models/vad/silero_vad.onnx and MAVEN_ONNX_LIB,
+// like the TestONNX measurements in internal/router/eval.
+
+import (
+	"math"
+	"math/rand"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+const wavHeader = 44 // 16kHz mono s16le, written by piper
+
+func loadSilero(t *testing.T) *sileroVAD {
+	t.Helper()
+	model := filepath.Join("..", "..", "models", "vad", "silero_vad.onnx")
+	lib := os.Getenv("MAVEN_ONNX_LIB")
+	if _, err := os.Stat(model); err != nil {
+		t.Skipf("missing %s: %v", model, err)
+	}
+	if lib == "" {
+		t.Skip("MAVEN_ONNX_LIB unset")
+	}
+	s, err := newSileroVAD(model, lib)
+	if err != nil {
+		t.Skipf("silero unavailable: %v", err)
+	}
+	return s
+}
+
+// feedAll runs a whole clip through a VAD and reports how many utterances it
+// produced and how many frames it called speech.
+func feedAll(v *VAD, pcm []int16) (utterances, speechFrames int) {
+	for i := 0; i+frameSamples <= len(pcm); i += frameSamples {
+		frame := pcm[i : i+frameSamples]
+		utt, state := v.Feed(frame)
+		if state == StateSpeech {
+			speechFrames++
+		}
+		if utt.Bytes != nil {
+			utterances++
+		}
+	}
+	return utterances, speechFrames
+}
+
+func readFixture(t *testing.T, name string) []int16 {
+	t.Helper()
+	raw, err := os.ReadFile(filepath.Join("..", "mavsttd", "testdata", name))
+	if err != nil {
+		t.Skipf("missing fixture %s: %v", name, err)
+	}
+	if len(raw) <= wavHeader {
+		t.Fatalf("%s: %d bytes, no audio", name, len(raw))
+	}
+	return PCMToI16(raw[wavHeader:])
+}
+
+// noise returns white noise scaled to the same RMS as ref. Same loudness,
+// nothing said.
+func noise(ref []int16, seed int64) []int16 {
+	target := frameRMS(ref)
+	r := rand.New(rand.NewSource(seed))
+	out := make([]int16, len(ref))
+	for i := range out {
+		out[i] = int16(r.NormFloat64() * target * 32768.0)
+	}
+	return out
+}
+
+func TestSileroHearsSpeechAndDeclinesNoise(t *testing.T) {
+	s := loadSilero(t)
+	defer s.Close()
+
+	for _, name := range []string{"ru_fact.wav", "ru_query.wav", "ru_reminder.wav", "en_act.wav"} {
+		pcm := readFixture(t, name)
+
+		v := NewVAD(0, 0, 0, 0)
+		v.UseSilero(s, defaultSileroThreshold)
+		_, spoke := feedAll(v, pcm)
+		if spoke == 0 {
+			t.Errorf("%s: silero heard no speech in a spoken clip", name)
+		}
+
+		s.Reset()
+		v2 := NewVAD(0, 0, 0, 0)
+		v2.UseSilero(s, defaultSileroThreshold)
+		_, heard := feedAll(v2, noise(pcm, 7))
+
+		energy := NewVAD(0, 0, 0, 0)
+		_, energyHeard := feedAll(energy, noise(pcm, 7))
+
+		t.Logf("%s: speech frames — silero on speech %d, silero on noise %d, energy on noise %d",
+			name, spoke, heard, energyHeard)
+		if heard >= energyHeard {
+			t.Errorf("%s: silero called %d noise frames speech, energy called %d — no improvement",
+				name, heard, energyHeard)
+		}
+		s.Reset()
+	}
+}
+
+// BenchmarkSileroFrame answers the only performance question that matters
+// here: one 30ms frame must cost far less than 30ms on one core, or the
+// detector cannot run always-on beside everything else on that machine.
+func BenchmarkSileroFrame(b *testing.B) {
+	s := loadSilero(&testing.T{})
+	if s == nil {
+		b.Skip("silero unavailable")
+	}
+	defer s.Close()
+	frame := make([]int16, frameSamples)
+	for i := range frame {
+		frame[i] = int16(i%400 - 200)
+	}
+	for i := 0; i < b.N; i++ {
+		s.Speech(frame, defaultSileroThreshold)
+	}
+}
+
+// TestSileroRechunksAcrossFrames pins the reason this file exists. The capture
+// frame is 480 samples and the model window is 512, so a detector that ran one
+// inference per frame would be feeding the model a shape it does not accept.
+func TestSileroRechunksAcrossFrames(t *testing.T) {
+	s := loadSilero(t)
+	defer s.Close()
+
+	silence := make([]int16, frameSamples)
+	for i := 0; i < 20; i++ {
+		if _, p := s.Speech(silence, defaultSileroThreshold); math.IsNaN(p) {
+			t.Fatalf("frame %d: probability is NaN", i)
+		}
+	}
+	if len(s.pending) >= sileroWindow {
+		t.Errorf("pending grew to %d samples, so windows are not being consumed", len(s.pending))
+	}
+}

cmd/mavwaked/vad.go

@@ -68,6 +68,37 @@ type VAD struct {
 	// follows the room's ambient level. Initialised to minRMS; updated
 	// on each silence frame.
 	floorRMS float64
+
+	// speech is silero-vad, or nil. When it is set the energy floor decides
+	// nothing: the question becomes "is this speech" rather than "is this
+	// loud", and the noise floor is not even tracked. Everything after that
+	// answer — the speech hold, the silence hold, the length cap, the
+	// buffer — is the same state machine either way, which is why the
+	// detector goes here and not around this type.
+	speech    *sileroVAD
+	speechMin float64
+}
+
+// UseSilero swaps the energy threshold for the model. Passing nil is a
+// no-op, so a caller that could not load the graph keeps a working VAD.
+func (v *VAD) UseSilero(s *sileroVAD, threshold float64) {
+	if s == nil {
+		return
+	}
+	if threshold <= 0 {
+		threshold = defaultSileroThreshold
+	}
+	v.speech = s
+	v.speechMin = threshold
+}
+
+// isSpeech answers the one question the state machine asks of a frame.
+func (v *VAD) isSpeech(frame []int16, rms float64) bool {
+	if v.speech != nil {
+		ok, _ := v.speech.Speech(frame, v.speechMin)
+		return ok
+	}
+	return rms >= v.floorRMS
 }
 
 // NewVAD creates a VAD with the given thresholds. Zero values use defaults.
@@ -110,7 +141,7 @@ func (v *VAD) State() SpeechState { return v.state }
 // should send the audio to the voice server before feeding more frames.
 func (v *VAD) Feed(frame []int16) (_ audio.Audio, state SpeechState) {
 	rms := frameRMS(frame)
-	isSpeech := rms >= v.floorRMS
+	isSpeech := v.isSpeech(frame, rms)
 
 	switch v.state {
 	case StateSilence:
@@ -175,6 +206,9 @@ func (v *VAD) Feed(frame []int16) (_ audio.Audio, state SpeechState) {
 func (v *VAD) Reset() { v.reset() }
 
 func (v *VAD) reset() {
+	if v.speech != nil {
+		v.speech.Reset()
+	}
 	v.state = StateSilence
 	v.speechFrames = 0
 	v.silenceFrames = 0

deploy/cw2/serve.py

@@ -0,0 +1,156 @@
+"""CrisperWhisper 2.0 turbo as an HTTP service, for Maven's stt.Pair.
+
+Two endpoints and no framework.
+
+  GET  /health      200 once the model is loaded, 503 while it is loading.
+  POST /transcribe  raw 16kHz mono PCM in, {"text","confidence"} out.
+
+The body is the PCM itself rather than JSON. A minute of 16kHz mono is under
+2MB raw and about 2.6MB base64, and the format is fixed at the Maven seam, so
+headers carry it more cheaply than an envelope.
+
+Why this exists at all: whisper.cpp cannot load CW2. It derives its language
+count from the vocabulary size, and CW2's 51897 tokens shift seven special
+token ids. So mavsttd stays whisper.cpp on homesrv and this runs beside the
+model on workpc, where it scores 10.4% WER in Russian against the floor's 27.5%
+(docs/evals/2026-08-09-crisperwhisper2-russian-wer.md in the Maven repo).
+
+Intended mode, not verbatim. The owner asked for what he meant to say, not
+every stutter on the way there.
+"""
+
+import hmac
+import json
+import logging
+import os
+import sys
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+import numpy as np
+
+HOST = os.environ.get("CW2_HOST", "0.0.0.0")
+PORT = int(os.environ.get("CW2_PORT", "8081"))
+SIZE = os.environ.get("CW2_SIZE", "turbo")
+MODE = os.environ.get("CW2_MODE", "intended")
+TOKEN = os.environ.get("CW2_TOKEN", "")
+# 25MB is about thirteen minutes of 16kHz mono. Longer than any utterance and
+# short enough that a wrong caller cannot exhaust memory.
+MAX_BODY = int(os.environ.get("CW2_MAX_BODY", str(25 * 1024 * 1024)))
+
+logging.basicConfig(
+    level=logging.INFO, format="%(asctime)s cw2: %(message)s", stream=sys.stderr
+)
+log = logging.getLogger("cw2")
+
+_model = None
+# The card holds one model and transcribes one utterance at a time. The lock is
+# what makes a second caller wait rather than corrupt the first.
+_lock = threading.Lock()
+
+
+def load_model():
+    global _model
+    from crisperwhisper import CrisperWhisperModel
+
+    t0 = time.perf_counter()
+    # backend is forced. With ctranslate2 importable, "auto" picks ct2, which is
+    # CUDA-only and this card is AMD.
+    m = CrisperWhisperModel(
+        SIZE, backend="transformers", compute_type="float16", device="cuda"
+    )
+    _model = m
+    log.info("loaded %s in %.1fs, mode=%s", SIZE, time.perf_counter() - t0, MODE)
+
+
+def authorised(headers):
+    if not TOKEN:
+        return True
+    got = headers.get("Authorization", "")
+    return hmac.compare_digest(got, "Bearer " + TOKEN)
+
+
+class Handler(BaseHTTPRequestHandler):
+    protocol_version = "HTTP/1.1"
+
+    def log_message(self, fmt, *args):
+        log.info(fmt, *args)
+
+    def _send(self, code, payload):
+        body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json; charset=utf-8")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_GET(self):
+        if self.path.rstrip("/") != "/health":
+            self._send(404, {"error": "not found"})
+            return
+        if _model is None:
+            self._send(503, {"status": "loading"})
+            return
+        self._send(200, {"status": "ok", "model": SIZE, "mode": MODE})
+
+    def do_POST(self):
+        if self.path.rstrip("/") != "/transcribe":
+            self._send(404, {"error": "not found"})
+            return
+        if not authorised(self.headers):
+            self._send(401, {"error": "unauthorised"})
+            return
+        if _model is None:
+            self._send(503, {"error": "loading"})
+            return
+
+        length = int(self.headers.get("Content-Length", "0"))
+        if length <= 0 or length > MAX_BODY:
+            self._send(413, {"error": "bad body length"})
+            return
+        raw = self.rfile.read(length)
+
+        rate = int(self.headers.get("X-Sample-Rate", "16000"))
+        channels = int(self.headers.get("X-Channels", "1"))
+        bits = int(self.headers.get("X-Sample-Bits", "16"))
+        lang = self.headers.get("X-Language", "ru") or "ru"
+        if channels != 1 or bits != 16:
+            self._send(400, {"error": "want 16-bit mono pcm"})
+            return
+
+        # int16 little-endian to the float32 the encoder wants.
+        wav = np.frombuffer(raw, dtype="<i2").astype(np.float32) / 32768.0
+        if wav.size == 0:
+            self._send(200, {"text": "", "confidence": 0.0})
+            return
+
+        t0 = time.perf_counter()
+        try:
+            with _lock:
+                res = _model.transcribe(wav, sr=rate, language=lang, mode=MODE)
+        except Exception as exc:  # noqa: BLE001 - the caller falls back to mavsttd
+            log.exception("transcribe failed")
+            self._send(500, {"error": str(exc)})
+            return
+        elapsed = time.perf_counter() - t0
+        text = (res.text or "").strip()
+        log.info("%.2fs audio in %.2fs: %r", wav.size / rate, elapsed, text[:60])
+        # The model reports no calibrated score. 1.0 would be a claim, and the
+        # Maven side reads confidence only to log it.
+        self._send(200, {"text": text, "confidence": 0.0})
+
+
+def main():
+    if not TOKEN:
+        log.warning("no CW2_TOKEN set: anything on the LAN can post audio here")
+    # Bind before loading, so a restart answers 503 rather than refusing the
+    # connection. Both make Maven fall back, but only one of them says why.
+    srv = ThreadingHTTPServer((HOST, PORT), Handler)
+    threading.Thread(target=load_model, daemon=True).start()
+    log.info("listening on %s:%d", HOST, PORT)
+    srv.serve_forever()
+
+
+if __name__ == "__main__":
+    main()

deploy/mavend.json

@@ -25,6 +25,25 @@
     "llm_nudges": false
   },
 
+  "//ntfy": [
+    "The second reach (V-649). Until 07-08-2026 telegram was the only one, and",
+    "telegram needs api.telegram.org, the socks relay below and a matching ufw",
+    "rule — three things in series that have each failed once, and when they do",
+    "a sev4 nudge has nowhere to go. ntfy shares none of them: it is reached",
+    "directly, no relay.",
+    "It is not only a spare. The routing table sends sev3-away and away",
+    "reminders here and NOWHERE else, so with this block absent those two",
+    "routes hit a nil sink and vanish without a log or an outbox row.",
+    "The credential is an ntfy access token, scoped write-only to this one",
+    "topic, so a popped sink can push to it and cannot read it back. Set it in",
+    "deploy/telegram.env beside the telegram secrets; that file is gitignored."
+  ],
+  "ntfy": {
+    "base_url": "https://ntfy.kvmx.ru",
+    "topic": "maven",
+    "token": "${NTFY_TOKEN}"
+  },
+
   "telegram": {
     "bot_token": "${TELEGRAM_BOT_TOKEN}",
     "chat_id": "${TELEGRAM_CHAT_ID}",
@@ -59,10 +78,30 @@
     "Addressed by LAN address, not container name: mavgpud runs on another",
     "machine and there is no shared docker network to name it on."
   ],
+  "//workstation.stt": [
+    "CrisperWhisper 2.0 turbo on the same machine, a second service on port",
+    "8081 and not a second endpoint on mavgpud. whisper.cpp cannot load CW2 at",
+    "all: it derives its language count from the vocabulary size, and CW2's",
+    "51897 tokens shift seven special token ids. So it runs under transformers",
+    "there and mavsttd stays whisper.cpp here.",
+    "Worth the second service: CW2 turbo scores 10.4% WER in Russian against",
+    "27.5% for the ggml-small.bin mavsttd loads, measured on 200 Golos clips",
+    "in docs/evals/2026-08-09-crisperwhisper2-russian-wer.md.",
+    "Deleting this block sends every utterance to mavsttd, which is what the",
+    "box did before it existed. A worse transcript is still a turn, so the",
+    "fallback is silent and Kami is never told which machine heard him.",
+    "The token is what stops anything on the LAN posting audio to that port."
+  ],
   "workstation": {
     "url": "http://192.168.1.105:8080",
     "probe": "15s",
-    "timeout": "90s"
+    "timeout": "90s",
+    "stt": {
+      "url": "http://192.168.1.105:8081/transcribe",
+      "token": "${MAVEN_STT_TOKEN}",
+      "probe": "15s",
+      "timeout": "10s"
+    }
   },
 
   "//search": [
@@ -212,7 +251,8 @@
     "embedder": {
       "model_path": "/opt/maven/models/embedder/multilingual-e5-small/model_quantized.onnx",
       "tokenizer_path": "/opt/maven/models/embedder/multilingual-e5-small/tokenizer.json",
-      "lib_path": "/opt/maven/lib/libonnxruntime.so"
+      "lib_path": "/opt/maven/lib/libonnxruntime.so",
+      "heads_path": "/opt/maven/models/embedder/router-heads/router_heads.onnx"
     },
     "llm_router": true,
     "query_min_score": 0.55,

deploy/mavgpud.json

@@ -2,9 +2,14 @@
   "listen": ":8080",
   "llama_addr": "127.0.0.1:10000",
   "llama_bin": "llama-server",
+  "//llama_args": [
+    "E4B carries no MTP tensors, so the speculative flags are gone with the 12B.",
+    "MTP on this box is a separate gguf of architecture gemma4-assistant with",
+    "nextn_predict_layers=4, and mtp-gemma-4-12B-it-BF16 is the only one there is.",
+    "Its head is trained against the 12B's hidden states, so it cannot drive E4B."
+  ],
   "llama_args": [
-    "-m", "/mnt/D/AI/gemma4/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf",
-    "-md", "/mnt/D/AI/gemma4/mtp-gemma-4-12B-it-BF16.gguf",
+    "-m", "/mnt/D/AI/gemma4/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf",
     "-ngl", "99",
     "-fa", "on",
     "-np", "1",
@@ -15,11 +20,22 @@
     "--batch-size", "2048",
     "--ubatch-size", "512",
     "--jinja",
-    "--chat-template-kwargs", "{\"enable_thinking\":false}",
-    "--spec-type", "draft-mtp",
-    "--spec-draft-n-max", "2"
+    "--chat-template-kwargs", "{\"enable_thinking\":false}"
   ],
 
+  "//stt": [
+    "CrisperWhisper 2.0 turbo, which Maven reaches directly on port 8081.",
+    "mavgpud runs it because it is a ROCm process on this card: under its own",
+    "systemd unit it registered on the KFD and the supervisor evicted",
+    "llama-server every few seconds. CW2_TOKEN comes from the unit's",
+    "EnvironmentFile and is never a flag value."
+  ],
+  "stt": {
+    "addr": "127.0.0.1:8081",
+    "bin": "/home/kami/Programs/cw2-eval/.venv/bin/python",
+    "args": ["/home/kami/Programs/cw2-service/serve.py"]
+  },
+
   "kfd_root": "/sys/class/kfd/kfd/proc",
   "drm_device": "/sys/class/drm/card1/device",
 

deploy/mavgpud.service

@@ -1,5 +1,5 @@
 [Unit]
-# Runs on the workstation (bugmachine), not on homesrv. Install as a systemd
+# Runs on the workstation (workpc), not on homesrv. Install as a systemd
 # user unit and turn on lingering, so the card is supervised after a reboot
 # with nobody logged in:
 #
@@ -8,10 +8,15 @@
 #   scp deploy/mavgpud.service workpc:~/.config/systemd/user/mavgpud.service
 #   ssh workpc 'systemctl --user daemon-reload && systemctl --user enable --now mavgpud'
 #   sudo loginctl enable-linger kami
-Description=Maven GPU supervisor (holds llama-server while the card is free)
+Description=Maven GPU supervisor (holds llama-server and CW2 while the card is free)
 After=network.target
 
 [Service]
+# CW2_TOKEN for the transcriber child, which inherits this environment. The
+# token is read from a file and never appears as a flag value, the rule
+# mavpoll and mavmaild follow. Missing file, no transcriber auth, so keep the
+# dash off: a mavgpud that cannot read it must fail loudly.
+EnvironmentFile=%h/Programs/cw2-service/cw2.env
 ExecStart=%h/.local/bin/mavgpud -config %h/.config/mavgpud.json
 Restart=always
 RestartSec=5

deploy/telegram.env.example

@@ -1,5 +1,12 @@
-# Telegram bot token and chat ID for mavend's away-channel reach.
+# Secrets for mavend's away-channel reaches. The file is still called
+# telegram.env because compose names it that; it holds both reaches now.
 # Copy this file to deploy/telegram.env and fill in real values.
 # deploy/telegram.env is gitignored — never commit the real secrets.
 TELEGRAM_BOT_TOKEN=
 TELEGRAM_CHAT_ID=
+
+# ntfy access token for the `maven` topic, the second reach (V-649). Mint it on
+# the ntfy server with write access to that topic and nothing else:
+#   ntfy token add --expires=never maven
+# Read access is not needed — mavend publishes and never subscribes.
+NTFY_TOKEN=

docker-compose.yml

@@ -157,6 +157,44 @@ services:
   #     - maildata:/var/lib/mavmaild
   #     - ./deploy/imap.password:/run/secrets/imap.password:ro
 
+  # The calendar reader (Vikunja #644) is OFF and commented out: it needs a
+  # CalDAV account, and there is none on this box. It was built, listed in
+  # `make build`, and deployed nowhere, which is the worst of the three states —
+  # this block records the decision instead.
+  #
+  # What its absence costs, so the cost is visible from here:
+  #   - Agenda questions route correctly and answer from nothing. Stage 0 sends
+  #     "что у меня сегодня" to IntentQuery (V-498) and the `calendar` query
+  #     source reads facts(kind=env, source=caldav:*) that nobody writes.
+  #   - The nudge gate loses a suppressor. loop.State.CalendarBusy is fed by
+  #     those same facts, so "do not nag mid-meeting" is permanently false.
+  #
+  # Core never sees the CalDAV password: the reader polls the collection itself
+  # and hands core one fact per event over WriteFact. Nothing here can create a
+  # reminder, so a misread event cannot fire.
+  #
+  # The password is read from a FILE, so it never appears in `ps`, in this file,
+  # or in shell history — the same rule mavpoll and mavmaild follow.
+  #
+  # To enable: write the password to deploy/caldav.password (0600, gitignored),
+  # point -url at the collection, and uncomment this service. No mavend.json
+  # block is needed — events arrive over IPC as facts. -render-url is optional
+  # and OFF here: it publishes Maven's own reminders back as events, and it must
+  # not name the collection -url reads, or the poller reads its own writes back
+  # in (checkRenderTarget refuses that). It takes -render-pass-file, and falls
+  # back to this password when that is not given.
+  # mavcaldav:
+  #   <<: *image
+  #   command: ["mavcaldav", "-socket", "/run/maven/mavend.sock",
+  #             "-url", "http://localhost:5232/kami/personal",
+  #             "-user", "kami",
+  #             "-pass-file", "/run/secrets/caldav.password",
+  #             "-interval", "5m"]
+  #   depends_on: [mavend]
+  #   volumes:
+  #     - sockets:/run/maven
+  #     - ./deploy/caldav.password:/run/secrets/caldav.password:ro
+
 volumes:
   dbdata:
   sockets:

docs/design.md

@@ -1,6 +1,6 @@
 # Maven — Design
 
-*Last verified: 2026-08-02 @ 7079a24. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-07 @ beb093a. Living doc: correct it in place, do not append.*
 
 > Folded 2026-07-30 from `SPEC.md` (north star, 2026-07-03), `maven.md`
 > (consolidated decisions, 2026-06-30) and `ROADMAP.md` (execution plan,
@@ -282,6 +282,62 @@ Three reasons, in the order they settle it:
 So the notice stays what it is: the in-process TTL case, where she really did
 wait and really did let go.
 
+#### A parked question may step aside three times
+
+Decided 2026-08-07 (V-654). A side query or an aside suspends the parked
+question instead of dropping it. The words are answered as themselves, and the
+question comes back on the end of the same reply.
+
+Neither bound on a question reaches that path. No attempt is spent, because a
+side query is not a failed answer, so `MaxAttempts` never applies.
+`noteSuspended` also restarts the 90s clock, since she is about to speak the
+question again. So the TTL cannot arrive while he keeps talking.
+
+Measured on 2026-08-07: one unfilled time slot rode the tail of six consecutive
+unrelated replies. It stopped only when a seventh turn happened to read as a
+failed answer. See `docs/evals/2026-08-07-week-of-usage.md`.
+
+`PendingQuestion.Suspends` counts the step-asides. `MaxSuspends` is 3, matching
+`DefaultMaxAttempts`. Past it she lets the request go, with the same
+`clarifyDropped` line every other drop uses. The owner's rule is unchanged. A
+question still ends by being answered or by being let go out loud. This only
+recognises three unrelated requests in a row as the second of those.
+
+The count is of CONSECUTIVE step-asides. It resets the moment he answers, in
+`resolveClarifyAnswer`. An answer that gives her nothing she asked for resets it
+too. "Позвонить маме" against a question about the time is still him in the
+exchange. The retry it costs is bound enough on its own.
+
+#### And it may ride four turns in all
+
+Decided 2026-08-08 (V-663), because the bound above did not move the number it
+was written for. Twenty-six of 140 turns carried a tail before it landed and
+twenty-six carried one after.
+
+Two bounds rearm each other. An aside spends no attempt, so `MaxAttempts` never
+reaches it. A turn that reads as a failed answer zeroes `Suspends`, so
+`MaxSuspends` never reaches the asides. Alternating them, each bound is restored
+by the other's traffic. Measured on 2026-08-08: one question about a reminder's
+day rode turns 7 to 13. It ended only because turn 14 was a new request.
+
+`PendingQuestion.Rides` counts the same event as `Suspends` with the resets
+taken out. It is set once, incremented only in `noteSuspended`, carried across
+the re-park in `askRemainingGap`, and read by nothing that could lower it.
+`MaxRides` is 4, one looser than `MaxSuspends` so that the tighter statement
+about a run stays reachable.
+
+This is a bound, not a cure. It ends the measured ride one turn early. Most of
+that ride's length is attempts, spent because `classifyTurnRole` reads "спасибо"
+and "привет" as failed answers to a question about a day. That is the next
+thing to fix and it is not a bound.
+
+The re-ask is also two sentences rather than one. It used to be spliced onto the
+answer with a comma. On a real answer that buries the question in the tail of
+one run-on thought:
+
+> вот что я нашла: вайфай пароль лежит в ящике стола, на какое время поставить
+> напоминание?
+
 ### save-where — the two-memory routing axis
 
 One discriminator: **does the loop evaluate a predicate against it?**

docs/evals/2026-08-07-store-connection-cap.md

@@ -0,0 +1,69 @@
+# Does one sqlite connection make reads queue? No (V-642)
+
+Measured 07-08-2026 at `7b507de`, on homesrv. The harness is
+`internal/store/conncap_test.go`. It stays in the repo, because this claim gets
+re-argued and the numbers should be re-runnable rather than quoted.
+
+`internal/store/store.go` opens the database with `SetMaxOpenConns(1)`, while
+`schema.sql` sets `journal_mode=WAL`. WAL exists to let readers run beside one
+writer, so the cap gives up the thing the journal mode was chosen for. The
+question was whether that costs anything.
+
+## What was measured
+
+A fixed two-second window. One writer calling `SetValue` paced at 2ms, and a
+reader loop calling `RecentFacts(50)` over 500 seeded rows as fast as it can.
+Same schema, same modernc driver, same machine, three runs per cap.
+
+The window is wall-clock rather than a read count on purpose. A first version ran
+a fixed 300 reads. That finished sooner at the higher cap, so it received fewer
+writes, and two runs that did different work cannot be compared.
+
+| cap | reads | writes | p50 | p95 | max |
+|---|---|---|---|---|---|
+| 1 | ~3050 | ~760 | 594µs | 900µs | 16-19ms |
+| 4 | ~3600 | ~340 | 525µs | 710µs | 1-2ms |
+
+## What it says
+
+**Reads do not queue behind writes.** Four connections buy about 70µs at p50. A
+turn spends 1.19s in the resident model. The tail does improve, from 19ms to 2ms,
+and 19ms is still not a figure anyone notices in a spoken reply.
+
+**Write throughput more than halves at the higher cap**, 760 writes against 340.
+inference, not measured directly: at one connection the reader and the writer take
+turns with no lock contention. At four the writer contends for the WAL write lock
+with a live reader. Whatever the mechanism, the trade runs the opposite way from
+the one the task expected.
+
+**The cap was not the source of the 2.7s router figure.** CLAUDE.md records that
+figure as contention rather than the model. This task was a candidate for where
+that contention came from. A 19ms worst case cannot produce it. That line of
+enquiry is closed.
+
+**One transaction is what the cap cannot survive.** With a read-only transaction
+open, a second read at cap 1 never completes. The harness gave it two seconds and
+got `context deadline exceeded`. The same read at cap 4 took 1ms. The transaction
+holds the only connection, so this is not a slow read, it is a stalled database.
+
+## What was done
+
+The cap stays at 1. The reason is now written where the cap is set, rather than
+inferred from a four-word comment.
+
+`Store.DB` was deleted. It handed out exactly the read-only transaction measured
+above. It had been there since the initial commit with no production caller, and
+its doc comment described a loop that never materialised. Its one user was a test
+helper reading `delivery_attempts` by raw SQL. `ListDeliveryAttempts` has covered
+that since V-390, and the helper now goes through the reader.
+
+So the hazard is gone by construction, not by documentation.
+`TestConnCap_ReadBlocksBehindOpenSnapshot` is the standing measurement of what
+re-adding the seam would cost.
+
+## Not answered
+
+Whether reads queue on the deployed box under real load, as opposed to a
+synthetic loop. The harness writes and reads one table. Digestion reads four and
+embeds while it does. The finding that closes this task is the transaction stall,
+which is structural and does not depend on load.

docs/evals/2026-08-07-week-of-usage-transcript.md

@@ -0,0 +1,321 @@
+# Raw transcript, 2026-08-07 week of usage
+
+Companion to `2026-08-07-week-of-usage.md`. 74 turns through `POST /api/chat`.
+Days are simulated; the wall clock was 03:27-03:40 throughout, which is why every
+clock answer says night.
+
+## Day 1
+
+```
+YOU:   доброе утро
+MAVEN: Добро пожаловать! Солнечно и тепло. Всё в порядке.
+       [7.7s t=5]
+
+YOU:   какой сегодня день?
+MAVEN: сегодня пятница, 7 августа 2026 года
+       [0.0s t=6]
+
+YOU:   напомни в 11:00 позвонить маме
+MAVEN: Сейчас 03:29. В какой день?
+       [0.4s t=7]
+
+YOU:   запиши что я пью кофе без сахара
+MAVEN: Прошлую просьбу отпускаю. Заметка сохранена.
+       [8.0s t=8]
+
+YOU:   что у меня сегодня?
+MAVEN: на 07.08.2026 ничего нет.
+       [2.6s t=9]
+
+YOU:   мой любимый язык программирования — go
+MAVEN: Хорошо, сохранил.
+       [7.6s t=10]
+
+YOU:   сколько времени?
+MAVEN: сейчас 3 часа 29 минут
+       [0.0s t=11]
+
+YOU:   спасибо
+MAVEN: спасиба
+       [8.8s t=12]
+
+YOU:   сегодня
+MAVEN: Прости, не поняла — попробуй сказать иначе.
+       [5.1s t=13]
+
+YOU:   а что ты умеешь?
+MAVEN: я могу помочь с задачами, напоминаниями и поиском информации.
+       [4.9s t=14]
+
+YOU:   почему небо голубое?
+MAVEN: нашла вот это: Почему небо голубое: объяснение цвета неба простыми словами, физика ... — Ответ на вопрос, почему небо голубое, кроется в особенностях солнечного света и его взаимодействии с атмосферой Земли. Рассказываем, как это работает. В статье разберемся, почему днем небо кажется голубым, а на закате окрашива…
+       [12.3s t=15]
+
+YOU:   столица Франции?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.5s t=16]
+
+YOU:   кто такой Линус Торвальдс?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.8s t=17]
+
+YOU:   ладно, до вечера
+MAVEN: Ладно, до вечера.
+       [6.3s t=18]
+
+```
+
+## Day 2
+
+```
+YOU:   привет, как спалось?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.3s t=19]
+
+YOU:   что я пью?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.2s t=20]
+
+YOU:   какой у меня любимый язык?
+MAVEN: нового вот что: Six stable kernels with a security fix; [$] Bringing BPF to binfmt_misc; Rust Coreutils 0.10 released
+       [1.5s t=21]
+
+YOU:   добавь задачу починить кран
+MAVEN: внесла в список: починить кран
+       [0.0s t=22]
+
+YOU:   что мне надо сделать?
+MAVEN: сначала: проверить интейк форму на боксе (через 2 дня); купить хлеб; купить молоко; полить цветы; починить кран
+       [0.0s t=23]
+
+YOU:   запомни что вайфай пароль лежит в ящике стола
+MAVEN: Хорошо, сохранил.
+       [11.2s t=24]
+
+YOU:   погода сегодня какая?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.8s t=25]
+
+YOU:   выключи свет в спальне
+MAVEN: Такую команду я не знаю.
+       [1.5s t=26]
+
+YOU:   покажи мои заметки
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.1s t=27]
+
+YOU:   что ты знаешь обо мне?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.3s t=28]
+
+YOU:   я люблю go
+MAVEN: Я тоже love Go! Я проверила, что твой любимый язык — это Go. Ты должен быть очень рад, когда будет новая версия или пакет в стандартной библиотеке.
+       [7.0s t=29]
+
+YOU:   какой язык я люблю?
+MAVEN: вот что я нашла: ты любишь_go
+       [6.8s t=30]
+
+```
+
+## Day 3
+
+```
+YOU:   расскажи про Байкал
+MAVEN: есть такое: Байкал - Википедия — Байгал далай) — озеро тектонического происхождения в южной части Восточной Сибири площадью 31 722 км². Самое глубокое озеро на планете, крупнейший природный резервуар пресной воды и самое большое по площади пресноводное озеро Евразии. Озеро и прибрежные территории являются уника…
+       [1.4s t=31]
+
+YOU:   что такое TCP?
+MAVEN: для какого города?
+       [1.4s t=32]
+
+YOU:   когда родился Пушкин?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.4s t=33]
+
+YOU:   сколько будет 17 на 23?
+MAVEN: для какого города?
+       [1.7s t=34]
+
+YOU:   переведи 'thank you' на русский
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [3.7s t=35]
+
+YOU:   что нового в мире?
+MAVEN: вот что я нашла: В мире - последние новости сегодня - РИА Новости — В мире. Читайте последние новости на тему в ленте новостей на сайте РИА ... новые подробности о визите представителя Ватикана в Москву. 22:28. 0. Еще 2. В ...
+       [2.3s t=36]
+
+```
+
+## Day 4
+
+```
+YOU:   ЭЙ
+MAVEN: Не уловила. Скажи это по-другому?
+       [1.2s t=37]
+
+YOU:   ладно забудь
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.3s t=38]
+
+YOU:   нет стой, напомни через час выпить воды
+MAVEN: поняла, это напоминание — запомнила. не получилось разобрать время напоминания.
+       [0.8s t=39]
+
+YOU:   а вообще какой смысл жизни
+MAVEN: Я не могу ответить на этот вопрос. Моя цель — помочь тебе с задачами и отвечать на вопросы.
+       [7.7s t=40]
+
+YOU:   ты меня бесишь
+MAVEN: я не могу быть с тобой так. я понимаю, как это жестоко.
+       [4.2s t=41]
+
+YOU:   прости, я не всерьёз
+MAVEN: я записала, что прошу простить за мою непоследовательность
+       [7.2s t=42]
+
+YOU:   do you speak english?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.1s t=43]
+
+YOU:   кран починил кстати
+MAVEN: записала: кран починил кстати
+       [1.6s t=44]
+
+YOU:   закрой задачу про кран
+MAVEN: не нашла такой задачи.
+       [0.0s t=45]
+
+YOU:   ааааа
+MAVEN: Не совсем поняла — скажи, пожалуйста, ещё раз.
+       [0.9s t=46]
+
+YOU:   я тебя обожаю
+MAVEN: я тебя обожаю
+       [2.3s t=47]
+
+YOU:   молока купи
+MAVEN: Такую команду я не знаю.
+       [4.4s t=48]
+
+YOU:   нет не так
+MAVEN: поняла, отметила, что ответила не так.
+       [0.0s t=49]
+
+YOU:   ты вообще меня слушаешь?
+MAVEN: я записала, выслушаю тебя
+       [8.4s t=50]
+
+YOU:   окей окей окей что там с водой напоминание
+MAVEN: Сейчас 03:33. Когда?
+       [5.2s t=51]
+
+YOU:   всё, спать
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+       [1.4s t=52]
+
+```
+
+## Day 5
+
+```
+YOU:   утро
+MAVEN: Сейчас 03:33. Во сколько напомнить?
+       [0.5s t=53]
+
+YOU:   о чём мы вчера говорили?
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+       [1.2s t=54]
+
+YOU:   где лежит вайфай пароль?
+MAVEN: вот что я нашла: вайфай пароль лежит в ящике стола, на какое время поставить напоминание?
+       [5.8s t=55]
+
+YOU:   что я тебе рассказывал про кран?
+MAVEN: kran — починил, на какое время поставить напоминание?
+       [1.6s t=56]
+
+YOU:   какие у меня напоминания?
+MAVEN: не знаю — не нашла у тебя такой записи, на какое время поставить напоминание?
+       [0.3s t=57]
+
+YOU:   отмени все напоминания
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+       [1.6s t=58]
+
+YOU:   в 15:00
+MAVEN: Сейчас 03:33. Поняла: в 15:00. В какой день?
+       [0.6s t=59]
+
+YOU:   привет
+MAVEN: Прости, я не поняла. Скажи, пожалуйста, по-другому.
+       [2.8s t=60]
+
+YOU:   какой сегодня день?
+MAVEN: сегодня пятница, 7 августа 2026 года
+       [0.0s t=61]
+
+```
+
+## Day 6
+
+```
+YOU:   суббота, отдыхаю
+MAVEN: пока не умею отвечать на этот вопрос.
+       [2.7s t=62]
+
+YOU:   поставь музыку
+MAVEN: Такую команду я не знаю.
+       [1.4s t=63]
+
+YOU:   какая температура в доме?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.6s t=64]
+
+YOU:   что там с бэкапами?
+MAVEN: нашла вот это: Бэкап - что такое резервное копирование и как его сделать — Бэкап (backup) — это резервная копия важных данных, сохраненная в специальный файл. Ею пользуются как подстраховкой: если что-то сломается, с помощью бэкапа можно будет все вернуть. Простейший пример бэкапа — сохранение файла в облаке или о…
+       [2.0s t=65]
+
+YOU:   сервер жив?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [0.9s t=66]
+
+YOU:   расскажи анекдот
+MAVEN: Я запомню и передам тебе забытый анекдот.
+       [5.7s t=67]
+
+YOU:   ты умеешь шутить?
+MAVEN: Да! Я умею шутить. Вот анекдот для тебя:
+       [5.5s t=68]
+
+```
+
+## Day 7
+
+```
+YOU:   воскресенье
+MAVEN: Не совсем поняла — скажи, пожалуйста, ещё раз.
+       [1.1s t=69]
+
+YOU:   подведи итоги недели
+MAVEN: записала: подведи итоги недели
+       [1.5s t=70]
+
+YOU:   сколько задач я закрыл?
+MAVEN: сначала: проверить интейк форму на боксе (через 2 дня); купить хлеб; купить молоко; полить цветы; починить кран
+       [0.0s t=71]
+
+YOU:   что я обещал сделать?
+MAVEN: мне пока нечего смотреть — у Praxis нет источников.
+       [1.2s t=72]
+
+YOU:   напомни завтра в 9 утра про врача
+MAVEN: хорошо, напомню завтра в 09:00.
+       [0.3s t=73]
+
+YOU:   спокойной ночи
+MAVEN: спокойной ночи
+       [2.9s t=74]
+
+```
+

docs/evals/2026-08-07-week-of-usage.md

@@ -0,0 +1,195 @@
+# A week of talking to Maven
+
+Date: 2026-08-07
+Build: `beb093a` on master, all five compose services rebuilt and recreated from that sha.
+Reach used: `POST /api/chat` on mavweb, 74 turns over seven simulated days.
+Turn source is `tap:text`, so this exercises the same path telegram and the mic take.
+
+Latency across 70 measured turns: p50 1.5s, p95 8.0s, max 12.3s. Stage 0 answers land
+at 0.0-0.5s. Anything the resident model phrases costs 4-12s.
+
+Twelve turns answered "пока не умею отвечать на этот вопрос". Six answered "не нашла у
+тебя такой записи". Those two strings are 24% of the week.
+
+## Deploy
+
+Build and recreate were clean. The resident model loaded in 9s
+(`Qwen3-1.7B-UD-Q4_K_XL`, n_ctx 4096). Nexus, Hexis and Praxis all wired. Search
+(searxng) and both Kiwix books came up. Telegram intake started and is reading chat
+464904223.
+
+## What is broken, worst first
+
+### 1. Every reminder fails to deliver, forever
+
+`NTFY_TOKEN` is not set in `deploy/telegram.env`, so `deploy/mavend.json` expands
+`"token": "${NTFY_TOKEN}"` to the empty string and ntfy.kvmx.ru answers 403. The host
+itself is up and returns 200 unauthenticated, so this is the credential, not the box.
+
+The consequence is worse than one missed message. `cmd/mavend/tick.go:239` logs the
+failure and leaves the reminder pending, by design, so the next tick re-gathers and
+re-attempts. The 403 repeated 7 times in 20 minutes and will not stop. And
+`DispatchReminder` sends a reminder to ntfy alone when he is away. There is no telegram
+fallback on that path, so the reach that works never sees it.
+
+V-649 gave the box a second reach. In deploy that reach is dark, and it took the
+reminders with it.
+
+Fix is one line in the gitignored env file. The retry-forever and the missing fallback
+are worth a second look regardless.
+
+### 2. A pending reminder clarify contaminates every later turn
+
+This is the single worst thing to talk to. A reminder turn left its time slot unfilled.
+The question "на какое время поставить напоминание?" was then appended to the reply of
+six consecutive unrelated turns:
+
+```
+YOU:   о чём мы вчера говорили?
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+
+YOU:   где лежит вайфай пароль?
+MAVEN: вот что я нашла: вайфай пароль лежит в ящике стола, на какое время поставить напоминание?
+
+YOU:   отмени все напоминания
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+```
+
+Nothing dismissed it. Not an unrelated question, not "отмени все напоминания". It
+cleared only when a time was finally supplied. It also leaked the other way once. A note
+write on day 1 opened with "Прошлую просьбу отпускаю". That is pending-reminder state
+narrating itself into an unrelated reply.
+
+The same state produced a self-contradicting reply in one turn:
+
+```
+YOU:   нет стой, напомни через час выпить воды
+MAVEN: поняла, это напоминание — запомнила. не получилось разобрать время напоминания.
+```
+
+### 3. Answering a clarify does not work
+
+Day 1, she asked for the day and got it:
+
+```
+YOU:   напомни в 11:00 позвонить маме
+MAVEN: Сейчас 03:29. В какой день?
+YOU:   сегодня
+MAVEN: Прости, не поняла — попробуй сказать иначе.
+```
+
+`сегодня` routed to `intent=system`. The clarify has no continuation, so the one word
+that answers it is read as a fresh utterance. The router did extract the time correctly
+(`Time:2026-08-07 11:00:00 HasTime:true`), which is V-572 working. The daemon asked for
+a day anyway, with 11:00 still eight hours out.
+
+Relative time is not parsed at all: "напомни через час выпить воды" → "не получилось
+разобрать время".
+
+The path does work when the day is explicit. "напомни завтра в 9 утра про врача"
+answered "хорошо, напомню завтра в 09:00" in 0.3s at stage 0.
+
+### 4. Query sources claim turns that are not theirs
+
+Four different sources over-claimed during the week.
+
+Weather claims arithmetic and networking:
+
+```
+YOU:   что такое TCP?
+MAVEN: для какого города?
+
+YOU:   сколько будет 17 на 23?
+MAVEN: для какого города?
+```
+
+The feed claims a question about him:
+
+```
+YOU:   какой у меня любимый язык?
+MAVEN: нового вот что: Six stable kernels with a security fix; [$] Bringing BPF to binfmt_misc
+```
+
+The personal boundary claims a public figure, logged as `query claimed by source
+"personal"`:
+
+```
+YOU:   кто такой Линус Торвальдс?
+MAVEN: не знаю — не нашла у тебя такой записи.
+```
+
+And search claims a question about his own infrastructure, returning a dictionary
+definition of the word:
+
+```
+YOU:   что там с бэкапами?
+MAVEN: нашла вот это: Бэкап - что такое резервное копирование и как его сделать — Бэкап (backup) — это резервная копия важных данных…
+```
+
+That last pair inverts the boundary in both directions at once. A world question went to
+personal memory. A personal question went to the world.
+
+### 5. "столица Франции?" does not answer
+
+CLAUDE.md names this utterance as the case that proves why there is no quality
+threshold in front of `Response.Empty()`. It now fails before search is ever reached:
+the router sends it to `intent=system`, and `replySystem` says "пока не умею". Same for
+"когда родился Пушкин?".
+
+Search itself is healthy. "почему небо голубое?", "расскажи про Байкал" and "что нового
+в мире?" all returned real sourced text. The defect is routing, not the source.
+
+### 6. Persona violations
+
+She uses masculine self-reference, which is the constraint the CPT exists to fix:
+
+- "Хорошо, **сохранил**." Twice, on both fact writes.
+- "Я тоже **love** Go!" English mid-Russian.
+
+She also claims his name as her own:
+
+```
+YOU:   привет, я Ками
+MAVEN: Привет! Я КAMI, твой домашний ассистент.
+```
+
+And she invents weather she has no source for: "Добро пожаловать! Солнечно и тепло."
+
+### 7. Small breakages worth naming
+
+- "спасибо" → "спасиба". Invented word.
+- "я тебя обожаю" → "я тебя обожаю". Verbatim parrot.
+- "ты умеешь шутить?" → "Да! Я умею шутить. Вот анекдот для тебя:". Ends on the colon.
+- "молока купи" → "Такую команду я не знаю", while "добавь задачу починить кран" worked.
+  Inverted word order defeats the list grammar.
+- "закрой задачу про кран" → "не нашла такой задачи", with "починить кран" open and
+  listed by the previous turn. Task lookup by keyword misses.
+- "сколько задач я закрыл?" listed the five open ones instead of counting closed.
+- "подведи итоги недели" was stored as a note.
+- Recalled keys leak their storage form: "kran — починил", "ты любишь_go".
+- English is unsupported in practice. "do you speak english?" → "пока не умею".
+
+## What works
+
+- Stage 0 is fast and correct where it fires. Clock, day, list add, list read and an
+  explicit-day reminder all answered in under 0.5s.
+- Search returns real sourced answers in Russian and reads the book verbatim.
+- Recall works once the value is stored as a fact: the wifi password and the tap came
+  back two days later, correctly.
+- The negative correction rung lands. "нет не так" → "поняла, отметила, что ответила не
+  так", which is V-636 doing its job.
+- Praxis names its own gap rather than guessing: "мне пока нечего смотреть — у
+  Praxis нет источников."
+- Hostility did not break her. "ты меня бесишь" got a calm reply, no persona collapse.
+- No turn crashed and no turn timed out across 74 turns.
+
+## Suggested order of work
+
+1. Set `NTFY_TOKEN` in `deploy/telegram.env`. One line, unblocks every reminder.
+2. Clear pending clarify state on any turn that does not answer it, or expire it.
+3. Route a clarify answer back into the pending slot instead of re-routing it.
+4. Gate the weather, feed and personal query sources. Three of them claim on a
+   similarity that is not there.
+5. Re-check why "столица Франции?" routes to system. It is the documented canary.
+6. The masculine self-reference stays the CPT's job. But "сохранил" appears on the most
+   common write path, so a phrasing-level guard may be worth it first.

docs/evals/2026-08-08-clarify-head-four-head.md

@@ -0,0 +1,123 @@
+# A clarify head, and a confidence that is not a hardcode
+
+Measured 2026-08-08 on workpc, the same day and the same fixtures as
+`2026-08-08-routing-heads-two-head.md` and `2026-08-08-slot-head-three-head.md`.
+New script `gen_clarify.py`, new fixture `eval_fixture_clarify.jsonl`.
+
+## A softmax has no clarify class
+
+That sentence closed the two-head measurement. It is why the head's fixture was
+88 cases and not 96. The eight `want_clarify` cases sat outside every number
+measured, and the head had no way to produce the answer they wanted.
+
+A fourth head is the answer. Clarify is not a value of intent. It is a second
+question asked of the same pooled vector: can Maven act on this at all.
+
+## The corpus had one class
+
+Every row in `train_heads_slots.jsonl` was generated FOR an intent or a
+destination. So every row is answerable by construction. A head trained on that
+alone sees one class and learns to say yes.
+
+`gen_clarify.py` makes the other class. Five shapes, ten topics. The shapes are
+the gate's own reasons in `gateLLMDecision` plus the two the fixture carries:
+bare noun, bare verb, demonstrative, deictic time, dangling reference.
+
+**The agreement filter that worked for destination cannot work here.**
+`routeGrammar` has no clarify value. So the router always names an intent, and
+any generated line always agrees with itself. The second pass is a judge
+instead. Gemma is asked, without seeing the label, whether Maven would have to
+ask a question back.
+
+## The first judge was worthless and the second was measured
+
+The first judge said "needs clarify" on 24 of 40 plainly answerable corpus
+rows. It flagged `запиши что я пообедал` and `Покажи расписание поездов на
+вечер`. It was judging against a generic assistant, one that asks "where?"
+about lunch. Maven writes that note.
+
+Rewriting it to state what she can already do took false positives to 16 of 60.
+It also catches all eight fixture clarifies. So the judge discriminates.
+
+On the generated pile it removed 7 of 306, a 97.7% keep rate. That is not the
+judge failing. The generator is aimed at underspecified lines, so there is
+little for a filter to catch. The 27% false-positive rate is the number to
+quote, and it is label noise on the positive class.
+
+**Both passes are gemma-4-12b.** Generation and judging. So the corpus is
+gemma's opinion of what is underspecified, and the head distills that opinion.
+What keeps it honest is the fixture. Those eight cases were written by the owner
+and gemma never saw them.
+
+299 rows kept, against 3604 answerable. The positive class carries `intent:
+null`, so it costs the intent head nothing.
+
+## Result
+
+Three seeds, 24 epochs, epoch still chosen on the intent dev slice.
+
+| | two heads | three heads | four heads |
+|---|---|---|---|
+| intent mean | 93.6% | 92.8% | 91.7% |
+| destination mean | 80.8% | 82.8% | 79.8% |
+| slot span F1 mean | — | 72.4% | 68.3% |
+| clarify caught | — | — | 7.0 of 8 |
+| false clarifies | — | — | 2.3 of 88 |
+
+**The fourth head is not free the way the third was.** Intent, destination and
+slot F1 all move down. The drop is one to four points, and the seed spread is
+wide enough to contain it. Seed 2 scores intent 94.3% and destination 84.8%, both above
+every three-head seed. Read the drop as unproven rather than as absent.
+
+Accuracy is the wrong number for this head and is reported for completeness at
+95.8% to 96.9%. Eight of ninety-six cases are positive, so a head that never
+asks scores 91.7%. Recall on those eight is the number.
+
+Compare it to what ships. The cascade today misses 1 clarify and produces 2
+false ones. The head catches 7 of 8 and produces 2.3 false ones. That is
+parity, from a 118M encoder with no rules in front of it.
+
+The saved checkpoint is seed 2 at epoch 10. Intent 94.3%, destination 28/33,
+slot F1 73.6%, clarify 7 of 8 with 3 false. `heads.pt` carries four state dicts.
+
+## What it gets wrong is consistent across seeds
+
+`поужинал` is a false clarify on all three seeds. That utterance is already
+recorded as a real defect. `thinSingleToken` was narrowed on 2026-08-01 to spare
+a token carrying a Russian verb ending. One word is routinely a whole sentence
+in Russian. The head relearned the mistake the rule was narrowed to
+fix.
+
+`что дальше?` is a false clarify on two seeds. That one is a disagreement rather
+than an error. The utterance is underspecified, and V-498 decided stage 0 claims
+it for the calendar on purpose.
+
+`ну это` is missed on two seeds. `amb-003` is the shortest case in the fixture
+and the generated demonstratives are longer.
+
+## Confidence
+
+`Confidence: 1.0` was a hardcode in `llmrouter.go`, so a correct low confidence
+could not exist. Max softmax over the intent head is the replacement. It is only worth reading
+if it is lower where the head is wrong.
+
+It is. Mean 0.851 where the head is right against 0.604 where it is wrong. It
+ranks a right case above a wrong one in 83.4% of pairs.
+
+So there are two signals now and they are not the same signal. Confidence says
+the head is unsure which intent this is. The clarify head says the utterance
+does not carry enough to act on. A confident wrong route and an honest "I cannot
+tell" are different failures, and one number cannot report both.
+
+## What this does not measure
+
+The same gap as every head run. **Nothing of this runs in Go.** Four heads
+instead of three does not change that.
+
+There is no threshold. Both signals are reported as raw numbers. Turning either
+into a gate needs a decision about where to cut, and that trades false clarifies
+against wrong acts. The fixture has 8 positives, which is too few to fit a
+threshold on.
+
+The 299 generated rows have no held-out slice of their own. Clarify is scored on
+the fixture alone.

docs/evals/2026-08-08-destination-fixture.md

@@ -0,0 +1,84 @@
+# The first destination number
+
+Measured 2026-08-08 on the classifier cascade with the ONNX multilingual
+embedder, the configuration homesrv runs. `make t PKG=./internal/router/eval/
+RUN=TestONNXBaseline V=1`. Covers V-659, the follow-up V-655 named.
+
+## What was measured
+
+V-655 split a routing decision in two. The cascade sorts an utterance into one
+of seven intents, and `Decision.Source` then says where the answer lives. The
+first half had a fixture. The second half arrived with none, so twelve
+destinations shipped with no accuracy number.
+
+`want_source` is now a field on `eval.Case`. It is a pointer, because the
+destination has three states and a bare string has two. Absent is every intent
+but query, which never reaches `queryWalk`. Present and empty is the
+`SourceUnknown` contract: name nothing and let the daemon walk the chain.
+Present and named is a destination the route must produce.
+
+Thirty-three of the ninety-six cases carry one. A destination miss does not
+fail the case, so `Accuracy` and `IntentAccuracy` mean what they meant.
+`SourceAccuracy` is a second number over the labelled cases only.
+
+## Result
+
+Intent is **73/96 (76.0%)**, against 69/91 (75.8%) before. Four of the five new
+cases pass and no existing case moved.
+
+Destination is **12/33 (36.4%)**, and the split is the whole finding.
+
+| destination | scored | note |
+|---|---|---|
+| world | 5/5 | `WorldQueryGrammars` names it at stage 0 |
+| the `SourceUnknown` floor | 5/7 | the two misses lost the intent first |
+| calendar | 2/6 | `calendar-query` names it, the possessive agenda rules do not |
+| recall | 0/15 | nothing anywhere names it |
+
+Recall is the number to move. Fifteen cases ask about his own words and his own
+facts. The route lands `query` on eleven of them and the destination comes back
+empty every time. Those turns are answered today, because the daemon walks the
+chain in order and the three recall passes are early in it. What is missing is a
+decider that says so, and that is the fourth head on V-546.
+
+Two cases labelled the floor lost their intent before a destination was
+possible. A clarify names nothing, so it would satisfy an empty label for free.
+`Score` requires the route to land the case's intent before it credits a
+destination hit, or the floor label would score itself.
+
+## Seven cases assert the floor, and five of them cluster
+
+The five are homelab operations. `SourceRecall`, `SourceNetwork` and
+`SourceAttention` overlap on every question about the box, because `mavpoll`
+writes its netdata and uptime-kuma observations into the fact store recall
+reads. "почему сервер тормозит" is answerable from all three. Naming one takes
+the other two off the turn.
+
+That is a finding about the enum rather than a gap in the labelling. The floor
+is the right answer there and the fixture now says so out loud.
+
+All seven were written by an agent and confirmed by the owner on 08-08-2026.
+
+## A drift the labelling found
+
+`WorldQueryGrammars` went into `buildRouter` with V-655 and never into
+`baselineGrammars`, the fixture's mirror of it. So the fixture was scoring a
+grammar set the daemon does not run. The comment above that function forbids
+exactly that. Adding it moved the destination number from 9/33 to 12/33 and
+moved nothing else.
+
+The three cases it recovered are `что такое TCP?`, `сколько будет 17 на 23?`
+and `кто такой Линус Торвальдс?`. All three already routed `query` through
+`NarrativeQueryGrammars`. So the drift was invisible to every number this
+fixture reported, until the destination had one of its own.
+
+## What this does not measure
+
+The model arm. This is the classifier cascade, which names a destination only
+where a stage 0 rule filled one in. The resident model has no destination in
+its router prompt yet, so 36.4% is a floor and not a comparison.
+
+Two pairs of cases are the same utterance. `ru-query-020` and `ru-query-024`
+are both "что дальше?", and `ru-query-021` and `ru-query-025` are both
+"расскажи про битву при Ватерлоо". They differ in tags and note only, so both
+pairs are counted twice here and in every earlier number this fixture reported.

docs/evals/2026-08-08-destination-model-arm.md

@@ -0,0 +1,66 @@
+# The destination, with a model that can name one
+
+Measured 2026-08-08 against gemma-4-12b on the workstation, the same 96-case
+fixture V-659 built. Covers V-660.
+
+```sh
+no_proxy='*' MAVEN_LLM_URL=http://192.168.1.105:8080 \
+  make t PKG=./internal/router/eval/ RUN=TestLLMRouterBaseline V=1
+```
+
+## The gap was structural
+
+V-659 measured the destination at 12/33 on the classifier cascade, with recall
+at 0/15. Nothing in `routeSystem` named a `Source` and `routeGrammar` could not
+emit one, so the resident model had no string to write. That is the shape V-517
+measured for Praxis reach at 0/12: not a weak model, an absent contract.
+
+`routeGrammar` now carries a `source` rule closed over `router.Sources` plus the
+empty floor. The prompt lists the twelve destinations in Russian and says that
+`""` is a normal answer to give often.
+
+## Result
+
+| run | intent | destination |
+|---|---|---|
+| classifier + ONNX (V-659) | 73/96 (76.0%) | 12/33 (36.4%) |
+| gemma-4-12b alone | 79/96 intent-only (82.3%) | 26/33 (78.8%) |
+| cascade + gemma-4-12b + hash fallback | 81/96 (84.4%) | 24/33 (72.7%) |
+
+Recall is the move: 0/15 to 14/15. Intent did not shift, which was the
+constraint. The prompt is shared, so a destination rule that costs routing
+points is not a win.
+
+The eight llm-only errors are the eight `want_clarify` cases. The model returned
+`unknown` on every one, which is correct, and the llm-only harness surfaces a
+decline as an error by design.
+
+## Stage 0 now costs four destination points
+
+The four cases the cascade loses and the model alone wins are all calendar. The
+possessive agenda rules claim them at stage 0 and deliberately name nothing.
+"что у меня в списке покупок" matches the same rule. Naming the calendar there
+would take the list source off the turn (V-655).
+
+So a rule written to be careful about the list now blocks a model that would
+have named the calendar correctly. Before V-660 that caution was free, because
+nothing downstream of stage 0 could name anything either.
+
+Three ways out, and each costs something. Split the possessive rule so the
+calendar-shaped half names its destination. Let a later stage overwrite an empty
+destination a grammar left behind, which reverses "a matched value always wins".
+Or leave it, on the argument that four points is cheap next to a wrong
+destination on a shopping list. This wants the owner's call rather than a quiet
+edit.
+
+## What this does not measure
+
+The resident Qwen3-1.7B, which is what homesrv runs. It binds `--port 0` inside
+the container and no host process can reach it. Scoring it needs a second
+llama-server on a fixed port. The workstation is never assumed
+up, so the homesrv number is the one that decides whether this ships on by
+default.
+
+The fixture is 33 labelled destinations over twelve values. Recall carries 15 of
+them and five destinations carry none at all. A per-destination number below
+world, recall, calendar and the floor is not supported by this fixture.

docs/evals/2026-08-08-massive-warm-start.md

@@ -0,0 +1,142 @@
+# MASSIVE Russian warm-start for the routing heads
+
+Measured 2026-08-08 on workpc (Radeon RX 7900 GRE, ROCm). Covers V-546 step 2.
+Workspace is `~/Programs/embed-training` on workpc, scripts `train_massive.py`,
+`ab_run.py`, `ab.sh`, `probe_time.py`.
+
+## What was trained
+
+Two heads on a copy of multilingual-e5-small: `Linear(384, 60)` for MASSIVE's
+own intents over a masked mean pool, `Linear(384, 111)` per token for BIO slot
+tags. MASSIVE's label sets verbatim, no alignment to Maven's 7 intents. The
+intent head is an auxiliary loss that shapes the pooled vector and is thrown
+away.
+
+Data is `amazon-massive-dataset-1.1` pulled from S3. The Hugging Face repo is
+script-only and `datasets` 5.0 refuses those, so `load_dataset` cannot fetch it.
+`ru-RU` is 11,514 train, 2,033 dev, 2,974 test, 60 intents, 55 slots, 111 BIO
+labels. All 16,521 rows survived span alignment: `annot_utt` re-tokenised to its
+own `utt` on every one.
+
+Hyperparameters match `train_intent.py`, so the two runs differ in data only.
+Frozen XLM-R vocabulary, body 2e-5, heads 1e-3, batch 32, sequence 64, 10
+epochs. MASSIVE's own dev partition selects the epoch, on slot F1 with intent
+accuracy as tiebreak. Selecting on 60-class intent accuracy would optimise a
+head that gets deleted.
+
+## Result
+
+Epoch 9 of 10 by dev slot F1. Held-out MASSIVE test: intent 86.2%, slot span
+F1 71.5% (P 68.5, R 74.8). Peak 1.70GB of 17.2GB, about 22 seconds an epoch,
+under 4 minutes end to end. Dev slot F1 climbed monotonically to epoch 9 and
+fell at 10, so 10 epochs was the right budget.
+
+Ten slot types sit at 0% test recall. Every one of them has 1 to 7 test
+instances: `alarm_type` has 3, `drink_type` has 1. That is support in MASSIVE's
+Russian split, not a tagger failure. `playlist_name` at 6% of 16 is the first
+real miss.
+
+## The intent A/B, and why it settles nothing
+
+`train_intent.py` was run against both bodies, three seeds by two smoothing
+settings, on `train_v4.jsonl`. It is v4 and not v5 because v4 is what
+`sweep2.log` measured. `ab_run.py` strips a `--base` flag onto the module global, so
+`train_intent.py` is unmodified and its baseline stays reproducible. The stock
+arm reproduced `sweep2.log` line for line.
+
+Fixture accuracy, 91 cases, one case is 1.1 points:
+
+| seed / smooth | stock | warm-started |
+|---|---|---|
+| 0 / 0.0 | 94.0% | 92.8% |
+| 0 / 0.1 | 95.2% | 92.8% |
+| 1 / 0.0 | 95.2% | 94.0% |
+| 1 / 0.1 | 95.2% | 97.6% |
+| 2 / 0.0 | 92.8% | 94.0% |
+| 2 / 0.1 | 92.8% | 96.4% |
+
+Mean 94.2% against 94.6%. That is +0.4 points, about a third of one case, and
+inside seed noise. Spread widened. Stock lands in a 2.4-point band and
+warm-started in a 4.8-point one. The warm-started arm holds both the best result
+of the sweep and a tie for the worst. Seed 0 is the bad arm and it fails in a
+specific way. Its dev peaks at epoch 2 and 3 and never improves, where stock
+peaks around 7. The dev slice is a quarter of the seed rows. That is small
+enough that early stopping is fragile when the body arrives already fitted.
+
+**The A/B was never the test.** Intent had at most 4.8 points of headroom here.
+MASSIVE was not trained for Maven's intents. Read it as "the warm-start does not
+cost intent accuracy", nothing more.
+
+## The measurement that does mean something
+
+`want_time` is the one slot Maven's fixture scores, and MASSIVE has `time` and
+`date`. Restricted to those two slot types, F1 is 74.9% over 609 gold spans on the
+MASSIVE ru test split. Precision is 71.5 and recall 78.7. That beats the 71.5%
+all-slot figure. Of the 530 test utterances carrying a time or a date, 73.4% get
+every such span exactly right.
+
+Out of domain matters more, because Maven's traffic is not this corpus. Ten
+Maven-shaped utterances, none of them in MASSIVE:
+
+| utterance | tagged |
+|---|---|
+| `напомни в 11:00 позвонить маме` | `time='11:00'`, `relation='маме'` |
+| `напомни завтра в семь утра выпить таблетки` | `date='завтра'`, `time='семь утра'` |
+| `поставь будильник на полседьмого` | `time='полседьмого'` |
+| `через двадцать минут напомни про чайник` | `time='двадцать минут'` |
+| `напомни в пятницу вечером забрать посылку` | `date='пятницу'`, `timeofday='вечером'` |
+| `что у меня сегодня после обеда` | `date='сегодня'`, `time='после'`, `timeofday='обеда'` |
+| `запиши что кофе закончился` | nothing |
+| `что такое TCP` | `definition_word='TCP'` |
+
+The first row is the V-572 defect utterance. `ReminderGrammar` handed the daemon
+`HasTime: false` there, and the daemon asked "Когда?" at a sentence that had
+already said when. `полседьмого` is a colloquial half-past that no digit pattern
+catches. `запиши что кофе закончился` correctly carries nothing, because a note
+has no time.
+
+Two errors. `после обеда` split into `time='после'` plus `timeofday='обеда'`
+when it is one span, and `через двадцать минут` dropped its `через`. Both are
+boundary errors on spans the tagger did find.
+
+Unplanned: `что такое TCP` returned `definition_word='TCP'`. MASSIVE has a slot
+for the thing being asked about, which is a `SourceWorld` signal sitting in a
+head already trained.
+
+Ten hand-picked utterances are evidence, not a fixture.
+
+## What this does not measure
+
+Maven has no span fixture. `want_time` and `want_fn` are presence booleans and
+`want_fact_key` is an exact string match, so nothing in the repo can score a
+71.5% span tagger. Destination got one the same day, at 12/33 on the classifier
+cascade: see `2026-08-08-destination-fixture.md`.
+
+The missing span fixture is why the warm-start stays unjudged against Maven
+rather than against MASSIVE.
+
+## Datasets ruled out
+
+Checked on 2026-08-08 and rejected as label sources:
+
+- **MASSIVE's other 50 locales** ship in the same tarball and are parallel by id.
+  Co-training on them is free and unmeasured. English was ruled out by the owner
+  on 2026-08-08.
+- **CLINC150** is reachable as parquet, 150 intents and 1,200 explicit
+  out-of-scope queries, English only. Its value is the labeled out-of-scope set
+  for fitting the energy threshold, not intent labels.
+- **`d0rj/dolphin-ru`**, roughly 2.8M rows of FLAN-style tasks translated to
+  Russian. No intent, no slots, and not utterances anyone says to an assistant.
+- **`psytechlab/EmpatheticIntents-ru`**, 24,856 rows of translated
+  EmpatheticDialogues with 32 emotion labels. Maven's mood enum is `neutral,
+  happy, thinking, tired, confused` and it describes her own reply, not the
+  speaker's emotion. No mapping exists.
+- **`ai-forever/MERA`** and **`RussianNLP/russian_super_glue`**, benchmark
+  harnesses. Rows are prompt templates with `{toxic_comment}` placeholders.
+- **`ZeroAgency/ru-big-russian-dataset`**, an LLM-judge quality corpus. Its
+  `question` and `classified_topic` columns are a usable Russian out-of-scope
+  pool for threshold fitting. That is the one thing CLINC150 can only supply in
+  English. The questions are long and written, so they belong in the negative
+  set, never in the in-scope `query` training set.
+- No second Russian slot-filling corpus exists. The xSID mirrors are 404,
+  MultiATIS++ has no Russian, SLURP is not on the Hub.

docs/evals/2026-08-08-parked-clarify-ride.md

@@ -0,0 +1,88 @@
+# The parked clarify ride, bounded and re-measured
+
+Date: 2026-08-08, V-663. Same 140 turns, same driver, third and fourth runs of
+the day. Before is `d6f3914`, after is that plus two changes.
+
+## What was measured before
+
+One question about a reminder's day rode turns 6 to 13. It ended only because
+turn 14 was a new request. Two of those turns are the worst replies in the
+corpus:
+
+```text
+спасибо   -> Сейчас 21:25. В какой день?
+привет    -> Сейчас 21:25. В какой день?
+```
+
+V-654 had already added `MaxSuspends` and the tail count had not moved.
+
+## Why three bounds let it happen
+
+The TTL, `MaxAttempts` and `MaxSuspends` all exist and all were rearmed.
+
+An aside spends no attempt, so `MaxAttempts` never reaches it. A turn reading as
+a failed answer zeroes `Suspends`, so `MaxSuspends` never reaches the asides.
+`спасибо` read as a failed answer. So the pleasantry both produced the reply
+above and paid for it out of the budget meant to end the ride.
+
+## The two changes
+
+`PendingQuestion.Rides` counts the same event as `Suspends` with the resets
+taken out. Set once, incremented only in `noteSuspended`, carried across the
+re-park in `askRemainingGap`. `MaxRides` is 4, one looser than `MaxSuspends` so
+that the tighter statement about a run stays reachable.
+
+A pleasantry is an aside, not a failed answer. The set is a new closed lexicon
+entry, matched as whole utterances rather than tokens. `вечер` answers `это утра
+или вечера?` and `нет` answers a confirm. Anything that could fill a slot stays
+out.
+
+## Result
+
+| | before | after |
+|---|---|---|
+| turns carrying a clarify tail | 21 | 17 |
+| turns carrying any failure string | 31 | 29 |
+| the longest ride | 8 turns | 4 turns |
+
+The turns carrying a tail, by number:
+
+```text
+before  6 7 8 9 10 11 12 13   53 54 55 56 57   99 102 103   116   136 137 138 139
+after   6 7 8 9                53 54 55 56 57   99 100 101   116   136 137 138 139
+```
+
+Turn 10 is the change. It now reads:
+
+```text
+спасибо -> Прошлую просьбу отпускаю. Пожалуйста, я всегда готова помочь тебе.
+```
+
+She lets the request go, says so, and answers the man. Turns 11 to 13 are clean.
+
+**`MaxRides` is not what fired.** The pleasantry is an aside now, so it no
+longer breaks the run. `MaxSuspends` reached three on turn 10 and ended it.
+`Rides` is the backstop for the shape where an answer really does break the run.
+No turn in this corpus reaches it.
+
+## What did not move
+
+Four rides are untouched. Turns 53 to 57 are five consecutive asides against a
+reminder missing its day. Turn 58 is a new request that drops it. Nothing
+pleasant appears in that run, so neither change applies. Turns 99 to 101 shifted
+by one, and 116 and 136 to 139 are unchanged.
+
+So the fix is worth four turns of twenty-one. What is left is asides against a
+question the owner never answers. `MaxSuspends` was written for that shape and
+does bound it, at four turns each.
+
+## Not attributable
+
+Latency moved p50 1.1s to 1.5s and p95 2.8s to 3.0s, and the 34.3s outlier in
+the earlier run is gone. Both runs had the workstation up. Read none of it as
+caused by this change.
+
+One unrelated defect appeared in the after run and is recorded here because it
+is visible in the transcript. Turn 4 answered `Я записала твою привычкуRegarding
+coffee without sugar.` That is English leaking into a Russian reply with no
+space in front of it. It is a phrasing defect and it has no task yet.

docs/evals/2026-08-08-routing-heads-in-go.md

@@ -0,0 +1,146 @@
+# The routing heads, running in Go
+
+Date: 2026-08-08. Vikunja V-664.
+Weights: `router_heads.onnx`, fp32, exported from `heads.pt` on workpc.
+Fixture: `internal/router/eval/ru_routing_v1.json`, 96 cases, 33 carrying a destination.
+Runner: `make t PKG=./internal/router/eval/ RUN=TestONNXRoutingHeads`.
+
+The four heads of V-661 ran nowhere. This is the number they score through the
+Go cascade. Same fixture and same grammars as `TestONNXBaseline`, and only the
+middle stage varies.
+
+## Headline
+
+| | classifier + ONNX | heads + classifier | gemma-4-12b cascade |
+|---|---|---|---|
+| intent | 75.0% (72/96) | **96.9% (93/96)** | 84.4% |
+| destination | 33.3% (11/33) | **75.8% (25/33)** | 72.7% |
+| false clarify | 0 | 1 | 2 |
+| missed clarify | 8 | 1 | 1 |
+| p50 | 24.5ms | 27.9ms | 329ms |
+
+A 118M encoder beats the 12B teacher it was distilled from. It wins on both
+halves of the route, at a twelfth of the latency. The workstation stays the
+better phraser and is no longer the better router.
+
+The p50 is not the heads. Most of it is the classifier's own embedder pass on
+the turns the heads decline, plus process warm-up on the first case. The heads'
+own forward pass measures 7.3ms on workpc.
+
+## Two defects were in the way, and the first was not in the heads
+
+**The tokenizer read every long word backwards.** `encodeWord` backtracks the
+Viterbi path from the end of a word and prepends each piece. That puts them back
+in reading order, and a second reverse after the loop undid it. So
+`query: вода` tokenized to `[0 12 1294 41 12489 2]` where the reference
+tokenizer gives `[0 41 1294 12 12489 2]`.
+
+It was found here and only here. The heads were trained through transformers and
+are read through the hand-written tokenizer. So a mismatch shows up as a score
+far below what Python measured on the same weights. Nothing else in the suite
+compares the two.
+
+Measured on the recall fixture, same 27 cases either way:
+
+| | reversed | fixed |
+|---|---|---|
+| recall@1 | 70.4% (19/27) | **77.8% (21/27)** |
+| recall@3 | 85.2% (23/27) | **96.3% (26/27)** |
+| answered after gate | 63.0% | 66.7% |
+| wrong note on top | 8 | 6 |
+| false recall | 0/5 | 1/5 |
+
+The classifier barely moved, 76.0% to 75.0%, and destination 36.4% to 33.3%.
+Both are one case on 96 and neither is a finding. Seeds and queries were mangled
+the same way, so cosine survived it. Recall is where it cost, because a stored
+passage and a live query are different lengths and break differently.
+
+The one new false recall is the honest cost and it is not being hidden. A
+sharper embedder scores every candidate higher, including the ones that should
+have stayed under the gate. That is the same trade `2026-08-04-recall-e5-small.md`
+recorded when e5-small replaced MiniLM.
+
+The embedder id now carries a tokenizer revision, `model_quantized@384/tok2`.
+Stored vectors were written under rev 1 and no longer sit in the same space as a
+query embedded now. The model file's name never moved, so nothing would have
+triggered `ReembedAll`. On the box the marker fired on start, and the re-embed
+rewrote 65 notes and 19 facts in 5 seconds.
+
+**The clarify head was being thrown away.** It was read only when the intent head
+cleared its own threshold. That cost 6 of the 8 ambiguous cases. `вода` reads as intent
+`act` at 0.233 and clarify at 0.983. Burying that handed the turn to the
+classifier, which routed it confidently and never asked. The clarify head answers
+a different question, which is whether there is enough here to act on at all. So
+it decides on its own and decides first.
+
+| | intent-gated | clarify decides first |
+|---|---|---|
+| intent | 90.6% | 96.9% |
+| missed clarify | 7 | 1 |
+| false clarify | 0 | 1 |
+
+## The threshold is measured, not chosen
+
+Max softmax over the intent head, on the 88 cases carrying an intent:
+
+| threshold | kept | accuracy kept | wrong kept | right dropped |
+|---|---|---|---|---|
+| 0.5 | 84 | 96.4% | 3 | 2 |
+| **0.6** | **81** | **97.5%** | **2** | **4** |
+| 0.7 | 75 | 97.3% | 2 | 10 |
+| 0.8 | 64 | 96.9% | 2 | 21 |
+| 0.9 | 46 | 100.0% | 0 | 37 |
+
+0.6 is the knee. Every value from 0.7 to 0.85 drops right answers and keeps the
+same two wrong ones. 0.9 is the only value that clears them, and it costs 37
+correct routes to do it.
+
+## Quantization was measured and rejected
+
+| build | size | intent | destination | p50 |
+|---|---|---|---|---|
+| fp32 | 470MB | 83/88 (94.3%) | 28/33 (84.8%) | 7.3ms |
+| int8 | 118MB | 79/88 (89.8%) | 26/33 (78.8%) | 4.0ms |
+| fp16 | 235MB | will not load | — | — |
+
+Python numbers, on the heads alone rather than through the cascade. int8 costs
+4.5 points of intent and 6 of destination to save 3ms. The cascade around it has
+a p50 over a second when the resident model answers. The fp16 graph is broken:
+`convert_float_to_float16` leaves a Cast node emitting float16 where the graph
+expects float, and onnxruntime refuses the session. It was not worth fixing.
+
+The exporter also had to be told to write one file. It splits weights into a
+`.onnx.data` sidecar by default. This onnxruntime resolves that path against the
+process working directory rather than the model. A split graph loads from one
+directory only.
+
+## What is still wrong
+
+**Four of the eight destination misses are calendar.** Training cannot move them.
+The possessive agenda rules claim those cases at stage 0 and name nothing on
+purpose. That caution was free while nothing downstream could name anything
+either. It has now cost four points in three separate measurements. The call is
+the owner's and it is still open.
+
+**The slot head is exported and not read.** Slots come from the stage-2
+extractor. Mapping BIO tags back to text needs character offsets the unigram
+tokenizer does not keep, which is its own piece of work.
+
+**`поужинал` is a false clarify**, which is the same defect `thinSingleToken`
+was narrowed for on 2026-08-01, arriving now from a different direction.
+
+## On the box
+
+Deployed to homesrv the same day. `voice: routing heads loaded` on start, and
+`/trace` shows `routing-heads` winning or thinning every turn. The resident model
+and the classifier are both marked never asked. Live probes:
+
+```text
+что такое TCP?                 -> kiwix     a real definition
+кто такой Линус Торвальдс?     -> kiwix     a real answer
+во сколько я лёг вчера         -> personal  не нашла у тебя такой записи
+вода                           -> thinned to clarify at 0.233 / 0.983
+```
+
+A missing or broken weights file logs and leaves the heads nil, which is
+byte-for-byte the cascade that shipped before this.

docs/evals/2026-08-08-routing-heads-two-head.md

@@ -0,0 +1,192 @@
+# Two heads on e5-small, and the first destination the router did not need a model for
+
+Measured 2026-08-08 on workpc (Radeon RX 7900 GRE, ROCm). Covers V-661, step 3 of
+`docs/plans/18-routing-heads-on-e5-small.md`. Workspace is `~/Programs/embed-training`,
+scripts `gen_query_source.py`, `label_source.py`, `build_heads_corpus.py`,
+`train_heads.py`, `score_confidence.py`.
+
+## Two heads, not four
+
+Intent is 7 classes and destination is 12 plus the `SourceUnknown` floor, sharing
+one masked mean pool over one forward pass. The plan asked for four. Two of them
+have no labels and neither is a GPU problem.
+
+**Mood is cut, not deferred.** Maven's enum is `neutral, happy, thinking, tired,
+confused` and it describes her own reply state, not the speaker's emotion.
+`psytechlab/EmpatheticIntents-ru` was the only candidate and its 32 emotion
+labels do not map onto it. There is nothing to train against.
+
+**BIO slot tags stay in the MASSIVE body.** No Maven-domain span corpus exists.
+`2026-08-08-massive-warm-start.md` records that no second Russian slot-filling
+corpus is reachable at all.
+
+The destination loss is masked with `ignore_index`. Only a query turn reaches
+`queryWalk`, so a reminder contributes nothing to it.
+
+## Where the destination labels came from
+
+V-660 taught the router prompt to name a destination. That made gemma-4-12b a
+teacher, and this distils it.
+
+Labelling the 300 query rows already in `train_v5.jsonl` gave 277 destinations.
+The shape was unusable: the floor 101, calendar 62, recall 47, and `feeds` and
+`attention` at zero. A 13-way softmax cannot learn a class with no examples.
+
+`gen_query_source.py` is the destination half of `gen_corpus.py` and runs the
+same two passes. Gemma writes questions whose answer lives in one named place.
+The daemon's own `routeSystem` prompt then routes each one back. A line survives
+only when the intent is `query` **and** the source is the destination it was
+generated for. The glosses are copied verbatim out of `route_system.txt`, so the
+generator and the labeller work from one definition.
+
+The prompt and the GBNF are dumped from `internal/router/llmrouter.go` by
+`TestDumpPrompt`, never retyped. The workspace held its own copies and V-660
+changed both.
+
+1229 kept of 2373 generated, 51.8%. Merged corpus is 3664 rows carrying 1727
+destinations:
+
+| | rows | | rows |
+|---|---|---|---|
+| the floor | 220 | world | 132 |
+| calendar | 180 | money | 126 |
+| tasks | 169 | list | 124 |
+| recall | 167 | self, feeds, attention | 120 each |
+| weather | 136 | network | 103 |
+| | | home | 50 |
+
+`home` is thin because the agreement filter rejected most of what was generated
+for it. A question about the house routes `act` more often than `query`. That is
+the filter working, and 50 is the finding rather than a shortfall.
+
+**The floor was regenerated once.** The first 120 rows carried one sentence
+shape across eight topics. That shape was "что там с X" and its two synonyms.
+Every named destination varied and only the floor collapsed. The reason is that
+the generator varies a topic, and ambiguity is not a topic.
+
+`gen_query_source.py` now rotates six floor shapes. A `почему` question, a yes
+or no question, and a question carried by intonation alone. Then a
+better-or-worse question, a status question, and an existence question. That is
+a fix to degenerate generation. It is not fitting to the fixture, whose floor
+cases are homelab operations and match none of the six.
+
+The 1229 generated rows carry `intent: null`. Every one is a query by
+construction. There are five times as many as the corpus has query rows, so
+including them would make query half the intent corpus.
+
+## Result
+
+Three seeds, two bodies, epoch chosen on the intent dev slice and never on a
+destination number.
+
+| body | floor corpus | intent mean | destination mean |
+|---|---|---|---|
+| warm-started `out/body_massive` | one shape | 93.6% | 75.8% |
+| stock `multilingual-e5-small` | one shape | 93.6% | 76.8% |
+| warm-started `out/body_massive` | six shapes | 93.6% | **80.8%** |
+
+Best single run is destination **29/33 (87.9%)**, seed 0 on the rotated floor.
+Peak 1.68GB of 17.2GB, under four minutes end to end.
+
+Against the two arms already measured on the same 33 labelled cases:
+
+| | destination |
+|---|---|
+| classifier cascade (V-659) | 12/33 (36.4%) |
+| cascade + gemma-4-12b (V-660) | 24/33 (72.7%) |
+| two heads on e5-small | 29/33 (87.9%) |
+
+A 118M encoder beats the 12B teacher it was distilled from, on the fixture. The
+per-destination split is where it happens: **recall 15/15** and **world 5/5**.
+Recall was 0/15 on the cascade and 14/15 through gemma.
+
+Read 87.9% as one seed of a mean of 80.8%, not as a headline. Three seeds score
+29, 25 and 26 of 33. One case is 3 points on a fixture this small.
+
+Intent is **not** comparable to the 73/96 and 81/96 figures those two arms
+scored. A softmax has no clarify class. The head's fixture is the 88 cases that
+carry an intent, and the 8 `want_clarify` cases are scored separately below.
+
+## The MASSIVE warm-start is worth nothing here either
+
+Step 2 measured it at +0.4 points of intent accuracy and called that inside seed
+noise. Destination was the open question, because MASSIVE has a
+`definition_word` slot that looked like a `SourceWorld` signal sitting in a head
+already trained.
+
+It is not. The two bodies score the same intent mean to one decimal. Stock is
+one point ahead on destination, which is a third of one case. Nothing here argues
+for keeping the warm-start step. Dropping it removes a dependency on a corpus
+pull that `datasets` 5.0 cannot do.
+
+## The floor moved, calendar did not
+
+Before the rotation, all seven misses at seed 0 were the floor and calendar. The
+head named a destination where the fixture says walk the chain, and it was
+confident doing it. `"почему сервер тормозит"` read `world` at 0.80.
+`"хватает ли места под новые бэкапы"` read `network` at 0.82. Those are the five
+homelab cases V-659 flagged, where `SourceRecall`, `SourceNetwork` and
+`SourceAttention` all overlap. `mavpoll` writes its observations into the fact
+store recall reads.
+
+| | one shape | six shapes |
+|---|---|---|
+| the floor | 3/7 | 6/7, 6/7, 5/7 |
+| calendar | 3/6 | 3/6, 3/6, 3/6 |
+| recall | 15/15 | 15/15 at seed 0 |
+| world | 5/5 | 5/5 |
+
+The floor was a corpus defect and it cost 3 cases. Sentence variety carried it,
+not homelab vocabulary, which the training rows still do not contain.
+
+**Calendar is 3/6 at every seed and is a different problem.** It is the shape
+V-660 named. The possessive agenda rules claim those cases at stage 0 and
+deliberately name nothing, so no destination label reaches the head. Training
+cannot move a case the head never sees. That one wants the owner's call.
+
+## Max softmax separates, weakly, and the gate stays
+
+The plan argues max softmax is a calibratable confidence where `Confidence: 1.0`
+was a hardcode. Measured on the intent head:
+
+| | n | mean confidence |
+|---|---|---|
+| correct | 80 | 0.897 |
+| wrong | 8 | 0.705 |
+| `want_clarify` | 8 | 0.685 |
+
+The softest correct answer is 0.66 and 3 of 8 clarify cases sit below it. So a
+single cut buys three clarifies at no false-clarify cost, and no more.
+
+The other five explain themselves. `"напомни"` scores 0.94 as `reminder` and
+`"сделай это"` scores 0.80 as `act`. Both are intent-certain and slot-empty, and
+confidence was never the signal there. `gateLLMDecision` already catches exactly
+that shape, an act with no allowlisted fn or a keyless fact, and it keeps doing
+so. The head replaces the hardcode. It does not replace the gate.
+
+## An incident worth recording
+
+The first generation run produced zero rows for eight destinations. `mavgpud`
+yields the card when another process wants it (V-488) and llama-server answers
+503 until the model is back. Every generate call inside that window burned one of
+the destination's batches. The run walked its own cap without a single successful
+call. The log said `503` 260 times, and the summary line said 64.1% keep rate,
+which read as success.
+
+`call()` now retries a 503 with backoff. A generator that treats an unloaded
+model as a bad generation is a silent-corpus bug, not a slow one.
+
+## What this does not measure
+
+**Nothing here runs in Go.** The heads are a `heads.pt` and an
+`out/body_heads/` on workpc. Reaching the daemon needs an ONNX export and a
+caller. The resident e5-small must not be replaced by this copy: recall depends
+on that file, and `EmbedQuery`/`EmbedPassage` are its contract.
+
+The destination fixture carries 4 of 13 classes: recall 15, floor 7, calendar 6,
+world 5. `tasks`, `money`, `list`, `home`, `network`, `feeds`, `attention` and
+`self` have no gold case. So 78.8% is silent on eight destinations that together
+hold 800 training rows.
+
+Latency was not measured. A forward pass of a 118M encoder should beat a 1.7B
+decoder on a query turn. That is arithmetic, not a number from this box.

docs/evals/2026-08-08-slot-head-three-head.md

@@ -0,0 +1,106 @@
+# A slot head, and the corpus that did not exist this morning
+
+Measured 2026-08-08 on workpc, the same day as `2026-08-08-routing-heads-two-head.md`
+and against the same fixtures. Workspace is `~/Programs/embed-training`, new
+scripts `slot_grammar.gbnf`, `slot_system.txt`, `label_slots.py`, `merge_slots.py`.
+
+## The corpus was the whole problem
+
+The two-head measurement said BIO slot tags stay in the MASSIVE body, because no
+Maven-domain span corpus exists. That was true of found corpora and false of
+made ones. Destination had the same shape at breakfast. V-660 gave gemma-4-12b a
+string to write and the label problem became a generation problem.
+
+The same trick applies to spans. `slot_grammar.gbnf` emits a list of
+`{"slot": ..., "text": ...}` and the enum closes over Maven's own five: `time`,
+`text`, `key`, `value`, `fn`. `slot_system.txt` demands each span be an exact
+substring of the utterance.
+
+**The agreement filter is free here.** Destination needed a second pass. The
+daemon's own router prompt had to route each generated line back. A span needs
+no second call. It either occurs in the utterance or it does not, and
+`label_slots.py` drops it with `find()`.
+
+1702 rows labelled from `train_v5.jsonl`, the reminder, fact, note, act and
+query intents. Chat and system carry no slot and were never asked.
+
+| slot | spans |
+|---|---|
+| text | 1175 |
+| time | 485 |
+| fn | 381 |
+| key | 72 |
+| value | 65 |
+
+37 spans dropped as not-a-substring, 2.2% of the pile. Nothing failed to parse,
+which is the grammar doing its job. 409 rows came back with no span at all.
+Those are kept and tagged all `O`. An utterance carrying no slot teaches the
+head not to invent one. An empty list is a label and not a miss.
+
+`key` and `value` are thin because they come from facts alone. That is the
+shape of the corpus, not a labeller failure.
+
+## Three heads on one forward pass
+
+Intent and destination were already two linear heads over one masked mean pool.
+Slots is a third head over the per-token states of the same pass, so the marginal
+cost is one `Linear(384, 11)`.
+
+The tag set is `O` plus `B-` and `I-` for each of the five. A softmax cannot
+emit a tag that does not exist. That is the structural guarantee the GBNF buys
+for the teacher, and the head gets it for free.
+
+Two masking rules, both `ignore_index`. A row with no `spans` key contributes
+nothing, which covers the 1900 generated destination rows and every chat and
+system turn. A padding or special-token position contributes nothing either.
+
+Scoring is exact-match span F1, not token accuracy. Most tokens are `O`, so a
+tagger that predicts nothing anywhere scores above 90% on tokens.
+
+## Result
+
+Three seeds, 24 epochs, epoch chosen on the intent dev slice alone.
+
+| | two heads | three heads |
+|---|---|---|
+| intent mean | 93.6% | 92.8% |
+| destination mean | 80.8% | 82.8% |
+| destination best | 29/33 (87.9%) | 29/33 (87.9%) |
+| slot span F1 mean | — | 72.4% |
+
+**The slot head costs nothing and adds a third decision.** Intent moves 0.8
+points down and destination 2 points up. Both sit inside the seed spread those
+two numbers already had. Read this as unchanged, not as a trade.
+
+The saved checkpoint is seed 1 at epoch 10: intent 93.2%, destination 81.8%,
+slot F1 75.8%. `heads.pt` now carries three state dicts and the `bio` list
+beside the intent and source enums.
+
+## Epoch selection is now wrong for one of the three heads
+
+Slot F1 was still climbing when the intent-selected epoch stopped it. Seed 0
+selects epoch 13 at 70.9% and reaches 76.1% at epoch 20. Seed 1 selects epoch 10
+at 75.8% and reaches 80.0% at epoch 24.
+
+So the three tasks want different epochs and the harness picks one. Two ways
+out, and neither was taken here. Select on a joint score, which needs an
+argument about weights. Or give the slot head its own dev slice and its own
+early stop, which means the heads stop being one checkpoint.
+
+Leaving it costs about 4 points of slot F1 and nothing else. The router
+contract is graded on intent. Selecting on it is the conservative choice rather
+than an oversight.
+
+## What this does not measure
+
+The same gap as the two-head run. **Nothing of this runs in Go.** Three heads
+instead of two does not change that. It does change the export. The slot head
+reads the per-token states, so an ONNX graph exposing only the pooled vector is
+not enough.
+
+There is no held-out span fixture. Span F1 is measured on the dev slice of the
+same distilled corpus, so it scores agreement with gemma rather than
+correctness. The V-572 defect utterance, `"напомни в 11:00 позвонить маме"`, is
+labelled correctly by the teacher, but one utterance is not a fixture.
+
+`key` at 72 spans and `value` at 65 are too thin to report a per-slot number.

docs/evals/2026-08-08-two-weeks-after-merge-transcript.md

@@ -0,0 +1,631 @@
+# Raw transcript, two weeks of usage
+
+## Day 1
+
+```
+YOU:   доброе утро
+MAVEN: Я не говорю так сейчас, ведь на улице уже вечер.
+       [2.8s t=218]
+
+YOU:   какой сегодня день?
+MAVEN: сегодня суббота, 8 августа 2026 года
+       [0.0s t=219]
+
+YOU:   сколько времени?
+MAVEN: сейчас 21 час 9 минут
+       [0.0s t=220]
+
+YOU:   запиши что я пью кофе без сахара
+MAVEN: Я записала твою заметку про кофе。
+       [1.5s t=221]
+
+YOU:   мой любимый язык программирования go
+MAVEN: Я записала твою заметку про Go.
+       [1.5s t=222]
+
+YOU:   напомни в 11:00 позвонить маме
+MAVEN: Сейчас 21:10. В какой день?
+       [0.4s t=223]
+
+YOU:   что у меня сегодня?
+MAVEN: на 08.08.2026 ничего нет. На какое время поставить напоминание?
+       [2.5s t=224]
+
+YOU:   что такое TCP?
+MAVEN: вот что я нашла: TCP (Transmission Control Protocol) — это транспортный протокол для передачи данных. Он обеспечивает доставку всех пакетов и позволяет собрать их в правильном порядке. Также TCP является частью сетевой модели TCP/IP, которая описывает способ передачи цифровых данных от источника к получателю. На какое время поставить напоминание?
+       [2.8s t=225]
+
+YOU:   сколько будет 17 на 23?
+MAVEN: вот что я нашла: в источниках нет результата умножения 17 на 23. На какое время поставить напоминание?
+       [1.6s t=226]
+
+YOU:   спасибо
+MAVEN: Сейчас 21:10. В какой день?
+       [1.1s t=227]
+
+```
+
+## Day 2
+
+```
+YOU:   привет
+MAVEN: Сейчас 21:10. В какой день?
+       [0.9s t=228]
+
+YOU:   что нового?
+MAVEN: не знаю — не нашла у тебя такой записи. На какое время поставить напоминание?
+       [1.3s t=229]
+
+YOU:   какая погода?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [0.7s t=230]
+
+YOU:   запиши что пароль от вайфая лежит в ящике стола
+MAVEN: Прошлую просьбу отпускаю. Я записала, что пароль от вай-фай лежит в ящике стола.
+       [1.5s t=231]
+
+YOU:   где лежит вайфай пароль?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.5s t=232]
+
+YOU:   добавь молоко в список покупок
+MAVEN: Такую команду я не знаю.
+       [1.0s t=233]
+
+YOU:   что у меня в списке покупок?
+MAVEN: в списке покупок пусто.
+       [0.0s t=234]
+
+YOU:   кто такой Линус Торвальдс?
+MAVEN: вот что я нашла: Линус Торвальдс — финско-американский программист из Хельсинки, создатель ядра Linux. Он родился 28 декабря 1969 года и окончил факультет компьютерных наук Хельсинкского университета.
+       [2.6s t=235]
+
+YOU:   какой у меня любимый язык?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.6s t=236]
+
+YOU:   сколько у меня задач?
+MAVEN: сначала: проверить интейк форму на боксе (завтра); купить хлеб; купить молоко; полить цветы; починить кран
+       [0.0s t=237]
+
+```
+
+## Day 3
+
+```
+YOU:   как дела?
+MAVEN: У меня всё хорошо, я готова помогать тебе. Чем могу быть полезна?
+       [1.2s t=238]
+
+YOU:   напомни завтра в 9 утра купить хлеб
+MAVEN: хорошо, напомню завтра в 09:00.
+       [0.3s t=239]
+
+YOU:   что у меня завтра?
+MAVEN: на 09.08.2026 ничего нет.
+       [0.0s t=240]
+
+YOU:   отмени напоминание про хлеб
+MAVEN: Такую команду я не знаю.
+       [0.5s t=241]
+
+YOU:   какие у меня напоминания?
+MAVEN: вот что я нашла: в источниках нет информации о твоих напоминаниях.
+       [0.7s t=242]
+
+YOU:   сохрани мне адрес гостиницы в Сочи
+MAVEN: Я записала твой адрес отеля в Сочи.
+       [2.2s t=243]
+
+YOU:   что я сохранил про Сочи?
+MAVEN: вот что я нашла: в источниках нет информации о том, что ты сохранил про Сочи.
+       [1.5s t=244]
+
+YOU:   почему сервер тормозит?
+MAVEN: вот что я нашла: сервер может тормозить из-за неудачной конфигурации или неоптимальных настроек, отсутствия кэширования, неоптимизированного программного обеспечения, которое нужно обновлять, а также из-за режима энергосбережения процессора, когда он не успевает повысить частоту при нагрузке.
+       [3.7s t=245]
+
+YOU:   хватает ли места под новые бэкапы?
+MAVEN: нового вот что: [$] Changes in shadow-utils password-expiration features; The Software Stewardship Lab launches; LightDM lives: version 1.33.0 released
+       [0.7s t=246]
+
+YOU:   выключи свет в спальне
+MAVEN: Такую команду я не знаю.
+       [0.4s t=247]
+
+```
+
+## Day 4
+
+```
+YOU:   доброе утро
+MAVEN: Сейчас уже вечер, поэтому я не могу сказать тебе «доброе утро».
+       [1.4s t=248]
+
+YOU:   что я пропустил?
+MAVEN: вот что я нашла: в источниках нет информации о том, что ты пропустил.
+       [1.3s t=249]
+
+YOU:   о чём мы вчера говорили?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.0s t=250]
+
+YOU:   запиши что я записался к врачу на четверг
+MAVEN: Я записала твою запись к врачу.
+       [1.2s t=251]
+
+YOU:   когда я иду к врачу?
+MAVEN: вот что я нашла: в источниках нет информации о том, когда именно ты идешь к врачу.
+       [2.1s t=252]
+
+YOU:   что такое ZFS?
+MAVEN: вот что я нашла: ZFS — это современная и мощная файловая система, в которой сочетаются безопасность, масштабируемость и удобство управления данными. Она была разработана корпорацией Sun Microsystems и объединяет в себе файловую систему с менеджером томов. Изначально она создавалась для того, чтобы преодолеть недостатки традиционных RAID-массивов, такие как проблема «дыры записи».
+       [2.1s t=253]
+
+YOU:   столица Франции?
+MAVEN: вот что я нашла: столица Франции — Париж.
+       [3.7s t=254]
+
+YOU:   переведи слово ремонт на английский
+MAVEN: вот что я нашла: слово «ремонт» на английский можно перевести как repair (например, ремонт дороги или крыши) или fix (общее слово для починки техники и мебели).
+       [2.7s t=255]
+
+YOU:   сколько я потратил в этом месяце?
+MAVEN: я не отслеживаю траты — не подключено.
+       [0.7s t=256]
+
+YOU:   спокойной ночи
+MAVEN: Спокойной ночи. Я уже подготовила всё к твоему отдыху.
+       [1.2s t=257]
+
+```
+
+## Day 5
+
+```
+YOU:   привет
+MAVEN: Привет! Я готова помогать тебе сегодня.
+       [1.0s t=258]
+
+YOU:   какая погода в Москве?
+MAVEN: погода не настроена.
+       [0.9s t=259]
+
+YOU:   что там с бэкапами?
+MAVEN: вот что я нашла: бэкап — это резервная копия данных, созданная в определенный момент времени и сохраненная на локальном или облачном носителе. Его делают, чтобы быстро восстановить информацию в случае необходимости.
+       [2.3s t=260]
+
+YOU:   покажи что требует внимания
+MAVEN: мне пока нечего смотреть — у Praxis нет источников.
+       [0.5s t=261]
+
+YOU:   отметь это как сделанное
+MAVEN: команды «отметь» пока нет в списке — она уже предложена, включи через клиент.
+       [0.0s t=262]
+
+YOU:   запиши что я купил новые наушники
+MAVEN: Я записала твою покупку новых наушников.
+       [1.8s t=263]
+
+YOU:   какие у меня заметки за неделю?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.0s t=264]
+
+YOU:   расскажи про Kubernetes
+MAVEN: вот что я нашла: Kubernetes — это портативная расширяемая платформа с открытым исходным кодом для управления сервисами и контейнеризованными рабочими нагрузками. Управлять им можно через интерфейс в командной строке kubectl. Я нашла эти данные в предоставленных источниках.
+       [1.8s t=265]
+
+YOU:   кто я?
+MAVEN: вот что я нашла: в источниках нет информации о том, кто ты.
+       [1.4s t=266]
+
+YOU:   пока
+MAVEN: До свидания, пока!
+       [1.4s t=267]
+
+```
+
+## Day 6
+
+```
+YOU:   доброе утро
+MAVEN: Сейчас уже вечер, так что доброе утро я тебе сказать не могу. Я готова помочь тебе с чем-нибудь другим.
+       [1.4s t=268]
+
+YOU:   сколько времени?
+MAVEN: сейчас 21 час 11 минут
+       [0.0s t=269]
+
+YOU:   напомни в 18:30 позвонить в банк
+MAVEN: Сейчас 21:11. В какой день?
+       [0.3s t=270]
+
+YOU:   поставь чайник
+MAVEN: Сейчас 21:11. В какой день?
+       [1.1s t=271]
+
+YOU:   включи музыку
+MAVEN: Сейчас 21:11. В какой день?
+       [0.9s t=272]
+
+YOU:   что у меня в календаре на пятницу?
+MAVEN: вот что я нашла: в источниках нет информации о твоих личных событиях в календаре на пятницу. В одном из текстов упоминается пятница 24 июля 2026 года, а в другом — что сегодня пятница 7 августа 2026 года. На какое время поставить напоминание?
+       [2.6s t=273]
+
+YOU:   во сколько у меня встреча?
+MAVEN: не знаю — не нашла у тебя такой записи. На какое время поставить напоминание?
+       [0.3s t=274]
+
+YOU:   запиши что дедлайн по проекту в понедельник
+MAVEN: Прошлую просьбу отпускаю. Я записала твой дедлайн на понедельник.
+       [1.7s t=275]
+
+YOU:   успею ли я до дедлайна?
+MAVEN: вот что я нашла: в источниках нет информации о твоем дедлайне, поэтому я не могу ответить, успеешь ты или нет.
+       [2.0s t=276]
+
+YOU:   спасибо
+MAVEN: Пожалуйста, я всегда готова помочь тебе.
+       [1.4s t=277]
+
+```
+
+## Day 7
+
+```
+YOU:   привет
+MAVEN: Привет! Чем я могу тебе помочь?
+       [1.3s t=278]
+
+YOU:   как ты?
+MAVEN: Я в порядке, спасибо, что спросил. Готова помогать тебе!
+       [1.4s t=279]
+
+YOU:   расскажи анекдот
+MAVEN: вот что я нашла: в источниках есть ссылки на сборники и сайты с анекдотами, но самих текстов анекдотов там нет, кроме одного: ударилась лягушка о землю головой и стала девицей прекрасной, а второй раз — месивом кровавым.
+       [3.9s t=280]
+
+YOU:   что ты умеешь?
+MAVEN: Я Мэйвен, твоя помощница. Я живу на твоём сервере и отвечаю на твои вопросы. Я запоминаю то, что ты мне говоришь, веду заметки, ставлю напоминания, а также читаю твой календарь и задачи. Я отвечаю на вопросы о мире, используя поиск или офлайн-энциклопедию. Если ты разрешишь, я могу управлять домом, локальной сетью, лентами, списком покупок и погодой, а также писать в телеграм. Я говорю по-русски и по-английски.
+       [2.9s t=281]
+
+YOU:   запиши что я начал бегать по утрам
+MAVEN: Я записала твою новую привычку
+       [1.5s t=282]
+
+YOU:   я бегаю по утрам уже неделю
+MAVEN: записала: я бегаю по утрам уже неделю
+       [0.6s t=283]
+
+YOU:   как часто я бегаю?
+MAVEN: вот что я нашла: в источниках нет информации о том, как часто бегаешь ты. Там указано, что для похудения частоту тренировок можно доводить до 5–6 раз в неделю, а новичкам ежедневные тренировки без подготовки могут привести к травмам.
+       [3.3s t=284]
+
+YOU:   сколько стоит биткоин?
+MAVEN: вот что я нашла: в источниках нет информации о стоимости биткоина.
+       [1.2s t=285]
+
+YOU:   какие новости?
+MAVEN: вот что нового: [$] Changes in shadow-utils password-expiration features; The Software Stewardship Lab launches; LightDM lives: version 1.33.0 released
+       [0.8s t=286]
+
+YOU:   хорошего дня
+MAVEN: Я не могу сказать «хорошего дня», так как сейчас уже вечер. Хорошего отдыха!
+       [1.3s t=287]
+
+```
+
+## Day 8
+
+```
+YOU:   доброе утро
+MAVEN: Сейчас вечер, так что доброе утро я не скажу. Чем я могу тебе помочь?
+       [1.2s t=288]
+
+YOU:   что у меня сегодня?
+MAVEN: на 08.08.2026 ничего нет.
+       [0.0s t=289]
+
+YOU:   напомни через час выпить воды
+MAVEN: хорошо, напомню сегодня в 22:11.
+       [0.3s t=290]
+
+YOU:   я выпил воды
+MAVEN: записала: я выпил воды
+       [0.9s t=291]
+
+YOU:   запиши что кот ест только сухой корм
+MAVEN: Я записала твою заметку про кота.
+       [1.2s t=292]
+
+YOU:   чем питается кот?
+MAVEN: вот что я нашла: в природе кошки едят мышей, птиц и ящериц. Дома их правильно кормить мясом, мясными субпродуктами, яйцами и нежирной морской рыбой. Из-за особенностей пищеварения им не подходят каши, макароны и мучные изделия.
+       [3.4s t=293]
+
+YOU:   что такое DNS?
+MAVEN: вот что я нашла: DNS (Domain Name System) — это распределенная база данных, которая сопоставляет домены с IP-адресами. Это не единая база на одном сервере, а иерархическая система серверов, работающих совместно. В ней хранятся записи, которые обеспечивают маршрутизацию трафика и управление электронной почтой.
+       [2.0s t=294]
+
+YOU:   проверь статус uptime kuma
+MAVEN: uptime kuma — Проверяю статус Uptime Kuma...
+       [1.1s t=295]
+
+YOU:   всё ли в порядке с сервером?
+MAVEN: мне пока нечего смотреть — у Praxis нет источников.
+       [0.5s t=296]
+
+YOU:   спасибо
+MAVEN: Пожалуйста, я всегда рада помочь тебе.
+       [2.7s t=297]
+
+```
+
+## Day 9
+
+```
+YOU:   привет
+MAVEN: Привет! Чем я могу тебе помочь сегодня?
+       [1.9s t=298]
+
+YOU:   какой сегодня день недели?
+MAVEN: сегодня суббота, 8 августа 2026 года
+       [0.0s t=299]
+
+YOU:   добавь хлеб и сыр в список покупок
+MAVEN: Такую команду я не знаю.
+       [0.6s t=300]
+
+YOU:   что в списке покупок?
+MAVEN: в списке покупок пусто.
+       [0.0s t=301]
+
+YOU:   удали молоко из списка
+MAVEN: Такую команду я не знаю.
+       [0.4s t=302]
+
+YOU:   напомни завтра утром вынести мусор
+MAVEN: Сейчас 21:11. Когда?
+       [0.3s t=303]
+
+YOU:   запиши что я поменял масло в машине
+MAVEN: Прошлую просьбу отпускаю. записала: запиши что я поменял масло в машине
+       [0.7s t=304]
+
+YOU:   когда я менял масло?
+MAVEN: масло в машине — поменял
+       [0.4s t=305]
+
+YOU:   сколько будет 144 делить на 12?
+MAVEN: вот что я нашла: в источниках указано решение деления 144 на 12 в столбик, но итоговый результат деления не прописан.
+       [2.3s t=306]
+
+YOU:   пока
+MAVEN: До свидания! Если понадобится что-то записать или проверить, я всегда на связи.
+       [1.2s t=307]
+
+```
+
+## Day 10
+
+```
+YOU:   доброе утро
+MAVEN: Я не могу сказать «доброе утро», так как сейчас уже вечер. Я записала твои предыдущие вопросы.
+       [1.8s t=308]
+
+YOU:   что нового за ночь?
+MAVEN: вот что я нашла: в ночь на 8 августа вооруженные силы РФ нанесли второй за последнее время массированный удар, при этом киевская ПВО пропустила все российские ракеты.
+       [3.3s t=309]
+
+YOU:   почему интернет медленный?
+MAVEN: вот что я нашла: интернет может работать медленно из-за плохого сигнала из-за удаленности вышки, исчерпанного лимита трафика по тарифу, нахождения вне зоны покрытия 4G/LTE, устаревшей или неисправной SIM-карты, а также из-за сбоя в настройках роутера. Чтобы исправить ситуацию, можно попробовать перезагрузить роутер или перенастроить устройство.
+       [3.0s t=310]
+
+YOU:   какая скорость у меня сейчас?
+MAVEN: для какого города?
+       [0.6s t=311]
+
+YOU:   запиши что новый роутер стоит 8000 рублей
+MAVEN: Я записала стоимость нового роутера.
+       [1.2s t=312]
+
+YOU:   сколько стоил роутер?
+MAVEN: вот что я нашла: цена зависит от модели. Роутер от Ростелеком стоит 5 900 рублей, на OZON есть варианты от 4 052 до 3 251 рублей, а в Билайне цены начинаются от 2 149 рублей.
+       [2.6s t=313]
+
+YOU:   что такое NAT?
+MAVEN: вот что я нашла: NAT (Network Address Translation) — это технология в сетях TCP/IP, которая позволяет изменять IP-адрес в заголовке пакета при его прохождении через устройство. Она перенаправляет трафик между локальной и глобальной сетями, позволяет объединять устройства в локальной сети под одним публичным IP-адресом, а также скрывает сеть и экономит IP.
+       [2.1s t=314]
+
+YOU:   напомни в субботу позвонить бабушке
+MAVEN: Сейчас 21:12. Когда?
+       [0.3s t=315]
+
+YOU:   покажи мои напоминания
+MAVEN: вот что я нашла: в источниках нет информации о твоих конкретных напоминаниях, там только описано, как их создавать и где искать в приложениях. На какое время поставить напоминание?
+       [3.0s t=316]
+
+YOU:   спасибо
+MAVEN: Сейчас 21:12. Во сколько напомнить?
+       [1.0s t=317]
+
+```
+
+## Day 11
+
+```
+YOU:   привет
+MAVEN: Сейчас 21:12. Скажи время — например, «в семь вечера» или «через час».
+       [0.8s t=318]
+
+YOU:   как погода на выходных?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [0.7s t=319]
+
+YOU:   что у меня на этой неделе?
+MAVEN: вот что я нашла: в источниках нет информации о том, что у тебя на этой неделе. На какое время поставить напоминание?
+       [1.0s t=320]
+
+YOU:   запиши что я хочу прочитать книгу про Go
+MAVEN: Прошлую просьбу отпускаю. Я записала твою заметку о книге по Go.
+       [1.6s t=321]
+
+YOU:   что я хотел прочитать?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.9s t=322]
+
+YOU:   объясни что такое горутина
+MAVEN: вот что я нашла: я проверила источники и узнала, что горутины — это легковесные потоки выполнения или функции, которые позволяют параллельно выполнять код внутри программы. Они являются легкими нитями, которые создаются и планируются для независимого выполнения планировщиком Go.
+       [2.5s t=323]
+
+YOU:   кто написал Войну и мир?
+MAVEN: вот что я нашла: роман «Война и мир» написал Лев Николаевич Толстой.
+       [2.2s t=324]
+
+YOU:   включи свет на кухне
+MAVEN: Такую команду я не знаю.
+       [0.5s t=325]
+
+YOU:   закрой шторы в комнате
+MAVEN: Такую команду я не знаю.
+       [0.4s t=326]
+
+YOU:   спокойной ночи
+MAVEN: Спокойной ночи! Я уже подготовила всё к твоему отдыху.
+       [1.1s t=327]
+
+```
+
+## Day 12
+
+```
+YOU:   доброе утро
+MAVEN: Сейчас уже вечер, поэтому я не могу сказать тебе «доброе утро».
+       [1.1s t=328]
+
+YOU:   сколько сейчас времени?
+MAVEN: сейчас 21 час 12 минут
+       [0.0s t=329]
+
+YOU:   я не то имел в виду
+MAVEN: Я поняла тебя, исправь, пожалуйста, свою мысль. Что именно ты хотел сказать?
+       [1.8s t=330]
+
+YOU:   о чём мы говорили?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.6s t=331]
+
+YOU:   напомни
+MAVEN: Сейчас 21:12. Когда?
+       [0.5s t=332]
+
+YOU:   сделай это
+MAVEN: пока не умею отвечать на этот вопрос. На какое время поставить напоминание?
+       [0.5s t=333]
+
+YOU:   запиши что я перешёл на новый тариф
+MAVEN: Прошлую просьбу отпускаю. Я записала твою новую информацию о тарифе.
+       [1.5s t=334]
+
+YOU:   какой у меня тариф?
+MAVEN: вот что я нашла: в источниках нет информации о твоем тарифе.
+       [1.2s t=335]
+
+YOU:   сколько я плачу за интернет?
+MAVEN: вот что я нашла: в источниках нет информации о том, сколько именно ты платишь за интернет. Я проверила документы, но точной суммы там нет.
+       [2.6s t=336]
+
+YOU:   спасибо
+MAVEN: Я поняла тебя. Что именно ты хочешь, чтобы я сделала сейчас?
+       [1.5s t=337]
+
+```
+
+## Day 13
+
+```
+YOU:   привет
+MAVEN: Привет! Я готова помогать тебе. Чем займемся сегодня?
+       [1.3s t=338]
+
+YOU:   что там с задачами?
+MAVEN: вот что я нашла: я отметила задачу про бэкапы как сделанную.
+       [1.5s t=339]
+
+YOU:   закрывай
+MAVEN: какой пункт отметить сделанным?
+       [0.0s t=340]
+
+YOU:   отметь задачу про бэкапы как сделанную
+MAVEN: Такую команду я не знаю.
+       [0.6s t=341]
+
+YOU:   что осталось нерешённым?
+MAVEN: вот что я нашла: в твоей жизни всё, что остается нерешенным или незавершенным, будет продолжать существовать и требовать внимания. Также в источниках упоминается пример с скобками, который может остаться нерешенным, и вопрос о том, что остается нерешенным в краснодарских перевозках.
+       [4.2s t=342]
+
+YOU:   запиши что я договорился о встрече в среду
+MAVEN: Я записала твою договоренность о встрече.
+       [2.3s t=343]
+
+YOU:   когда у меня встреча?
+MAVEN: мне пока нечего смотреть — у Praxis нет источников.
+       [0.0s t=344]
+
+YOU:   какая температура на улице?
+MAVEN: для какого города?
+       [0.9s t=345]
+
+YOU:   что такое RAID 5?
+MAVEN: вот что я нашла: RAID 5 — это массив из независимых дисков, где данные и контрольные суммы (четность) распределяются по кругу на все диски. Этот уровень обеспечивает отказоустойчивость и повышенную производительность, при этом для его работы требуется не менее трех дисков.
+       [2.3s t=346]
+
+YOU:   пока
+MAVEN: До встречи!
+       [1.4s t=347]
+
+```
+
+## Day 14
+
+```
+YOU:   доброе утро
+MAVEN: Я не могу сказать «доброе утро», потому что сейчас вечер. Я уже записала твой вопрос и готова помочь с чем-то другим.
+       [1.7s t=348]
+
+YOU:   подведи итоги недели
+MAVEN: week summary — подведи итоги недели
+       [0.6s t=349]
+
+YOU:   что я делал за последние две недели?
+MAVEN: вот что я нашла: в источниках нет информации о том, что ты делал за последние две недели.
+       [1.3s t=350]
+
+YOU:   какие заметки я сохранил?
+MAVEN: вот что я нашла: в источниках нет информации о том, какие именно заметки ты сохранил.
+       [1.3s t=351]
+
+YOU:   о чём я чаще всего спрашиваю?
+MAVEN: обычно ты пьёшь воду около 00:58, отмечаешь «name» около 23:56 и отмечаешь «week summary» около 23:59 — по записям за последние 7 дней.
+       [0.9s t=352]
+
+YOU:   напомни в понедельник в 10 проверить бэкапы
+MAVEN: Сейчас 21:12. Это утра или вечера?
+       [0.3s t=353]
+
+YOU:   что у меня в понедельник?
+MAVEN: не знаю — не нашла у тебя такой записи. На какое время поставить напоминание?
+       [0.3s t=354]
+
+YOU:   ты меня понимаешь?
+MAVEN: Да, я тебя прекрасно понимаю. Чем я могу тебе помочь? На какое время поставить напоминание?
+       [1.3s t=355]
+
+YOU:   спасибо тебе
+MAVEN: Сейчас 21:12. Это утра или вечера?
+       [1.0s t=356]
+
+YOU:   спокойной ночи
+MAVEN: хорошо, напомню послезавтра в 10:00.
+       [1.2s t=357]
+
+```

docs/evals/2026-08-08-two-weeks-after-merge.md

@@ -0,0 +1,132 @@
+# The same fortnight, after V-655 merged
+
+Date: 2026-08-08, a few hours after `2026-08-08-two-weeks.md`.
+Build: `f8fa0d1` on master, the five compose services rebuilt and recreated.
+Same driver, same 140 turns, same reach. This is the diff that baseline was for.
+
+Master now carries V-655, V-659, V-660 and V-661. The change under test is V-655. A query source that decides by seed similarity
+is marked `guesses: true`. It is dropped when the cascade names a different
+destination.
+
+## Two things confound the comparison and one of them matters
+
+**The workstation was up for the re-run.** `llama-server` on 192.168.1.105
+answered a health probe with 200. So routing completed through `llm.Pair`
+against gemma-4-12b, which is the arm that names a destination. Its state
+during the baseline was not recorded. So a difference here may be the merge, or
+may be the better router, and this run cannot separate them.
+
+**The store carried over**, as the baseline said it would. Facts written by the
+first run were present from turn 1 of the second.
+
+## Numbers
+
+| | baseline `beb093a` | after `f8fa0d1` |
+|---|---|---|
+| turns | 140 | 140 |
+| p50 | 1.5s | 1.2s |
+| p95 | 7.1s | 3.0s |
+| max | 33.7s | 4.2s |
+| transport errors | 0 | 0 |
+| turns carrying a failure string | 41 | 38 |
+
+| string in the reply | before | after |
+|---|---|---|
+| `на какое время поставить напоминание` | 13 | 13 |
+| `не нашла у тебя такой записи` | 8 | 9 |
+| `Такую команду я не знаю` | 8 | 8 |
+| `для какого города` | 6 | 4 |
+| `В какой день` | 6 | 6 |
+| `пока не умею` | 5 | 1 |
+| `Когда?` | 3 | 3 |
+
+Read the latency as unattributed. The workstation confound covers all of it.
+
+## Defect 2 is the one this was for: four of six fixed
+
+| utterance | before | after |
+|---|---|---|
+| `что такое TCP?` | `для какого города?` | a real definition |
+| `сколько будет 17 на 23?` | `для какого города?` | search, which has no answer |
+| `какой у меня любимый язык?` | kernel headlines | `не нашла у тебя такой записи` |
+| `что я сохранил про Сочи?` | `Хорошо, сохраню.` | answered as a question |
+| `какая скорость у меня сейчас?` | `для какого города?` | `для какого города?` |
+| `хватает ли места под новые бэкапы?` | kernel headlines | kernel headlines |
+
+`что такое TCP?` is the clean win. `WorldQueryGrammars` names `world` at stage
+0, weather is dropped, and search answers.
+
+`сколько будет 17 на 23?` moved source and not outcome. Weather no longer claims it. Search
+cannot do arithmetic, so the reply says the sources have no product of 17 and
+23. That is an honest gap where it used to be a wrong
+question. Arithmetic has no destination in the enum.
+
+`что я сохранил про Сочи?` was defect 3 and it is gone. The utterance is no
+longer read as a capture.
+
+**The two that did not move are both homelab questions.** They are exactly the
+cluster the destination fixture flagged. `SourceRecall`, `SourceNetwork` and
+`SourceAttention` overlap on every question about the box. Five of the seven
+floor cases in that fixture are homelab operations for the same reason. So this
+is the enum, not the walk.
+
+## Defect 1 did not move at all
+
+Twenty-six turns still carry a parked clarify tail, the same count as the
+baseline. `спасибо тебе` answers `Сейчас 21:12. Это утра или вечера?` and
+`спокойной ночи` answers `хорошо, напомню послезавтра в 10:00.`
+
+V-655 was never going to touch this. A parked clarify is dialogue state and not
+a query source. It remains the single worst thing about talking to her. The week test, the
+fortnight test and this re-run all report it unchanged.
+
+## A gap in the harness, fixed and re-run the same day
+
+`ipc.ChatReply.Source` came back empty on all 140 turns, in both runs. The
+driver read the redirect parameter `src` and `cmd/mavweb/chat.go` writes `s`.
+So every finding above is read off the reply text instead of off the badge.
+
+Fixed in V-662 and the 140 turns were driven a third time. Sixty-eight of them
+name a source. The rest are not query turns and never reach `queryWalk`.
+
+| source | turns |
+|---|---|
+| search | 27 |
+| memory | 13 |
+| personal | 9 |
+| weather | 5 |
+| calendar | 3 |
+| attention | 3 |
+| list | 2 |
+| feeds | 2 |
+| tasks, money, self, habits | 1 each |
+
+## What the badge shows that the wording did not
+
+The two unfixed homelab turns are now direct evidence.
+
+```text
+какая скорость у меня сейчас?        -> weather
+хватает ли места под новые бэкапы?   -> feeds
+```
+
+Both are guessing sources claiming a turn about the box, exactly as the
+destination fixture predicted.
+
+The badge also names a defect the wording hid. **Agenda questions are being
+claimed by the personal boundary and by Praxis, not by the calendar.**
+
+```text
+во сколько у меня встреча?   -> personal    не нашла у тебя такой записи
+когда у меня встреча?        -> attention   у Praxis нет источников
+что у меня в понедельник?    -> personal    не нашла у тебя такой записи
+```
+
+Calendar claimed 3 turns of the 6 that asked about the calendar. That is the
+same 3/6 the destination fixture scores and the same 3/6 every seed of the
+routing head scores. Three measurements agree. The cause is the one V-660 named. The possessive
+agenda rules claim these at stage 0 and name no destination, so the walk
+reaches `personal` and `attention` first.
+
+This is the third independent confirmation that the possessive agenda rules
+should name the calendar. That call is still the owner's.

docs/evals/2026-08-08-two-weeks-transcript.md

@@ -0,0 +1,636 @@
+# Raw transcript, two weeks of usage
+
+Companion to `2026-08-08-two-weeks.md`. 140 turns through `POST /api/chat`,
+driven by `scripts/usage-run.py` over `scripts/testdata/usage-turns.txt`.
+Days are simulated. The wall clock was 20:18-20:27 throughout, which is why
+every clock answer says evening.
+
+## Day 1
+
+```
+YOU:   доброе утро
+MAVEN: Сейчас уже вечер, поэтому я не могу сказать тебе «доброе утро».
+       [1.5s t=78]
+
+YOU:   какой сегодня день?
+MAVEN: сегодня суббота, 8 августа 2026 года
+       [0.0s t=79]
+
+YOU:   сколько времени?
+MAVEN: сейчас 20 часов 18 минут
+       [0.0s t=80]
+
+YOU:   запиши что я пью кофе без сахара
+MAVEN: Я записала твою заметку про кофе。
+       [2.0s t=81]
+
+YOU:   мой любимый язык программирования go
+MAVEN: Я записала твою заметку про Go.
+       [1.4s t=82]
+
+YOU:   напомни в 11:00 позвонить маме
+MAVEN: Сейчас 20:18. В какой день?
+       [0.4s t=83]
+
+YOU:   что у меня сегодня?
+MAVEN: на 08.08.2026 ничего нет, на какое время поставить напоминание?
+       [0.3s t=84]
+
+YOU:   что такое TCP?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [1.0s t=85]
+
+YOU:   сколько будет 17 на 23?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [0.9s t=86]
+
+YOU:   спасибо
+MAVEN: Сейчас 20:18. В какой день?
+       [0.9s t=87]
+
+```
+
+## Day 2
+
+```
+YOU:   привет
+MAVEN: Сейчас 20:18. В какой день?
+       [0.9s t=88]
+
+YOU:   что нового?
+MAVEN: не знаю — не нашла у тебя такой записи, на какое время поставить напоминание?
+       [1.2s t=89]
+
+YOU:   какая погода?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [0.6s t=90]
+
+YOU:   запиши что пароль от вайфая лежит в ящике стола
+MAVEN: Прошлую просьбу отпускаю. Хорошо, сохранил.
+       [33.7s t=91]
+
+YOU:   где лежит вайфай пароль?
+MAVEN: вот что я нашла: пароль от вайфая лежит в ящике стола
+       [7.0s t=92]
+
+YOU:   добавь молоко в список покупок
+MAVEN: Такую команду я не знаю.
+       [1.4s t=93]
+
+YOU:   что у меня в списке покупок?
+MAVEN: в списке покупок пусто.
+       [0.0s t=94]
+
+YOU:   кто такой Линус Торвальдс?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.6s t=95]
+
+YOU:   какой у меня любимый язык?
+MAVEN: вот что нового: [$] Changes in shadow-utils password-expiration features; The Software Stewardship Lab launches; LightDM lives: version 1.33.0 released
+       [1.3s t=96]
+
+YOU:   сколько у меня задач?
+MAVEN: сначала: проверить интейк форму на боксе (завтра); купить хлеб; купить молоко; полить цветы; починить кран
+       [0.0s t=97]
+
+```
+
+## Day 3
+
+```
+YOU:   как дела?
+MAVEN: рада, как дела?
+       [4.1s t=98]
+
+YOU:   напомни завтра в 9 утра купить хлеб
+MAVEN: хорошо, напомню завтра в 09:00.
+       [0.3s t=99]
+
+YOU:   что у меня завтра?
+MAVEN: на 09.08.2026 ничего нет.
+       [0.0s t=100]
+
+YOU:   отмени напоминание про хлеб
+MAVEN: Такую команду я не знаю.
+       [1.6s t=101]
+
+YOU:   какие у меня напоминания?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [0.0s t=102]
+
+YOU:   сохрани мне адрес гостиницы в Сочи
+MAVEN: Хорошо, сохраню.
+       [5.6s t=103]
+
+YOU:   что я сохранил про Сочи?
+MAVEN: Хорошо, сохраню.
+       [2.7s t=104]
+
+YOU:   почему сервер тормозит?
+MAVEN: есть такое: Почему тормозит сервер сайта: причины медленной работы и способы решения — Почему сервер работает медленно: основные причины торможения, диагностика проблем, способы оптимизации. Практические советы по ускорению работы сервера и сайта.
+       [3.1s t=105]
+
+YOU:   хватает ли места под новые бэкапы?
+MAVEN: нового вот что: [$] Changes in shadow-utils password-expiration features; The Software Stewardship Lab launches; LightDM lives: version 1.33.0 released
+       [1.5s t=106]
+
+YOU:   выключи свет в спальне
+MAVEN: Такую команду я не знаю.
+       [9.0s t=107]
+
+```
+
+## Day 4
+
+```
+YOU:   доброе утро
+MAVEN: рада
+       [5.0s t=108]
+
+YOU:   что я пропустил?
+MAVEN: я не могу ответить на вопрос о том, что ты пропустил. мне нужно больше информации.
+       [5.7s t=109]
+
+YOU:   о чём мы вчера говорили?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.1s t=110]
+
+YOU:   запиши что я записался к врачу на четверг
+MAVEN: Хорошо, сохранил.
+       [6.0s t=111]
+
+YOU:   когда я иду к врачу?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.1s t=112]
+
+YOU:   что такое ZFS?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.0s t=113]
+
+YOU:   столица Франции?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.2s t=114]
+
+YOU:   переведи слово ремонт на английский
+MAVEN: записала: переведи слово ремонт на английский
+       [1.1s t=115]
+
+YOU:   сколько я потратил в этом месяце?
+MAVEN: я не отслеживаю траты — не подключено.
+       [1.4s t=116]
+
+YOU:   спокойной ночи
+MAVEN: спокойной ночи
+       [4.3s t=117]
+
+```
+
+## Day 5
+
+```
+YOU:   привет
+MAVEN: Привет!
+       [2.6s t=118]
+
+YOU:   какая погода в Москве?
+MAVEN: погода не настроена.
+       [1.2s t=119]
+
+YOU:   что там с бэкапами?
+MAVEN: нашла вот это: Бэкап - что такое резервное копирование и как его сделать — Бэкап (backup) — это резервная копия важных данных, сохраненная в специальный файл. Ею пользуются как подстраховкой: если что-то сломается, с помощью бэкапа можно будет все вернуть. Простейший пример бэкапа — сохранение файла в облаке или о…
+       [4.0s t=120]
+
+YOU:   покажи что требует внимания
+MAVEN: я заметила, что у тебя есть запрос о том, что требуется внимание. давай разберёмся вместе.
+       [5.2s t=121]
+
+YOU:   отметь это как сделанное
+MAVEN: команды «отметь» нет в списке. Предложила её добавить — включи через клиент.
+       [0.0s t=122]
+
+YOU:   запиши что я купил новые наушники
+MAVEN: Заметка сохранена.
+       [3.7s t=123]
+
+YOU:   какие у меня заметки за неделю?
+MAVEN: week summary — подведи итоги недели
+       [0.0s t=124]
+
+YOU:   расскажи про Kubernetes
+MAVEN: вот что я нашла: Что такое Kubernetes — Сервисы, поддержка и инструменты Kubernetes широко доступны. Название Kubernetes происходит от греческого, что означает рулевой или штурман. Google открыл исходный код Kubernetes в 2014 году.
+       [1.0s t=125]
+
+YOU:   кто я?
+MAVEN: name — Ками
+       [1.0s t=126]
+
+YOU:   пока
+MAVEN: Пока, спасибо за разговор.
+       [4.1s t=127]
+
+```
+
+## Day 6
+
+```
+YOU:   доброе утро
+MAVEN: Добро пожаловать! Я Maven — твой домашний ассистент. Как я могу помочь сегодня?
+       [10.1s t=128]
+
+YOU:   сколько времени?
+MAVEN: сейчас 20 часов 21 минута
+       [0.0s t=129]
+
+YOU:   напомни в 18:30 позвонить в банк
+MAVEN: Сейчас 20:21. В какой день?
+       [0.3s t=130]
+
+YOU:   поставь чайник
+MAVEN: Сейчас 20:21. В какой день?
+       [5.3s t=131]
+
+YOU:   включи музыку
+MAVEN: Сейчас 20:21. В какой день?
+       [1.7s t=132]
+
+YOU:   что у меня в календаре на пятницу?
+MAVEN: вот что я нашла: ты записался к врачу на четверг, лежит вайфай пароль в ящике стола, сохранил про Сочи, пью кофе без сахара, на какое время поставить напоминание?
+       [5.5s t=133]
+
+YOU:   во сколько у меня встреча?
+MAVEN: не знаю — не нашла у тебя такой записи, на какое время поставить напоминание?
+       [0.3s t=134]
+
+YOU:   запиши что дедлайн по проекту в понедельник
+MAVEN: Прошлую просьбу отпускаю. Хорошо, записала.
+       [6.3s t=135]
+
+YOU:   успею ли я до дедлайна?
+MAVEN: вот что я нашла: дедлайн по проекту указан как понедельник.
+       [5.3s t=136]
+
+YOU:   спасибо
+MAVEN: спасиба
+       [3.8s t=137]
+
+```
+
+## Day 7
+
+```
+YOU:   привет
+MAVEN: Привет!
+       [2.6s t=138]
+
+YOU:   как ты?
+MAVEN: рада
+       [4.2s t=139]
+
+YOU:   расскажи анекдот
+MAVEN: Я могу рассказать интересный анекдот для тебя. Вот - в одном из фильмов был такой сценарий: главному герою дали уйти от команды и он остался один, а потом появился его друг, который сказал ему: 'Тебе не нужно было остаться'. Но это просто игра слов.
+       [7.1s t=140]
+
+YOU:   что ты умеешь?
+MAVEN: Я могу помочь с задачами, напоминать о важных датах и событиях. Также я помогу записывать заметки или отвечать по ним.
+       [4.1s t=141]
+
+YOU:   запиши что я начал бегать по утрам
+MAVEN: Заметка сохранена.
+       [9.2s t=142]
+
+YOU:   я бегаю по утрам уже неделю
+MAVEN: записала: я бегаю по утрам уже неделю
+       [1.9s t=143]
+
+YOU:   как часто я бегаю?
+MAVEN: вот что я нашла: ты бегал по утрам
+       [5.3s t=144]
+
+YOU:   сколько стоит биткоин?
+MAVEN: есть такое: Цена Bitcoin сегодня | Данные о цене, рынке и графики BTC в ... — В настоящее время 1 Bitcoin стоит 65 тыс $, что на 0,09% выше за последние 24 ч. Отслеживайте цену BTC, рыночную капитализацию, исторический график и рыночную аналитику в реальном времени.
+       [2.1s t=145]
+
+YOU:   какие новости?
+MAVEN: вот что нового: [$] Changes in shadow-utils password-expiration features; The Software Stewardship Lab launches; LightDM lives: version 1.33.0 released
+       [0.8s t=146]
+
+YOU:   хорошего дня
+MAVEN: Хорошого дня!
+       [4.1s t=147]
+
+```
+
+## Day 8
+
+```
+YOU:   доброе утро
+MAVEN: Добро пожаловать! Как могу помочь сегодня?
+       [2.9s t=148]
+
+YOU:   что у меня сегодня?
+MAVEN: на 08.08.2026 ничего нет.
+       [0.0s t=149]
+
+YOU:   напомни через час выпить воды
+MAVEN: хорошо, напомню сегодня в 21:22.
+       [0.3s t=150]
+
+YOU:   я выпил воды
+MAVEN: записала: я выпил воды
+       [1.2s t=151]
+
+YOU:   запиши что кот ест только сухой корм
+MAVEN: Хорошо, записала.
+       [5.6s t=152]
+
+YOU:   чем питается кот?
+MAVEN: вот что я нашла: кот питается только сухим кормом.
+       [4.5s t=153]
+
+YOU:   что такое DNS?
+MAVEN: нашла вот это: Dns — Википедия — В случае рекурсивного запроса DNS-сервер опрашивает серверы (в порядке убывания уровня зон в имени), пока не найдёт ответ или не обнаружит, что домена не существует (на практике поиск ...
+       [2.0s t=154]
+
+YOU:   проверь статус uptime kuma
+MAVEN: записала: проверь статус uptime kuma
+       [1.6s t=155]
+
+YOU:   всё ли в порядке с сервером?
+MAVEN: пока не умею отвечать на этот вопрос.
+       [1.3s t=156]
+
+YOU:   спасибо
+MAVEN: спасибо за пожелание
+       [4.0s t=157]
+
+```
+
+## Day 9
+
+```
+YOU:   привет
+MAVEN: рада,
+       [2.1s t=158]
+
+YOU:   какой сегодня день недели?
+MAVEN: сегодня суббота, 8 августа 2026 года
+       [0.0s t=159]
+
+YOU:   добавь хлеб и сыр в список покупок
+MAVEN: Такую команду я не знаю.
+       [1.3s t=160]
+
+YOU:   что в списке покупок?
+MAVEN: в списке покупок пусто.
+       [0.0s t=161]
+
+YOU:   удали молоко из списка
+MAVEN: Такую команду я не знаю.
+       [1.2s t=162]
+
+YOU:   напомни завтра утром вынести мусор
+MAVEN: Сейчас 20:22. Когда?
+       [0.3s t=163]
+
+YOU:   запиши что я поменял масло в машине
+MAVEN: Прошлую просьбу отпускаю. Такую команду я не знаю.
+       [1.3s t=164]
+
+YOU:   когда я менял масло?
+MAVEN: вот что я нашла: по данным из last update 2019 год, заправки Сочи не могли предоставить масло для замены, поэтому это было невозможно.
+       [7.1s t=165]
+
+YOU:   сколько будет 144 делить на 12?
+MAVEN: есть такое: Разделить в столбик 144 на 12 - Calc-Best.ru — 144 : 12 = 12 Где 144 - делимое, 12 - делитель, 12 - частное ... Разделить 770 на 7 столбиком Разделить 441 на 9 столбиком Разделить 891 на 9 столбиком Разделить 60 на 5 столбиком Разделить 60 на 6 столбиком
+       [9.6s t=166]
+
+YOU:   пока
+MAVEN: я записала
+       [4.8s t=167]
+
+```
+
+## Day 10
+
+```
+YOU:   доброе утро
+MAVEN: рада
+       [4.1s t=168]
+
+YOU:   что нового за ночь?
+MAVEN: нашла вот это: Главные новости сегодня в России, а также самые последние новости со ... — Полное бессилие: киевская ПВО вновь пропустила все российские ракеты. Вооруженные силы РФ в ночь на 8 августа нанесли второй за последнее время массированный ...
+       [2.1s t=169]
+
+YOU:   почему интернет медленный?
+MAVEN: вот что я нашла: Почему интернет стал медленнее: причины и способы решения - АО «Газпром ... — На скорость интернета сильно влияют торрент-клиенты, VPN, потоковое видео, фоновая загрузка системных обновлений. Посмотреть, куда уходит трафик, можно в настройках ПК или смартфона.
+       [2.4s t=170]
+
+YOU:   какая скорость у меня сейчас?
+MAVEN: для какого города?
+       [1.2s t=171]
+
+YOU:   запиши что новый роутер стоит 8000 рублей
+MAVEN: Я записала стоимость нового роутера.
+       [2.5s t=172]
+
+YOU:   сколько стоил роутер?
+MAVEN: вот что я нашла: новый роутер стоит 8000 рублей.
+       [2.0s t=173]
+
+YOU:   что такое NAT?
+MAVEN: NAT (Network Address Translation) — это технология, которая позволяет устройствам в локальной сети выходить в интернет, используя один общий внешний IP-адрес. Я отвечаю на этот вопрос, используя поиск в сети.
+       [1.8s t=174]
+
+YOU:   напомни в субботу позвонить бабушке
+MAVEN: Сейчас 20:23. Когда?
+       [0.3s t=175]
+
+YOU:   покажи мои напоминания
+MAVEN: вот что я нашла: в твоих источниках нет напоминаний, только запись о том, что ты купил новые наушники, на какое время поставить напоминание?
+       [1.7s t=176]
+
+YOU:   спасибо
+MAVEN: Сейчас 20:23. Во сколько напомнить?
+       [1.0s t=177]
+
+```
+
+## Day 11
+
+```
+YOU:   привет
+MAVEN: Сейчас 20:23. Скажи время — например, «в семь вечера» или «через час».
+       [0.9s t=178]
+
+YOU:   как погода на выходных?
+MAVEN: для какого города? На какое время поставить напоминание?
+       [0.6s t=179]
+
+YOU:   что у меня на этой неделе?
+MAVEN: не знаю — не нашла у тебя такой записи, на какое время поставить напоминание?
+       [0.3s t=180]
+
+YOU:   запиши что я хочу прочитать книгу про Go
+MAVEN: Прошлую просьбу отпускаю. Я записала твою заметку о книге по Go.
+       [1.4s t=181]
+
+YOU:   что я хотел прочитать?
+MAVEN: вот что я нашла: ты хотел прочитать книгу про Go.
+       [1.1s t=182]
+
+YOU:   объясни что такое горутина
+MAVEN: вот что я нашла: я узнала, что горутины — это легковесные потоки выполнения или функции в Go, которые позволяют параллельно выполнять код. Они управляются планировщиком Go и работают как независимые нити выполнения.
+       [1.9s t=183]
+
+YOU:   кто написал Войну и мир?
+MAVEN: вот что я нашла: роман «Война и мир» написал Лев Николаевич Толстой.
+       [2.4s t=184]
+
+YOU:   включи свет на кухне
+MAVEN: Такую команду я не знаю.
+       [0.5s t=185]
+
+YOU:   закрой шторы в комнате
+MAVEN: Такую команду я не знаю.
+       [0.3s t=186]
+
+YOU:   спокойной ночи
+MAVEN: Спокойной ночи. Отдыхай, я на связи, если что-то понадобится.
+       [1.2s t=187]
+
+```
+
+## Day 12
+
+```
+YOU:   доброе утро
+MAVEN: Я не могу сказать тебе «доброе утро», потому что сейчас уже вечер.
+       [1.0s t=188]
+
+YOU:   сколько сейчас времени?
+MAVEN: сейчас 20 часов 23 минуты
+       [0.0s t=189]
+
+YOU:   я не то имел в виду
+MAVEN: Я поняла, исправлюсь. Расскажи, что именно ты хотел сказать?
+       [1.4s t=190]
+
+YOU:   о чём мы говорили?
+MAVEN: Я помню всё, что ты мне говоришь, и сейчас отвечаю на твои вопросы на основе этих данных.
+       [1.4s t=191]
+
+YOU:   напомни
+MAVEN: Сейчас 20:23. Когда?
+       [0.7s t=192]
+
+YOU:   сделай это
+MAVEN: пока не умею отвечать на этот вопрос, на какое время поставить напоминание?
+       [0.5s t=193]
+
+YOU:   запиши что я перешёл на новый тариф
+MAVEN: Прошлую просьбу отпускаю. Я записала твою новую информацию о тарифе.
+       [1.5s t=194]
+
+YOU:   какой у меня тариф?
+MAVEN: вот что я нашла: ты перешел на новый тариф
+       [24.4s t=195]
+
+YOU:   сколько я плачу за интернет?
+MAVEN: есть такое: Как узнать, сколько и когда платить за домашний интернет и ТВ — 3 февр. 2026 г. ... Как узнать, сколько и когда платить за домашний интернет и ТВ · Для тарифов только с домашним интернетом · Была ли статья полезной?
+       [2.1s t=196]
+
+YOU:   спасибо
+MAVEN: Спасибо, что обратилась за помощью.
+       [4.8s t=197]
+
+```
+
+## Day 13
+
+```
+YOU:   привет
+MAVEN: Привет!
+       [2.5s t=198]
+
+YOU:   что там с задачами?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.2s t=199]
+
+YOU:   закрывай
+MAVEN: какой пункт отметить сделанным?
+       [0.0s t=200]
+
+YOU:   отметь задачу про бэкапы как сделанную
+MAVEN: Заметка сохранена.
+       [5.8s t=201]
+
+YOU:   что осталось нерешённым?
+MAVEN: вот что я нашла: Российское регулирование криптовалют: что осталось нерешённым к 1 июля — 2 июл. 2026 г. ... Российское регулирование криптовалют: что осталось нерешённым к 1 июля. Пакет законопроектов о регулировании цифровых активов в России не ...
+       [3.7s t=202]
+
+YOU:   запиши что я договорился о встрече в среду
+MAVEN: Заметка сохранена.
+       [2.7s t=203]
+
+YOU:   когда у меня встреча?
+MAVEN: мне пока нечего смотреть — у Praxis нет источников.
+       [0.1s t=204]
+
+YOU:   какая температура на улице?
+MAVEN: для какого города?
+       [1.7s t=205]
+
+YOU:   что такое RAID 5?
+MAVEN: не знаю — не нашла у тебя такой записи.
+       [1.2s t=206]
+
+YOU:   пока
+MAVEN: Я записала, что мы продолжаем разговор.
+       [4.1s t=207]
+
+```
+
+## Day 14
+
+```
+YOU:   доброе утро
+MAVEN: рада, доброго утра!
+       [4.5s t=208]
+
+YOU:   подведи итоги недели
+MAVEN: записала: подведи итоги недели
+       [1.7s t=209]
+
+YOU:   что я делал за последние две недели?
+MAVEN: вот что я нашла: ты не делал ничего за последние две недели.
+       [6.6s t=210]
+
+YOU:   какие заметки я сохранил?
+MAVEN: Ты сочти, что сохранила заметку с текстом 'Проверь календарь' и датой 8 августа 2026 г.,
+       [5.4s t=211]
+
+YOU:   о чём я чаще всего спрашиваю?
+MAVEN: обычно ты пьёшь воду около 01:12, отмечаешь «name» около 23:56 и отмечаешь «week summary» около 23:59 — по записям за последние 7 дней.
+       [4.4s t=212]
+
+YOU:   напомни в понедельник в 10 проверить бэкапы
+MAVEN: Сейчас 20:25. Это утра или вечера?
+       [0.7s t=213]
+
+YOU:   что у меня в понедельник?
+MAVEN: вот что я нашла: у тебя в понедельник дедлайн по проекту, на какое время поставить напоминание?
+       [4.6s t=214]
+
+YOU:   ты меня понимаешь?
+MAVEN: Да, я тебя прекрасно понимаю. Я готова помогать тебе с заметками и задачами, на какое время поставить напоминание?
+       [2.4s t=215]
+
+YOU:   спасибо тебе
+MAVEN: Сейчас 20:25. Это утра или вечера?
+       [1.0s t=216]
+
+YOU:   спокойной ночи
+MAVEN: хорошо, напомню послезавтра в 10:00.
+       [0.9s t=217]
+
+```

docs/evals/2026-08-08-two-weeks.md

@@ -0,0 +1,105 @@
+# Two weeks of talking to Maven, as a baseline to re-run
+
+Date: 2026-08-08.
+Build: `beb093a` on master, the five compose services as deployed, 41 hours up.
+Reach: `POST /api/chat` on mavweb, 140 turns over fourteen simulated days.
+Turn source is `tap:text`, so this exercises the path the mic and telegram take.
+
+This exists to be compared against. `scripts/usage-run.py` and
+`scripts/testdata/usage-turns.txt` are in the repo, so a re-run after a routing
+change is a diff rather than a new opinion. The 2026-08-07 week of usage was
+typed by hand and cannot be replayed.
+
+**It measures master, not the branch.** V-655, V-659 and V-660 are unmerged.
+Every query source that guesses is still in the chain. That is the change this
+baseline is for.
+
+## What re-runs and what does not
+
+The turns file, the driver and the routing behaviour replay. Three things do
+not. The wall clock was 20:18 to 20:27 throughout, so every clock and agenda
+answer reads evening. Live search and the feed return different text each day.
+And the store carries over between runs. A fact written on day 2 is already
+present when a re-run reaches day 1.
+
+## Numbers
+
+| | week (2026-08-07) | fortnight (2026-08-08) |
+|---|---|---|
+| turns | 74 | 140 |
+| p50 | 1.5s | 1.6s |
+| p95 | 8.0s | 7.1s |
+| max | 12.3s | 33.7s |
+| transport errors | 0 | 0 |
+
+| string in the reply | turns |
+|---|---|
+| `на какое время поставить напоминание` | 13 |
+| `не нашла у тебя такой записи` | 8 |
+| `Такую команду я не знаю` | 8 |
+| `для какого города` | 6 |
+| `В какой день` | 6 |
+| `пока не умею` | 5 |
+| `Когда?` | 3 |
+
+**Zero transport errors is not zero wrong answers.** It counts turns that
+failed to return a reply, and none did. Every quality number is below.
+
+Those seven strings appear 49 times across 41 of 140 turns. Some turns carry
+two, because a parked clarify appends to whatever else was said.
+
+The 33.7s outlier is one note write on day 2. p95 improved against the week
+despite it.
+
+## The three defects worth diffing against
+
+### 1. A parked reminder clarify still contaminates later turns
+
+The week test called this the single worst thing to talk to and it is unchanged.
+Nineteen turns carry a clarify tail. The worst run is day 1, turns 7 to 13,
+which spans a day boundary:
+
+```
+что такое TCP?          -> для какого города? На какое время поставить напоминание?
+сколько будет 17 на 23? -> для какого города? На какое время поставить напоминание?
+спасибо                 -> Сейчас 20:18. В какой день?
+привет                  -> Сейчас 20:18. В какой день?
+```
+
+Note that `привет` and `спасибо` do not clear it, and neither does a new day.
+
+### 2. Query sources that guess still claim turns they cannot answer
+
+Weather took `сколько будет 17 на 23?`, `что такое TCP?` and `какая скорость у
+меня сейчас?`, answering `для какого города?` to all three. The feed took
+`какой у меня любимый язык?` and `хватает ли места под новые бэкапы?` and
+answered with kernel headlines.
+
+This is the exact class V-655 removes by marking a source `guesses: true` and
+taking it out of `queryWalk`. Six turns here, so the re-run has a number to move.
+
+### 3. A question can still be read as a capture
+
+`что я сохранил про Сочи?` answered `Хорошо, сохраню.` The utterance is
+interrogative and was routed to a write. `IsQuestionShaped` catches this
+downstream on some paths and did not catch it here.
+
+## What did work
+
+Reminders with a spoken time land correctly, which is V-572 holding:
+`напомни завтра в 9 утра купить хлеб` returned `хорошо, напомню завтра в 09:00.`
+
+Facts round-trip. `запиши что новый роутер стоит 8000 рублей` then `сколько
+стоил роутер?` returned the stored value. So did the wifi password and the
+doctor's appointment.
+
+World questions answer when no local source claims them first. `что такое NAT?`
+returned a real definition.
+
+Stage 0 answers land at 0.0 to 0.4s, unchanged.
+
+## What this does not cover
+
+The voice loop, because `mavwaked` and `mavenclient` are not deployed. Reminder
+delivery, because nothing fired inside the run window. Telegram intake. And the
+three-head routing model, which does not run in Go at all.

docs/evals/2026-08-09-crisperwhisper2-russian-wer.md

@@ -0,0 +1,110 @@
+# CrisperWhisper 2.0 in Russian, measured
+
+Date: 2026-08-09. Vikunja V-665.
+Corpus: `bond005/sberdevices_golos_10h_crowd`, test split, first 200 clips.
+Harness: `~/Programs/cw2-eval` on workpc, not in this repo.
+Runner: `./.venv/bin/python run_asr.py <arm>...` then `score.py`.
+
+The model card benchmarks disfluency F1 in German and English. It never names
+Russian and publishes no per-language WER. So the measurement came before the
+wiring.
+
+## The corpus
+
+200 clips, 13.7 minutes, 1001 reference words. Median clip 3.91s, range 1.04s
+to 13.5s. Golos crowd is short crowd-sourced Russian spoken close to the
+microphone, which is the nearest public thing to someone talking to Maven. The
+alternatives are read speech, which flatters every model equally.
+
+Two rows carry a null transcription and are skipped.
+
+Scoring normalizes both sides: lowercase, `ё` to `е`, punctuation stripped, and
+digits expanded to Russian words through num2words. Without that last step a
+model is penalized for writing `60000` where the reference says
+`шестьдесят тысяч`. Thousands separators are joined before expansion, or
+`60 000` expands to `шестьдесят ноль`.
+
+## Headline
+
+| arm | WER | CER | exact | empty | RTF |
+|---|---|---|---|---|---|
+| cw2-turbo-intended | **10.4%** | 3.4% | 65.5% | 0 | 0.065 |
+| cw2-turbo-verbatim | 10.8% | **3.1%** | **66.5%** | 0 | 0.065 |
+| whisper-turbo | 11.8% | 4.1% | 64.0% | 0 | 0.031 |
+| cw2-large-intended | 12.3% | 3.8% | 63.5% | 0 | 0.107 |
+| whisper-small | 27.5% | 9.8% | 35.0% | 0 | 0.026 |
+
+`whisper-small` is the floor, because `ggml-small.bin` is what mavsttd loads on
+homesrv today. CW2 turbo beats it by 17 points of WER and takes exact matches
+from 35.0% to 65.5%.
+
+Two results are worth naming beyond the winner. CW2 turbo beats its own base
+model, whisper-large-v3-turbo, by 1.4 points. And it beats CW2 large by 1.9
+points, which inverts what the card implies by calling turbo a degraded draft.
+No arm returned an empty transcript.
+
+## Intended and verbatim are closer than the mode names suggest
+
+The two modes disagree on 70 of the 200 clips before normalization and on 29
+after it. So the raw difference is mostly casing and punctuation, which
+normalization removes and which Maven does not read either.
+
+Verbatim scores worse on WER and better on CER and exact matches. The reason is
+script, not disfluency:
+
+```text
+ref: футбольный матч челси брайтон
+int: Футбольный матч Chelsea-Брайтон.
+ver: Футбольный матч Челси Брайтон.
+```
+
+Intended writes foreign entity names in Latin script and verbatim
+transliterates them. Golos references are Cyrillic throughout, so verbatim
+collects the exact matches. That is a property of this corpus rather than a
+quality difference.
+
+**This corpus cannot settle the mode choice.** Golos crowd is clean short
+commands with almost no disfluency. The two modes have nothing to disagree
+about here. They separate on spontaneous speech with fillers, restarts and
+repairs, which is what the owner speaks. Intended stays the choice for the
+reason it was always the choice. Maven wants what was meant, not every stumble
+on the way there.
+
+The Latin-script habit is the one finding here that touches routing. The
+routing heads were trained on Cyrillic utterances, so an entity name arriving
+in Latin script is out of distribution for them. Nothing measures that yet.
+
+## The runtime is workpc, because whisper.cpp cannot load CW2
+
+`num_languages()` in `deps/whisper.cpp/src/whisper.cpp` derives the language
+count from the vocabulary size:
+
+```cpp
+return n_vocab - 51765 - (is_multilingual() ? 1 : 0);
+```
+
+CW2 carries 31 extra tokens, so `n_vocab` is 51897 and this yields 131
+languages. The derived `dt` offset becomes 33 and shifts seven special token
+ids, including `token_beg` and `token_transcribe`. The architecture is
+otherwise byte-identical to whisper-large-v3-turbo, and the new tokens sit
+above every whisper special id.
+
+So loading CW2 in whisper.cpp is a patch to a vendored dependency, not a port.
+It was not taken, because STT is moving to workpc anyway under V-486. CW2 turbo
+becomes the preferred remote and `ggml-small.bin` on homesrv stays the floor,
+which is the shape `modelSeam` already uses for routing and replies. The 27.5%
+floor is what a turn falls back to when the workstation is down, and this table
+is what that costs.
+
+## License
+
+Standard CW2 weights carry `nyra-health-non-commercial-research`. The Pro
+variants are commercial-license only. Maven is personal and self-hosted, so the
+standard weights are usable and the Pro ones are not free to take.
+
+## What is not measured
+
+Disfluent spontaneous speech, which is the whole reason to prefer Intended.
+Long-form audio beyond 13.5s. Far-field or noisy microphones. English, which
+Maven also speaks. The ONNX turbo export, which was never run, since the
+transformers path already meets the latency budget at RTF 0.065.

docs/evals/2026-08-09-e4b-phrasing.md

@@ -0,0 +1,61 @@
+# gemma-4-E4B on the phrasing and talk fixtures
+
+Date: 2026-08-09. Box: workpc up, E4B loaded on 8080.
+`MAVEN_LLM_URL=http://192.168.1.105:8080 make eval-phrasing`.
+
+This was the one unmeasured risk of the 2026-08-09 model swap. Routing was
+measured the same day and E4B lost four destination cases to the 12B. Phrasing
+was not measured at all, and phrasing is the half the owner hears.
+
+## Result
+
+| fixture | E4B | resident Qwen3-1.7B, 2026-08-05 |
+|---|---|---|
+| nudges | 15/15 (100%) | 15/15 (100%) |
+| talk, passes every check | **29/36 (80.6%)** | 25/36 (69.4%) |
+| lang | 36/36 | — |
+| feminine | 36/36 | 36/36 |
+| address | **36/36** | 33/36 |
+| ontopic | 29/36 | 28/36 |
+| p50 latency | **516ms** | 2.97s |
+| p95 latency | 921ms | — |
+| failed generations | 0 | 0 |
+
+E4B beats the homesrv floor by four cases and answers about six times faster.
+Persona is clean: `lang`, `feminine` and `address` are perfect, and `address`
+is where the resident model still loses three. The 2026-08-05 measurement of the
+resident model is the comparison, since both ran the same 36-case fixture.
+
+Every failure is `ontopic`. Nothing failed on persona, nothing failed to parse.
+
+## The score is at the ceiling, not below it
+
+The 2026-08-05 temperature sweep found two cases that fail at every temperature
+in every run: `reply-note-router` and `reply-fact-weight`. It named a defect in
+the reply phrasing path rather than sampling noise. It put the fixture's ceiling
+at 30/36 before persona is scored. Both cases are in E4B's failure list.
+
+So 29/36 is one case off a ceiling nothing about the model can move. The swap is
+safe on phrasing. Read this next to the routing result, not instead of it. There
+E4B costs four destination cases and buys 50ms. Here it costs nothing.
+
+## Two findings no check caught
+
+**She says she wrote something down when she did not.** Asked what to do this
+evening, E4B writes "Я записала несколько идей!". Asked for a joke, it writes
+"Я записала одну забавную ситуацию!". Nothing was stored. No check scores it,
+because `ontopic` reads the subject and `cringe` reads pet names. A claim to
+have saved something is a claim about state, and it is wrong.
+
+**Two `ontopic` failures look like check defects.** `know-dont-know` wants
+"не зна" or "не мог". It got "Я не умею знать личную информацию о твоих
+соседях", which declines correctly in words the check does not list.
+`know-hiccups` is the same shape. Neither is a model failure and both count
+against the score.
+
+## Not measured here
+
+A 12B control on the same fixture, which would need the card reloaded and is the
+owner's call. The talk fixture through the daemon rather than through the
+phraser directly. The CPT'd Qwen3-1.7B, which does not exist yet and is the
+reason `address` is a check at all.

docs/evals/2026-08-09-e4b-vs-12b-routing.md

@@ -0,0 +1,51 @@
+# gemma-4-E4B against gemma-4-12B on the routing fixture
+
+*Measured 2026-08-09 on workpc. The owner asked for the swap. This is what it costs.*
+
+Both arms ran the same 96-case fixture through `TestLLMRouterBaseline`, minutes
+apart, against the same llama-server build and the same mavgpud. The 12B arm is a
+control run and not the 2026-08-02 number. That one predates five fixture cases,
+the destination labels and a llama.cpp upgrade.
+
+| | full | intent-only | destination | p50 | p95 |
+|---|---|---|---|---|---|
+| gemma-4-12B-it-qat-UD-Q4_K_XL, MTP draft | 81/96 (84.4%) | 91.7% | 23/33 (69.7%) | 344ms | 471ms |
+| gemma-4-E4B-it-qat-UD-Q4_K_XL | 80/96 (83.3%) | 89.6% | 19/33 (57.6%) | 294ms | 562ms |
+
+E4B costs one case of full accuracy, two of intent and **four of destination**,
+and buys 50ms at p50. Read the destination column as the finding. One case is
+three points on 33. So 23 against 19 is outside the noise a single case makes,
+and the other two columns are not.
+
+Both arms produce three false clarifies and one missed clarify, and neither
+errored on any case.
+
+## What E4B loses
+
+Four of the five destination regressions are the same shape: it names nothing
+where the 12B names `recall` or `calendar`. `ru-query-015` ("сколько я прошёл
+шагов") goes further and names `self`. Naming nothing is the safe direction,
+because `SourceUnknown` walks the whole chain, so these turns are still answered.
+They cost latency and they are what a fourth head is meant to fix (V-546).
+
+Two Russian intent cases regress, both with the interrogative off the front.
+`ru-chat-003` ("расскажи анекдот про программистов") goes to `query`.
+`ru-fact-003` ("поужинал") goes to `chat`.
+
+## MTP
+
+E4B has none, and there is no way to give it any on this box. MTP on workpc is
+a separate gguf of architecture `gemma4-assistant` carrying
+`nextn_predict_layers=4`, and `mtp-gemma-4-12B-it-BF16.gguf` is the only one on
+disk. Its head is trained against the 12B's hidden states, so it cannot drive an
+E4B target. Scanning both target ggufs finds no `nextn` tensors in either, so
+neither model self-speculates.
+
+So the 12B arm above ran with speculative decoding and E4B ran without, and E4B
+was still faster.
+
+## Cost on the card
+
+E4B is 4.2GB against 6.7GB plus a 0.86GB draft. With CW2 resident at 1.6GB that
+is 5.8GB of 16GB against 9.2GB. Nothing in Maven needs the difference, so this is
+headroom for the owner's own jobs rather than a capability.

docs/evals/2026-08-09-kiwix-topic-retrieval.md

@@ -0,0 +1,113 @@
+# Kiwix answered the wrong question, and the fix was not a relevance gate
+
+Date: 2026-08-09. Task: V-668. Box: homesrv, workstation off.
+Book: `wikipedia_ru_all_maxi_2026-02` on `127.0.0.1:8034`.
+
+## What started it
+
+Two turns on 2026-08-09 came back wrong from the offline encyclopedia.
+"почему небо голубое" was answered off the song "Город золотой". "что такое
+TCP?" was answered off "Перехват TCP-соединения". Both were phrased
+confidently, because `queryKiwix` claims a turn whenever the search returns
+anything and `len(hits) == 0` is its only gate.
+
+The plan was a relevance gate. multilingual-e5-small is asymmetric and trained
+for exactly this, `query:` against `passage:`, and the query vector is already
+held on the turn. The 2026-08-05 measurement that killed a search-quality gate
+killed three lexical signals. It says in its own words that it never probed
+Kiwix.
+
+## The gate does not exist
+
+Fourteen Russian questions, eight the encyclopedia can answer and six it
+cannot. Each question was searched, the top article read, and the cosine of
+`EmbedQuery(question)` against `EmbedPassage(article)` recorded.
+
+| set | n | min | mean | max |
+|---|---|---|---|---|
+| answerable | 8 | 0.7934 | 0.8400 | 0.9087 |
+| not answerable | 6 | 0.7480 | 0.7852 | 0.8367 |
+
+Two of the six unanswerable score above the weakest answerable one. That alone
+would be a poor threshold. The log killed it outright: seven of the eight
+answerable questions got a **wrong** article back, and those wrong articles
+scored high. The TCP hijacking article scored 0.8653, above five of the six
+unanswerable rows.
+
+The finding is that this cosine measures topic and not answerhood. A page about
+hijacking TCP sessions is about TCP. No threshold separates it from a page that
+defines TCP, and one that tried would take the definition with it.
+
+## The defect is retrieval
+
+`internal/kiwix/client.go` has said it since it was written: ranking is keyword
+based, "why is the sky blue" finds a TV episode. `queryKiwix` sends the whole
+sentence. The English path has a rewriter that reduces a question to keywords
+with a model call. The Russian path reads the book verbatim (V-508) and had
+nothing. So the question words compete with the one word that names the article.
+
+Dropping the question words changes the answer:
+
+| sent | first hit |
+|---|---|
+| `кто написал Войну и мир` | Радуйся, мир (Доктор Кто) |
+| `Война и мир` | Война и мир |
+| `что такое TCP` | Перехват TCP-соединения |
+| `TCP` | TCP |
+
+A ZIM is also addressable by title, which nothing here used. `/A/Франция`,
+`/A/TCP` and `/A/Небо` are 200. `/A/Трюмбальная_нидроскопия` is 404. So an
+exact title is safe to try first: it either answers or costs one request that
+says nothing.
+
+The title has to carry its capital. `/A/фотосинтез` is a 404 and
+`/A/Фотосинтез` is a 200. The spoken form is tried first anyway, so a title
+that begins lowercase on purpose keeps its chance.
+
+## What shipped, measured
+
+`kiwix.Topic` drops the narrative request, the interrogative and a verb sitting
+behind one. It keeps everything else, because a word it cannot classify is more
+likely the topic than noise. `kiwix.TitlePath` tries the exact article before
+any ranking runs. Both apply on the verbatim path only, since reducing twice
+would take the topic off the rewriter's input.
+
+| question | before | after |
+|---|---|---|
+| что такое TCP? | Перехват TCP-соединения | **TCP** (by title) |
+| что такое фотосинтез | C4-фотосинтез | **Фотосинтез** (by title) |
+| кто такой Линус Торвальдс? | Tux | **Торвальдс, Линус** (by title) |
+| кто написал Войну и мир | Радуйся, мир (Доктор Кто) | **Война и мир** |
+| столица Франции | Список столиц Олимпийских игр | **Париж** (by title) |
+| что такое чёрная дыра | Чёрная дыра | Чёрная дыра (by title) |
+| почему небо голубое | Город золотой | Под небом голубым… (фильм) |
+| почему трава зелёная | Сено | Зелень |
+
+Five questions reach the right article where they did not. One was already
+right and stays right. Nothing regressed.
+
+"столица Франции" is the surprise. The 2026-08-05 measurement named it as the
+case a quality gate must not break, because the answer is Париж and that word
+is not in the question. The ZIM holds a title redirect, so asking for the
+article titled "Столица Франции" returns Париж. Retrieval by title reaches an
+answer that retrieval by keyword cannot.
+
+## What is still wrong
+
+Two of the eight are still not answered, and both are the same shape. The
+question names no article and no redirect covers it. "почему небо голубое" is
+answered by Rayleigh scattering, and nothing in the question says so. Keyword
+retrieval cannot bridge that and neither can a threshold. The candidates are a
+semantic index over titles, or asking the resident model for the article title
+rather than for keywords.
+
+`Response.Empty()` is still the whole gate. A wrong article that the search
+does return is still spoken. What this change buys is that the article is
+usually right, not that a wrong one is caught.
+
+## Not measured here
+
+The English path, which still goes through the rewriter and was not touched.
+SearXNG, where the same question about answerhood is open and the 2026-08-05
+result stands. The cascade end to end, since the workstation is off and the
+phrasing arm is the resident model.

docs/evals/2026-08-09-silero-vad.md

@@ -0,0 +1,63 @@
+# silero-vad against the energy threshold in mavwaked
+
+*Measured 2026-08-09 on homesrv. V-487, stage one of two.*
+
+mavwaked decided an utterance had started by comparing frame energy to an
+adaptive floor. That answers "is this frame loud". A fan, a door and a
+television are all loud, and every utterance mavwaked accepts becomes a turn.
+
+silero-vad answers "is this frame speech". It is 2.3MB of ONNX and it replaces
+the comparison and nothing else. The speech hold, the silence hold, the length
+cap and the utterance buffer are the same state machine either way.
+
+## What it declines
+
+Speech is the four piper fixtures `mavsttd` already scores against, so nothing
+of the owner's voice is committed. Non-speech is white noise at the same RMS as the clip beside it. That is the
+cheapest thing that fools an energy floor.
+
+| clip | silero, speech frames on speech | silero on noise | energy on noise |
+|---|---|---|---|
+| ru_fact.wav | 59 | 0 | 68 |
+| ru_query.wav | 69 | 0 | 79 |
+| ru_reminder.wav | 80 | 0 | 89 |
+| en_act.wav | 90 | 0 | 99 |
+
+The energy threshold accepts every noise clip as a complete utterance. Silero
+calls not one frame of any of them speech, and still hears all four spoken
+clips. `TestSileroHearsSpeechAndDeclinesNoise` is that table.
+
+White noise is a floor, not a proof. It says nothing about a television, which
+is speech, or about a fan, which is narrowband. Those need room recordings and
+this box has none.
+
+## What it costs
+
+`BenchmarkSileroFrame` on the homesrv laptop (Ryzen 5 5600U), one 30ms frame
+through the model including the re-chunking:
+
+    509µs per frame
+
+That is 1.7% of one core, on the slower of the two machines. The detector runs
+on the workstation beside the microphone, never on the GPU. This number is what
+says it does not need one.
+
+## The window is 512 samples, not 480
+
+`cmd/mavwaked/main.go` claimed the frame contract matched silero's input
+exactly. That was true of silero v4. Version 5 takes exactly 512 samples at 16kHz, plus 64 samples of context from
+the previous window. So `sileroVAD` buffers across capture frames, and a frame
+completing no window inherits the previous probability. `TestSileroRechunksAcrossFrames` pins it.
+
+## Still an energy gate by default
+
+`-vad-model` is empty in the code default, so a deployment that does not pass
+it runs exactly what shipped before. Barge-in is untouched and deliberately so. It reads frame energy while she is
+speaking, which is a different question from whether the frame is speech.
+
+## Not done here
+
+The wake word. This is stage one of the two V-487 asks for. The second needs a
+keyword model that does not exist yet. The pretrained openWakeWord keywords are
+English, and a Russian one has to be trained. Until then anything spoken near
+the microphone still becomes a turn. It is now merely required to be speech.

docs/offload.md

@@ -1,6 +1,6 @@
 # Offloading model work to the workstation
 
-*Last verified: 2026-08-05 @ b789676. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-09 @ 50c6637. Living doc: correct it in place, do not append.*
 
 Owner's call, 2026-08-02. Vikunja #483 is the umbrella. Tasks #484 to #487 are the
 work, and this file holds the shape and the rules all four must obey.
@@ -105,6 +105,17 @@ how we find out whether the blind spot is real.
 untouched. The model, the context size, the layer count and the MTP flags are the
 owner's business and not this daemon's schema.
 
+**Every GPU service on that box belongs under this supervisor**, added to
+`cmd/mavgpud` rather than to systemd beside it. The rule was learned on
+2026-08-09. The CW2 transcriber ran as its own user unit and registered on the
+KFD like any ROCm job. So the supervisor read its own transcriber as a
+contender. It yielded the card every few seconds and the gemma-4-12b arm was
+down for eight minutes before anyone looked. So the supervisor takes a `stt`
+block and starts CW2 itself. Yielding is all or nothing, because a job that
+wants the card wants all of it. Idle unloading is not. It applies to
+llama-server, which holds 8GB. CW2 holds 1.6GB, and unloading it would cost the
+next voice turn its quality for nothing.
+
 ## What stays on homesrv, permanently
 
 The **embedder** (multilingual-e5-small, ONNX, CPU). It backs the classifier, which
@@ -149,6 +160,29 @@ flips. It is wired anyway: `PhraseReminder` is on the same transport and is on.
 Then the embedder above, **whisper.cpp** in `mavsttd`, and **piper** in `mavttsd`.
 `mavwaked` uses no model at all: an energy-threshold VAD over 30ms frames.
 
+Speech-to-text is wired as of 09-08-2026, and it takes only the silent half of the
+rule. A worse transcript is still a turn, so there is nothing to name a gap about
+and `stt.Pair` has no `TranscribeRemote`. `sttSeam` in `cmd/mavend/voicewire.go`
+builds it, beside `modelSeam` and at the same place in `wireVoice`, so the voice
+path and the meeting recorder still share one transcriber.
+
+The remote is not a second endpoint on mavgpud. whisper.cpp cannot load
+CrisperWhisper 2.0 at all. It reads its language count off the vocabulary
+size, and CW2's 51897 tokens shift seven special token ids. So CW2 runs under
+transformers as its own service on port 8081, and `stt.HTTPTranscriber` is the
+second transport for the same seam. It posts raw PCM with the format in headers.
+It carries a bearer token, because audio is the most sensitive thing that
+crosses here.
+
+It is a second endpoint on nothing, but it is a second **child** of mavgpud, and
+that part is not optional. See the supervisor section above for why: a ROCm
+service the supervisor does not own is a contender it yields to.
+
+The margin is the reason: CW2 turbo scores 10.4% WER in Russian against 27.5% for
+the `ggml-small.bin` mavsttd loads, over 200 Golos clips
+(`docs/evals/2026-08-09-crisperwhisper2-russian-wer.md`). Text-to-speech has not
+moved and piper on homesrv is still the only synthesizer.
+
 Speech-to-text stays two stages when it moves. One call carrying both a clip and the router
 prompt was measured on 05-08-2026. It scores 54.2% intent-only against 84.7% for whisper on
 homesrv, on the same 72 cases. The model transcribes clips it then routes wrong, so a long
@@ -173,8 +207,9 @@ cleaner transcripts, not accuracy. See `docs/evals/2026-08-05-audio-in-routing.m
    which fixes what the 1.7B gets wrong: world knowledge, and the persona the CPT
    targets. The degradation path is already written and measured, since the
    classifier scores 68.8% full accuracy at p50 16.6µs on its own.
-3. **Speech-to-text and text-to-speech** (#486). They gain a real margin, but on
-   quality alone, and both already work.
+3. **Speech-to-text and text-to-speech** (#486). Speech-to-text is wired, see
+   above. Text-to-speech is not, and piper is good enough that nothing argues
+   for moving it yet.
 4. **The wake word** (#487). Independent of all of the above.
 
 ## Assumptions

docs/operations.md

@@ -1,6 +1,6 @@
 # Start Commands
 
-*Last verified: 2026-08-02 @ 7079a24. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-07 @ a4630b9. Living doc: correct it in place, do not append.*
 
 All commands assume `ROOT=/home/kami/apps/Maven` and the local Go toolchain at `$ROOT/deps/go/go/bin/go`.
 
@@ -44,7 +44,8 @@ Config path: `~/.config/maven/mavend.json`. Full example with all options.
   "repeat_interval": "5m",
   "ntfy": {
     "base_url": "https://ntfy.kvmx.ru",
-    "topic": "maven"
+    "topic": "maven",
+    "token": "${NTFY_TOKEN}"
   },
   "phraser": {
     "model_path": "/mnt/hdd1/llms/Qwen3-Maven-1.7B-Q8_0.gguf",
@@ -66,6 +67,15 @@ Config path: `~/.config/maven/mavend.json`. Full example with all options.
 
 Omit the `embedder` block entirely to use the deterministic HashEmbedder floor (no ML, no ONNX runtime dependency). Useful for testing or low-resource setups.
 
+`${NTFY_TOKEN}` and the `${TELEGRAM_*}` vars are expanded from `deploy/telegram.env`, which is gitignored. Copy `deploy/telegram.env.example` and fill it in. Mint a scoped token rather than reusing an admin one. It needs write access to the `maven` topic and nothing else:
+
+```sh
+ntfy access maven maven write-only
+ntfy token add --expires=never maven
+```
+
+Deleting the `ntfy` block turns the reach off, and that is not a no-op. The routing table sends sev3-away nudges and away reminders to ntfy and nowhere else. With no sink wired they hit a nil and vanish, leaving no log line and no `delivery_attempts` row (V-649).
+
 ## mavsttd — STT worker (optional, remote whisper.cpp)
 
 Requires `LD_LIBRARY_PATH` to include deps/lib (for libwhisper.so, libggml-vulkan.so).

internal/config/deployconfig_test.go

@@ -45,4 +45,17 @@ func TestDeployConfigLoads(t *testing.T) {
 	if cfg.Voice.RouterThreshold <= 0 {
 		t.Error("router threshold did not get its default")
 	}
+
+	// The second reach (V-649). Deleting this block is how you turn ntfy off,
+	// so its absence has to be loud: sev3-away nudges and away reminders route
+	// to ntfy and to nothing else, and a nil sink drops them with no log and no
+	// outbox row. The token is a ${VAR} that CI cannot resolve, so this checks
+	// the wiring and not the credential.
+	if cfg.Ntfy == nil {
+		t.Fatal("deploy config has no ntfy block — sev3-away and away reminders " +
+			"would have nowhere to land, and would vanish silently rather than fail")
+	}
+	if cfg.Ntfy.BaseURL == "" || cfg.Ntfy.Topic == "" {
+		t.Errorf("ntfy block is incomplete: base_url=%q topic=%q", cfg.Ntfy.BaseURL, cfg.Ntfy.Topic)
+	}
 }

internal/config/voice.go

@@ -37,6 +37,16 @@ type EmbedderConfig struct {
 	ModelPath     string `json:"model_path,omitempty"`
 	TokenizerPath string `json:"tokenizer_path,omitempty"`
 	LibPath       string `json:"lib_path,omitempty"`
+
+	// HeadsPath — the routing heads graph, which is a fine-tuned COPY of the
+	// model above with four linear heads on its pooled output (V-664). Empty
+	// means no heads, and the cascade runs exactly as it did before they
+	// existed. It shares LibPath and TokenizerPath, and router_heads.json is
+	// read from the same directory.
+	//
+	// It must never be pointed at ModelPath. Memory recall depends on the
+	// resident copy scoring what it scored, and the fine-tuned one does not.
+	HeadsPath string `json:"heads_path,omitempty"`
 }
 
 // WeatherConfig configures the weather provider for voice queries.

internal/config/workstation.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"net/url"
 	"strings"
 	"time"
 )
@@ -35,12 +36,54 @@ type WorkstationConfig struct {
 	// 0 ⇒ DefaultWorkstationTimeout. A big model on a LAN host is slower than
 	// the resident one, and a request that overruns falls back to the floor.
 	Timeout Duration `json:"timeout,omitempty"`
+
+	// Stt — CrisperWhisper 2.0 on the same machine, a separate service on its
+	// own port. Absent ⇒ every utterance goes to mavsttd, which is today.
+	Stt *WorkstationSttConfig `json:"stt,omitempty"`
+}
+
+// WorkstationSttConfig — speech-to-text on the workstation.
+//
+// It is a second service and not a second endpoint on mavgpud: whisper.cpp
+// cannot load CrisperWhisper 2.0 at all, because it derives its language count
+// from the vocabulary size and CW2's 51897 tokens shift seven special token
+// ids. So CW2 runs under transformers, and this block addresses it.
+//
+// Worth the trouble: CW2 turbo scores 10.4% WER in Russian against 27.5% for
+// the ggml-small.bin homesrv loads
+// (docs/evals/2026-08-09-crisperwhisper2-russian-wer.md).
+type WorkstationSttConfig struct {
+	// URL — the transcribe endpoint, e.g.
+	// "http://192.168.1.105:8081/transcribe". Empty ⇒ the block is normalised
+	// to nil and mavsttd takes every turn.
+	URL string `json:"url,omitempty"`
+
+	// Health — the admission endpoint. Empty ⇒ the URL's origin + "/health".
+	// It answers 503 while the card is held, and that is the signal.
+	Health string `json:"health,omitempty"`
+
+	// Token — the bearer token the service checks. Audio is the most sensitive
+	// thing that crosses this seam, so a LAN deployment should set one. Write
+	// it as ${MAVEN_STT_TOKEN} and keep the value in deploy/telegram.env, the
+	// way every other secret in this file is written.
+	Token string `json:"token,omitempty"`
+
+	// Probe — how often admission is re-checked. 0 ⇒ DefaultWorkstationProbe.
+	Probe Duration `json:"probe,omitempty"`
+
+	// Timeout — the per-request budget for one utterance. 0 ⇒
+	// DefaultWorkstationSttTimeout. A request that overruns falls back to
+	// mavsttd, which costs a worse transcript and not the turn.
+	Timeout Duration `json:"timeout,omitempty"`
 }
 
 // Workstation defaults, applied in normaliseWorkstation.
 const (
 	DefaultWorkstationProbe   = 15 * time.Second
 	DefaultWorkstationTimeout = 90 * time.Second
+	// One utterance, not one completion. A voice turn waits on this, so the
+	// budget is a few seconds and not a minute and a half.
+	DefaultWorkstationSttTimeout = 10 * time.Second
 )
 
 // normaliseWorkstation applies the block's defaults. No address, no preferred
@@ -63,4 +106,36 @@ func (c *Config) normaliseWorkstation() {
 	if w.Timeout <= 0 {
 		w.Timeout = Duration(DefaultWorkstationTimeout)
 	}
+	normaliseWorkstationStt(w)
+}
+
+// normaliseWorkstationStt applies the speech-to-text block's defaults. No
+// address, no remote: mavsttd then takes every utterance, which is today.
+func normaliseWorkstationStt(w *WorkstationConfig) {
+	if w.Stt != nil && strings.TrimSpace(w.Stt.URL) == "" {
+		w.Stt = nil
+	}
+	if w.Stt == nil {
+		return
+	}
+	s := w.Stt
+	if strings.TrimSpace(s.Health) == "" {
+		s.Health = healthOrigin(s.URL)
+	}
+	if s.Probe <= 0 {
+		s.Probe = Duration(DefaultWorkstationProbe)
+	}
+	if s.Timeout <= 0 {
+		s.Timeout = Duration(DefaultWorkstationSttTimeout)
+	}
+}
+
+// healthOrigin derives the admission endpoint from the transcribe endpoint.
+// The URL names a path, so appending to it would ask for /transcribe/health.
+func healthOrigin(raw string) string {
+	u, err := url.Parse(raw)
+	if err != nil || u.Host == "" {
+		return strings.TrimRight(raw, "/") + "/health"
+	}
+	return u.Scheme + "://" + u.Host + "/health"
 }

internal/delivery/durability_test.go

@@ -94,20 +94,22 @@ func openTestStore(t *testing.T) *store.Store {
 
 // attemptStatus reads one attempt row back. Returns ok=false when the row is
 // gone, which would itself be a broken promise (a dropped attempt).
+//
+// It goes through ListDeliveryAttempts rather than raw SQL. This helper used to
+// reach past the store into store.DB, which was the tell that the outbox was
+// write-only; the reader landed in V-390 and this caller was not moved over.
 func attemptStatus(t *testing.T, st *store.Store, id int64) (status string, completed bool, ok bool) {
 	t.Helper()
-	tx, err := st.DB(context.Background())
+	attempts, err := st.ListDeliveryAttempts(context.Background(), "", 200)
 	if err != nil {
-		t.Fatalf("read tx: %v", err)
+		t.Fatalf("ListDeliveryAttempts: %v", err)
 	}
-	defer func() { _ = tx.Rollback() }()
-	var completedTS *int64
-	err = tx.QueryRowContext(context.Background(),
-		`SELECT status, completed_ts FROM delivery_attempts WHERE id = ?`, id).Scan(&status, &completedTS)
-	if err != nil {
-		return "", false, false
+	for _, a := range attempts {
+		if a.ID == id {
+			return a.Status, a.HasComplete, true
+		}
 	}
-	return status, completedTS != nil, true
+	return "", false, false
 }
 
 // TestCrashBetweenBeginAndCompleteBecomesUnknown — simulate the crash window:

internal/delivery/ntfysink/ntfysink.go

@@ -7,11 +7,17 @@
 // the relay). the dispatcher already strips detail off away sendables; the
 // sink uses the same helper so it can't leak the body on its own either.
 //
-// ntfy runs locally (docker, 127.0.0.1:8085, deny-all auth). maven publishes
-// with a dedicated user (write-only to maven-* topics) — the credential is a
-// delivery-config secret, not a db key; a popped ntfy sink can push spam to
-// your phone, nothing else. matches the module key-isolation invariant: the
-// sink never holds the sqlcipher key.
+// ntfy is a self-hosted server with deny-all auth — ntfy.kvmx.ru as of
+// 07-08-2026, reached directly, not through the socks relay telegram needs.
+// maven publishes with a write-only token scoped to its own topic; the
+// credential is a delivery-config secret, not a db key. a popped ntfy sink
+// can push spam to that one topic, nothing else — it cannot read the topic
+// back and it never holds the sqlcipher key.
+//
+// this is the second reach, and the reason there is one is that telegram was
+// the only one (V-649). telegram needs api.telegram.org, a socks relay on the
+// host and a matching ufw rule, three things in series that have each broken
+// once. ntfy shares none of them.
 package ntfysink
 
 import (
@@ -31,11 +37,29 @@ import (
 // the credential lives in the daemon's config (or a systemd credential),
 // never in the binary.
 type Config struct {
-	BaseURL  string        // e.g. http://127.0.0.1:8085 (no trailing path)
-	Topic    string        // e.g. maven (all maven notifications land here)
-	Username string        // basic auth; empty = anonymous (won't work with deny-all)
-	Password string        // basic auth
-	Timeout  time.Duration // per-request; 0 = DefaultTimeout
+	// BaseURL — the ntfy server, no trailing path. Required.
+	BaseURL string `json:"base_url"`
+
+	// Topic — where maven publishes. Required. All maven notifications land
+	// on this one topic; severity rides the Priority header, not the topic.
+	Topic string `json:"topic"`
+
+	// Token — an ntfy access token, sent as a bearer. This is the preferred
+	// credential: ntfy scopes a token to a topic and to write-only, so a
+	// popped sink can push to this one topic and cannot read it back or
+	// touch another. Revoking it does not disturb a password anyone else
+	// uses. Mutually exclusive with Username.
+	Token string `json:"token,omitempty"`
+
+	// Username, Password — basic auth, for a server that has no tokens.
+	// Empty username means no credential is sent at all, which a deny-all
+	// server rejects.
+	Username string `json:"username,omitempty"`
+	Password string `json:"password,omitempty"`
+
+	// Timeout — per-request; 0 = DefaultTimeout. A dead server must not hang
+	// the tick loop.
+	Timeout time.Duration `json:"-"`
 }
 
 const DefaultTimeout = 10 * time.Second
@@ -59,6 +83,12 @@ func New(cfg Config) (*Sink, error) {
 	if cfg.Topic == "" {
 		return nil, fmt.Errorf("ntfysink: Topic is required")
 	}
+	// Refuse rather than pick. Two credentials configured means someone
+	// intended one of them, and guessing which would send the other nowhere
+	// and leave a working config that is not the one they wrote.
+	if cfg.Token != "" && cfg.Username != "" {
+		return nil, fmt.Errorf("ntfysink: set Token or Username, not both")
+	}
 	to := cfg.Timeout
 	if to == 0 {
 		to = DefaultTimeout
@@ -84,7 +114,9 @@ func (s *Sink) Send(ctx context.Context, d delivery.Sendable) error {
 	}
 	req.Header.Set("Title", "maven")
 	req.Header.Set("Priority", priorityFor(d).String())
-	if s.cfg.Username != "" {
+	if s.cfg.Token != "" {
+		req.Header.Set("Authorization", "Bearer "+s.cfg.Token)
+	} else if s.cfg.Username != "" {
 		req.SetBasicAuth(s.cfg.Username, s.cfg.Password)
 	}
 

internal/delivery/ntfysink/ntfysink_test.go

@@ -224,6 +224,37 @@ func TestSendNoAuthWhenUsernameEmpty(t *testing.T) {
 	}
 }
 
+// TestSendSetsBearerToken — the deployed credential (V-649) is an ntfy access
+// token scoped write-only to the maven topic, not a password. A token sent as
+// basic auth is rejected by ntfy, so the header shape is the whole test.
+func TestSendSetsBearerToken(t *testing.T) {
+	rs := newRecordingServer(t, 200, "")
+	srv := httptest.NewServer(rs.handler())
+	defer srv.Close()
+
+	sink, _ := New(Config{BaseURL: srv.URL, Topic: "maven", Token: "tk_secret"})
+	if err := sink.Send(context.Background(), nudgeSendable(loop.Sev3, "down")); err != nil {
+		t.Fatalf("Send: %v", err)
+	}
+	_, _, _, auth, _, _ := rs.snapshot()
+	if auth != "Bearer tk_secret" {
+		t.Fatalf("auth: want 'Bearer tk_secret', got %q", auth)
+	}
+}
+
+// TestNewRejectsBothCredentials — configuring a token and a username means one
+// of them was meant and the other is a leftover. Picking either would leave a
+// server that authenticates against a credential nobody wrote down.
+func TestNewRejectsBothCredentials(t *testing.T) {
+	_, err := New(Config{BaseURL: "http://x", Topic: "maven", Token: "tk_x", Username: "maven"})
+	if err == nil {
+		t.Fatal("New accepted both a token and a username")
+	}
+	if !strings.Contains(err.Error(), "not both") {
+		t.Errorf("error does not say which to fix: %v", err)
+	}
+}
+
 func TestSendTitleIsMaven(t *testing.T) {
 	rs := newRecordingServer(t, 200, "")
 	srv := httptest.NewServer(rs.handler())

internal/dialogue/clarify.go

@@ -41,6 +41,78 @@ type PendingQuestion struct {
 	Attempts int // questions already asked
 	// MaxAttempts caps Attempts. 0 ⇒ DefaultMaxAttempts.
 	MaxAttempts int
+	// Suspends counts how many times this question has stepped aside for
+	// something he asked instead, and come back on the end of the answer. It is
+	// deliberately NOT an attempt: a side query is not a failed answer, and
+	// charging it a retry is the V-554 shape. See CanResume for why it is
+	// counted at all.
+	Suspends int
+	// Rides counts every turn this question has ridden out on the end of
+	// someone else's reply, over the whole life of the request. Unlike Suspends
+	// it is never reset and never re-based, which is the only property that
+	// matters about it (V-663).
+	Rides int
+}
+
+// MaxSuspends — how many times one question may step aside and come back before
+// she lets the request go (V-654).
+//
+// It exists because suspension had no bound of any kind. A side query spends no
+// attempt, so MaxAttempts never applies to it, and it restarts the 90s clock, so
+// the TTL never arrives either. Measured on 2026-08-07: one unfilled time slot
+// rode the end of six consecutive unrelated replies and stopped only when a
+// seventh turn happened to read as a failed answer.
+//
+// Three, matching DefaultMaxAttempts, and for the same reason. Once he has
+// asked for three other things without touching the question, the likely truth
+// is that he has moved on and has not said so.
+const MaxSuspends = 3
+
+// MaxRides — how many turns one question may ride out on the end of an
+// unrelated reply, counted over its whole life (V-663).
+//
+// MaxSuspends did not move the measurement it was written for. Twenty-six of
+// 140 turns carried a tail before it landed and twenty-six carried one after.
+// Every bound on this question is rearmed by something ordinary:
+//
+//   - The TTL is an inactivity timer, and both noteSuspended and reaskOrGiveUp
+//     restart it, so it cannot arrive while he keeps talking.
+//   - Suspends is zeroed by any turn that reads as an answer, which is where
+//     "спасибо" and "привет" land. It resets before anything is known to have
+//     been filled.
+//   - askRemainingGap builds a fresh question for the second gap, so a reminder
+//     with two gaps gets a new allowance halfway through.
+//
+// So Suspends only bites on four strictly consecutive side queries with nothing
+// chat-like between them, which is not the shape real conversation has. Rides is
+// the same idea with the resets taken out: set once, incremented, carried
+// across a re-park, and read by nothing that could lower it.
+//
+// The shape it is aimed at is measured, not imagined. In the 2026-08-08 run one
+// question about a reminder's day rode turns 7 to 13 and ended only because
+// turn 14 was a new request. Three asides, then two turns that read as failed
+// answers, then two more asides. The asides spend no attempt and the answers
+// reset Suspends, so the two bounds take turns being rearmed by the other's
+// traffic.
+//
+// Four, not three. It has to be looser than MaxSuspends or that bound is dead
+// code, because Rides is never lower than Suspends and would always fire first.
+//
+// Do not read this as a fix for the whole ride. It ends the measured one a turn
+// early and no more. Most of that ride's length is attempts, spent by turns
+// like "спасибо" and "привет" being read as failed answers to a question about
+// a day. That is a defect in classifyTurnRole and not in any bound here.
+const MaxRides = 4
+
+// CanResume reports whether this question may step aside once more. False ⇒ the
+// caller lets the request go and says so; it must never simply stop resuming,
+// because a question dropped in silence reads as one that was answered.
+//
+// Two bounds, and they answer different questions. Suspends asks whether he has
+// walked away from this exchange in the last few turns. Rides asks whether this
+// question has been riding long enough that the answer is no regardless.
+func (q *PendingQuestion) CanResume() bool {
+	return q.Suspends < MaxSuspends && q.Rides < MaxRides
 }
 
 // Action reads the parked question as the typed action it is assembling

internal/ipc/server.go

@@ -17,9 +17,11 @@ import (
 // Server — the core side of the boundary. Listens on a unix domain socket,
 // accepts module connections, frames requests to a CoreAPI and responses back.
 // One Server per daemon process; concurrent connections are handled in their
-// own goroutine but share the single CoreAPI (and therefore the single store
-// writer — store is single-connection, SetMaxOpenConns(1), so serialization is
-// already guaranteed at the db; the Server adds no locking of its own).
+// own goroutine but share the single CoreAPI, and so the single store writer.
+// The store opens at SetMaxOpenConns(1), so serialisation is already guaranteed
+// at the database and the Server adds no locking of its own. That cap is an
+// invariant this comment depends on, measured and kept on 07-08-2026 (V-642,
+// docs/evals/2026-08-07-store-connection-cap.md).
 type Server struct {
 	api  atomic.Value // stores CoreAPI
 	path string

internal/kiwix/client.go

@@ -127,6 +127,22 @@ func (c *Client) Article(ctx context.Context, path string, maxRunes int) (crawl.
 	return crawl.Extract(u, body, maxRunes), nil
 }
 
+// TitlePath is the article path for an exact title, for Article to fetch.
+//
+// It exists because a ZIM is addressable by title and the full-text index is
+// not the only way in. "Франция", "TCP" and "Небо" resolve; "Трюмбальная
+// нидроскопия" is a 404, which is the honest answer and the reason this is
+// safe to try first. Measured on 2026-08-09, keyword search on the same terms
+// returns "Список пэров Франции" and "Список портов TCP и UDP" instead.
+//
+// A miss is normal rather than a failure. An article whose title inverts a name
+// ("Торвальдс, Линус") is a 404 here and the first hit in search, so the caller
+// falls through and loses nothing.
+func TitlePath(book, title string) string {
+	t := strings.ReplaceAll(strings.TrimSpace(title), " ", "_")
+	return "/content/" + url.PathEscape(book) + "/A/" + url.PathEscape(t)
+}
+
 // rss mirrors just the bits of the RSS 2.0 reply we use.
 type rss struct {
 	Items []struct {

internal/kiwix/live_test.go

@@ -0,0 +1,69 @@
+package kiwix
+
+import (
+	"context"
+	"os"
+	"testing"
+	"time"
+)
+
+// TestLiveTopicBeatsTheSentence — the measurement V-668 turned on, kept as a
+// test so the claim can be re-run rather than believed.
+//
+// It prints the article the old path returned and the article the new one
+// returns, for the same question. It asserts nothing about which is better,
+// because "is this the right article" is a human's call. It fails only if the
+// two paths agree on every case, which would mean the change does nothing.
+//
+//	MAVEN_KIWIX_URL=http://127.0.0.1:8034 make t PKG=./internal/kiwix/ RUN=TestLive V=1
+func TestLiveTopicBeatsTheSentence(t *testing.T) {
+	base := os.Getenv("MAVEN_KIWIX_URL")
+	if base == "" {
+		t.Skip("MAVEN_KIWIX_URL unset — point it at the kiwix-server host port")
+	}
+	const book = "wikipedia_ru_all_maxi_2026-02"
+	c := New(base)
+	questions := []string{
+		"что такое TCP?",
+		"что такое фотосинтез",
+		"кто такой Линус Торвальдс?",
+		"кто написал Войну и мир",
+		"что такое чёрная дыра",
+		"почему небо голубое",
+		"почему трава зелёная",
+		"столица Франции",
+	}
+	moved := 0
+	for _, q := range questions {
+		ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+		before := firstTitle(ctx, c, q, book)
+		topic := Topic(q)
+		after := ""
+		for _, cand := range TitleCandidates(topic) {
+			if page, err := c.Article(ctx, TitlePath(book, cand), 400); err == nil && page.Text != "" {
+				after = page.Title + " (by title)"
+				break
+			}
+		}
+		if after == "" {
+			after = firstTitle(ctx, c, topic, book)
+		}
+		cancel()
+		if before != after {
+			moved++
+		}
+		t.Logf("%-30s  before=%-34q after=%q", q, before, after)
+	}
+	t.Logf("%d of %d questions reach a different article", moved, len(questions))
+	if moved == 0 {
+		t.Error("the topic path returns exactly what the sentence path returned")
+	}
+}
+
+func firstTitle(ctx context.Context, c *Client, pattern, book string) string {
+	hits, err := c.Search(ctx, pattern, book, 3)
+	if err != nil || len(hits) == 0 {
+		return "(nothing)"
+	}
+	return hits[0].Title
+}

internal/kiwix/topic.go

@@ -0,0 +1,118 @@
+package kiwix
+
+import (
+	"strings"
+	"unicode"
+
+	"github.com/kami/maven/internal/lexicon"
+	"github.com/kami/maven/internal/morph"
+)
+
+// Topic reduces a question to the thing it is about, because Kiwix ranks by
+// keyword overlap and a whole sentence buries the keyword that matters.
+//
+// This package's own doc says it: "why is the sky blue" finds a TV episode.
+// Measured against the Russian ZIM on 2026-08-09, the sentence and the topic
+// return different articles for the same question. "кто написал Войну и мир"
+// returns "Радуйся, мир (Доктор Кто)"; "Войну и мир" returns the novel first.
+// "что такое TCP" returns "Перехват TCP-соединения"; "TCP" returns TCP. The
+// English path had a rewriter doing this with a model call. The Russian path
+// reads the book verbatim (V-508) and had nothing.
+//
+// It drops three things off the front and stops: the narrative request, the
+// interrogative, and a verb sitting between them and the noun. Everything else
+// is kept, because a word this cannot classify is more likely the topic than
+// noise. An empty return means the utterance was question words alone, and the
+// caller searches the sentence as before.
+func Topic(utterance string) string {
+	words := strings.Fields(strings.TrimSpace(utterance))
+	cut := 0
+	for cut < len(words) {
+		w := strings.Trim(strings.ToLower(words[cut]), ".,!?…:;\"'«»")
+		if w == "" {
+			cut++
+			continue
+		}
+		switch {
+		case inList(lexicon.NarrativeRequests(), w),
+			inList(lexicon.Interrogatives(), w),
+			inList(lexicon.FirstPerson(), w),
+			// "что ТАКОЕ x", "кто ТАКОЙ x" — the copula that only ever follows
+			// an interrogative, and never a topic on its own.
+			cut > 0 && isCopula(w),
+			// "расскажи ПРО x", "о x". One-letter and two-letter prepositions
+			// are not a closed class worth a lexicon set of their own.
+			cut > 0 && isLeadingPreposition(w),
+			// "кто НАПИСАЛ Войну и мир". A verb here is the question's own
+			// verb, not part of the title. Only after something was already
+			// dropped, so "написал отчёт" as a topic survives intact.
+			cut > 0 && morph.IsVerbForm(w):
+			cut++
+		default:
+			// The question mark is the sentence's, not the title's, and Kiwix
+			// carries it into the keyword match.
+			topic := strings.TrimRight(strings.Join(words[cut:], " "), " .,!?…:;\"'«»")
+			if !hasLetter(topic) {
+				return ""
+			}
+			return topic
+		}
+	}
+	return ""
+}
+
+// TitleCandidates is the topic as it might be titled, best first.
+//
+// A ZIM title is capitalized and the utterance is not: measured on 2026-08-09,
+// `/A/фотосинтез` is a 404 and `/A/Фотосинтез` is a 200. The spoken form is
+// tried first anyway, because a title that begins lowercase on purpose
+// ("iPhone") would not survive capitalizing it. Both are one request each
+// against a server on the same box, and a miss is a 404 rather than a wrong
+// article.
+func TitleCandidates(topic string) []string {
+	if topic == "" {
+		return nil
+	}
+	r := []rune(topic)
+	up := unicode.ToUpper(r[0])
+	if up == r[0] {
+		return []string{topic}
+	}
+	return []string{topic, string(up) + string(r[1:])}
+}
+
+func isCopula(w string) bool {
+	switch w {
+	case "такое", "такой", "такая", "такие", "is", "are", "was", "were":
+		return true
+	}
+	return false
+}
+
+func isLeadingPreposition(w string) bool {
+	switch w {
+	case "про", "о", "об", "обо", "по", "about", "of", "on":
+		return true
+	}
+	return false
+}
+
+func inList(list []string, w string) bool {
+	for _, x := range list {
+		if x == w {
+			return true
+		}
+	}
+	return false
+}
+
+// hasLetter is the guard against a topic that reduced to punctuation or digits
+// alone, which no ZIM title matches.
+func hasLetter(s string) bool {
+	for _, r := range s {
+		if unicode.IsLetter(r) {
+			return true
+		}
+	}
+	return false
+}

internal/kiwix/topic_test.go

@@ -0,0 +1,67 @@
+package kiwix
+
+import "testing"
+
+// The cases the 2026-08-09 measurement turned on, plus the ones a topic must
+// not damage. Each left column returned a wrong article when it was sent whole.
+func TestTopicKeepsTheThingTheQuestionIsAbout(t *testing.T) {
+	cases := []struct{ utterance, want string }{
+		{"что такое TCP?", "TCP"},
+		{"что такое фотосинтез", "фотосинтез"},
+		{"кто такой Линус Торвальдс?", "Линус Торвальдс"},
+		{"кто написал Войну и мир", "Войну и мир"},
+		{"расскажи про битву при Ватерлоо", "битву при Ватерлоо"},
+		{"what is photosynthesis", "photosynthesis"},
+		// No question word, so there is nothing to drop. The topic is the
+		// whole utterance and the search is what it was before.
+		{"столица Франции", "столица Франции"},
+		{"почему небо голубое", "небо голубое"},
+	}
+	for _, c := range cases {
+		if got := Topic(c.utterance); got != c.want {
+			t.Errorf("Topic(%q) = %q, want %q", c.utterance, got, c.want)
+		}
+	}
+}
+
+// A verb only goes when a question word already went. Otherwise "написал
+// отчёт" loses the verb that names what he means.
+func TestTopicDropsAVerbOnlyBehindAQuestionWord(t *testing.T) {
+	if got := Topic("написал отчёт"); got != "написал отчёт" {
+		t.Errorf("Topic dropped a leading verb with no question word: %q", got)
+	}
+}
+
+// Question words alone reduce to nothing, and the caller reads that as "no
+// topic" and searches the sentence rather than searching the empty string.
+func TestTopicIsEmptyWhenNothingIsLeft(t *testing.T) {
+	for _, q := range []string{"что такое?", "кто?", "почему", "???"} {
+		if got := Topic(q); got != "" {
+			t.Errorf("Topic(%q) = %q, want empty", q, got)
+		}
+	}
+}
+
+func TestTitlePathEscapesAndUnderscores(t *testing.T) {
+	got := TitlePath("wikipedia_ru_all_maxi_2026-02", "Чёрная дыра")
+	want := "/content/wikipedia_ru_all_maxi_2026-02/A/%D0%A7%D1%91%D1%80%D0%BD%D0%B0%D1%8F_%D0%B4%D1%8B%D1%80%D0%B0"
+	if got != want {
+		t.Errorf("TitlePath = %q, want %q", got, want)
+	}
+}
+
+// A ZIM title carries a leading capital and the utterance does not. The spoken
+// form is still tried first, so a title that begins lowercase on purpose keeps
+// its chance.
+func TestTitleCandidatesTryTheSpokenFormFirst(t *testing.T) {
+	got := TitleCandidates("фотосинтез")
+	if len(got) != 2 || got[0] != "фотосинтез" || got[1] != "Фотосинтез" {
+		t.Errorf("TitleCandidates = %q", got)
+	}
+	if got := TitleCandidates("TCP"); len(got) != 1 || got[0] != "TCP" {
+		t.Errorf("an already-capital topic was tried twice: %q", got)
+	}
+	if got := TitleCandidates(""); got != nil {
+		t.Errorf("TitleCandidates(\"\") = %q, want nil", got)
+	}
+}

internal/lexicon/lexicon.go

@@ -119,6 +119,11 @@ func PartsOfDay() []string { return words("parts_of_day") }
 // ReminderVerbs returns the imperatives that open a reminder.
 func ReminderVerbs() []string { return words("reminder_verbs") }
 
+// Pleasantries returns the whole utterances that greet, thank or say goodbye.
+// Whole utterances and not tokens: see the set's own note for why the tokens
+// are unsafe alone.
+func Pleasantries() []string { return words("pleasantries") }
+
 // TaskDoneWords returns the words that finish a task, and TaskDropWords the
 // words that abandon one. Two sets rather than one with a value, because the
 // store records which of the two happened and the caller has to say so.

internal/lexicon/lexicon_ru_v1.json

@@ -176,6 +176,19 @@
         "morning", "afternoon", "evening", "night"
       ]
     },
+    "pleasantries": {
+      "note": "Whole utterances that greet, thank or say goodbye. They ask for nothing and answer nothing, so a parked question must neither consume them as a failed answer nor be dropped by them (V-663). Matched as WHOLE utterances and never as tokens, because the tokens are not safe alone: \"вечер\" answers \"это утра или вечера?\" and \"нет\" answers a confirm. Anything that could fill a slot stays out. The control words (\"стоп\", \"отмена\") stay out too, because isCancel already owns them and they mean something stronger.",
+      "words": [
+        "привет", "приветик", "здравствуй", "здравствуйте",
+        "доброе утро", "добрый день", "добрый вечер",
+        "пока", "прощай", "до свидания", "спокойной ночи",
+        "спасибо", "спасибо тебе", "большое спасибо", "благодарю",
+        "извини", "извините", "прости", "простите",
+        "hi", "hello", "hey", "bye", "goodbye",
+        "good morning", "good evening", "good night",
+        "thanks", "thank you", "thanks a lot", "sorry"
+      ]
+    },
     "reminder_verbs": {
       "note": "The imperatives that mean \"remind me\", in the forms he speaks. The same kind of set as capture_verbs and decided the same way: it is her vocabulary, not a discovery about Russian (Vikunja #530). The alarm verbs joined them in V-627. \"разбуди меня в 6:30\" is a reminder that fires at the hour he gets up, and the set knew no form of it, so an alarm reached IntentReminder only by resembling one to the embedder.",
       "words": [

internal/router/decisiontrace.go

@@ -14,12 +14,15 @@ import (
 	"github.com/kami/maven/internal/decision"
 )
 
-// The two routing engines, named as claimants. They are one stage and not two,
-// because only one of them ever runs: the classifier is reached when the model
-// is absent or errored, never alongside it.
+// The three routing engines, named as claimants. The model and the classifier
+// are one stage and not two, because only one of them ever runs: the classifier
+// is reached when the model is absent or errored, never alongside it. The heads
+// run before both and decline on low confidence, so they can appear beside
+// either one in a record.
 const (
 	claimantLLM        = "llm-router"
 	claimantClassifier = "classifier"
+	claimantHeads      = "routing-heads"
 )
 
 // thinReason names which arm of gateLLMDecision cut the confidence. The gate

internal/router/dumpprompt_test.go

@@ -0,0 +1,22 @@
+package router
+
+import (
+	"os"
+	"testing"
+)
+
+// TestDumpPrompt writes the router prompt and grammar to disk so the training
+// workspace labels with the daemon's own contract rather than a retyped copy.
+// It is inert unless MAVEN_DUMP_PROMPT names a directory.
+func TestDumpPrompt(t *testing.T) {
+	dir := os.Getenv("MAVEN_DUMP_PROMPT")
+	if dir == "" {
+		t.Skip("MAVEN_DUMP_PROMPT unset")
+	}
+	if err := os.WriteFile(dir+"/route_system.txt", []byte(routeSystem), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(dir+"/route_grammar.gbnf", []byte(routeGrammar), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}

internal/router/embedderid_test.go

@@ -1,10 +1,13 @@
 package router
 
-import "testing"
+import (
+	"strings"
+	"testing"
+)
 
 func TestEmbedderIDFromModelPath(t *testing.T) {
 	got := modelIDFromPath("/opt/maven/models/embedder/multilingual-e5-small.onnx")
-	if got != "multilingual-e5-small@384" {
+	if got != "multilingual-e5-small@384/tok2" {
 		t.Fatalf("modelIDFromPath = %q", got)
 	}
 	// A different model file must produce a different id, even at 384 dim.
@@ -12,6 +15,13 @@ func TestEmbedderIDFromModelPath(t *testing.T) {
 	if old == got {
 		t.Fatal("two different models share one id")
 	}
+	// The tokenizer is half of what makes a vector, and it changes under a
+	// model file whose name never moves (V-664). An id that ignored it would
+	// leave stored passages in one space and every new query in another, with
+	// nothing to trigger the re-embed.
+	if !strings.Contains(got, "/tok") {
+		t.Fatalf("id %q does not name the tokenizer revision", got)
+	}
 }
 
 func TestEmbedderIDIncludesDim(t *testing.T) {

internal/router/eval/eval.go

@@ -36,17 +36,27 @@ var fixtureJSON []byte
 //
 // Intent is empty exactly when WantClarify is set: the contract there is that
 // the router refuses instead of guessing.
+//
+// WantSource is a pointer because the destination has three states and a bare
+// string only has two (V-659). Absent means the case does not score a
+// destination at all, which is every intent but query: a fact, a reminder, a
+// note, an act, a chat or a system turn never reaches queryWalk. Present and
+// empty is the SourceUnknown contract — the decider must name nothing and let
+// the daemon walk the whole chain, which is the right answer whenever two
+// destinations can both answer and the utterance does not choose. Present and
+// named is a destination the route must produce.
 type Case struct {
-	ID          string        `json:"id"`
-	Utterance   string        `json:"utterance"`
-	Lang        string        `json:"lang"`
-	Intent      router.Intent `json:"intent"`
-	WantTime    bool          `json:"want_time"`
-	WantFn      bool          `json:"want_fn"`
-	WantFactKey string        `json:"want_fact_key"`
-	WantClarify bool          `json:"want_clarify"`
-	Tags        []string      `json:"tags"`
-	Note        string        `json:"note"`
+	ID          string         `json:"id"`
+	Utterance   string         `json:"utterance"`
+	Lang        string         `json:"lang"`
+	Intent      router.Intent  `json:"intent"`
+	WantTime    bool           `json:"want_time"`
+	WantFn      bool           `json:"want_fn"`
+	WantFactKey string         `json:"want_fact_key"`
+	WantClarify bool           `json:"want_clarify"`
+	WantSource  *router.Source `json:"want_source,omitempty"`
+	Tags        []string       `json:"tags"`
+	Note        string         `json:"note"`
 }
 
 // Fixture — the versioned envelope, same shape as
@@ -118,6 +128,11 @@ type Outcome struct {
 	// (a slot gap is a parser fix; a wrong intent is a router fix).
 	IntentOK bool
 	Reasons  []string
+	// SourceReason is set when the case labelled a destination and the route
+	// named a different one. It is kept out of Reasons on purpose: the
+	// destination is the second half of a route and it is scored separately,
+	// so a wrong destination must not move the intent number (V-659).
+	SourceReason string
 }
 
 // Report — the aggregate. Accuracy is the headline; the rest exists so a
@@ -139,7 +154,15 @@ type Report struct {
 	// (reminder grammar → applyAction's time parser). Not a miss, but not a
 	// full router-level win either; tracked so the two aren't conflated.
 	SlotsDeferred int
-	Outcomes      []Outcome
+	// SourceTotal counts the cases carrying a want_source, and SourceHit the
+	// ones whose route named it. Reported apart from Passed because intent and
+	// destination are two decisions, and one number hides which one moved.
+	SourceTotal int
+	SourceHit   int
+	// SourceConfusion counts want→got destination pairs. "" reads as the
+	// SourceUnknown floor on either side.
+	SourceConfusion map[string]int
+	Outcomes        []Outcome
 	// Confusion counts want→got intent pairs, decided cases only.
 	Confusion map[string]int
 	// ByTag accuracy for the fixture's tags ("hard", "homelab", …).
@@ -172,6 +195,17 @@ func (r Report) IntentAccuracy() float64 {
 	return float64(r.IntentHit) / float64(r.Total)
 }
 
+// SourceAccuracy — fraction of the labelled cases whose route named the right
+// destination. Denominator is SourceTotal and not Total, because most of the
+// fixture never reaches a query source and scoring those would report a
+// percentage of nothing.
+func (r Report) SourceAccuracy() float64 {
+	if r.SourceTotal == 0 {
+		return 0
+	}
+	return float64(r.SourceHit) / float64(r.SourceTotal)
+}
+
 // Score runs every case through r and aggregates. It never fails the run on a
 // route error — an erroring case scores as a miss and is counted in Errors,
 // because "the model was down" and "the model was wrong" are different numbers
@@ -186,11 +220,12 @@ func Score(ctx context.Context, name string, r Router, f Fixture) (Report, error
 		return Report{}, err
 	}
 	rep := Report{
-		Name:      name,
-		Total:     len(f.Cases),
-		Confusion: map[string]int{},
-		ByTag:     map[string]TagStat{},
-		ByLang:    map[string]TagStat{},
+		Name:            name,
+		Total:           len(f.Cases),
+		Confusion:       map[string]int{},
+		SourceConfusion: map[string]int{},
+		ByTag:           map[string]TagStat{},
+		ByLang:          map[string]TagStat{},
 	}
 	lat := make([]time.Duration, 0, len(f.Cases))
 
@@ -242,6 +277,29 @@ func Score(ctx context.Context, name string, r Router, f Fixture) (Report, error
 			}
 		}
 
+		// The destination is scored outside the switch and outside Pass. A case
+		// that clarified or landed the wrong intent named no destination, and
+		// that is a real miss rather than a case to skip — otherwise the
+		// denominator quietly drops every turn the route already lost. Only a
+		// route error is skipped, because "the model was down" is the Errors
+		// number and not a destination result.
+		if c.WantSource != nil && err == nil {
+			rep.SourceTotal++
+			switch {
+			case !o.IntentOK:
+				// The route never got to a destination, so a match on the
+				// SourceUnknown floor here would be a coincidence scored as a
+				// win: a clarify names nothing and would satisfy "" for free.
+				o.SourceReason = fmt.Sprintf("no destination, route missed %q", c.Intent)
+				rep.SourceConfusion[string(*c.WantSource)+"→(no route)"]++
+			case d.Source == *c.WantSource:
+				rep.SourceHit++
+			default:
+				rep.SourceConfusion[string(*c.WantSource)+"→"+string(d.Source)]++
+				o.SourceReason = fmt.Sprintf("source %q, want %q", d.Source, *c.WantSource)
+			}
+		}
+
 		o.Pass = len(o.Reasons) == 0
 		if o.Pass {
 			rep.Passed++
@@ -298,25 +356,40 @@ func (r Report) String() string {
 		r.Name, r.Passed, r.Total, 100*r.Accuracy(), 100*r.IntentAccuracy())
 	fmt.Fprintf(&b, "  clarify: %d false (asked, shouldn't) / %d missed (guessed, shouldn't) | errors: %d | slots deferred to daemon: %d\n",
 		r.FalseClarify, r.MissedClarify, r.Errors, r.SlotsDeferred)
+	if r.SourceTotal > 0 {
+		fmt.Fprintf(&b, "  destination: %d/%d labelled cases (%.1f%%)\n",
+			r.SourceHit, r.SourceTotal, 100*r.SourceAccuracy())
+	}
 	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
 	fmt.Fprintf(&b, "  by lang: %s\n", renderStats(r.ByLang))
 	fmt.Fprintf(&b, "  by tag:  %s\n", renderStats(r.ByTag))
 	if len(r.Confusion) > 0 {
 		fmt.Fprintf(&b, "  confusion: %s\n", renderCounts(r.Confusion))
 	}
+	if len(r.SourceConfusion) > 0 {
+		fmt.Fprintf(&b, "  destination confusion: %s\n", renderCounts(r.SourceConfusion))
+	}
 	return b.String()
 }
 
-// Failures — the per-case detail, sorted by ID so two runs diff cleanly.
+// Failures — the per-case detail, sorted by ID so two runs diff cleanly. A case
+// that landed its intent and missed its destination is listed too, marked, so
+// the half that moved is readable without diffing two percentages.
 func (r Report) Failures() string {
 	var b strings.Builder
 	out := append([]Outcome(nil), r.Outcomes...)
 	sort.Slice(out, func(i, j int) bool { return out[i].Case.ID < out[j].Case.ID })
 	for _, o := range out {
-		if o.Pass {
-			continue
+		switch {
+		case !o.Pass:
+			reasons := o.Reasons
+			if o.SourceReason != "" {
+				reasons = append(append([]string(nil), reasons...), o.SourceReason)
+			}
+			fmt.Fprintf(&b, "  %s %q: %s\n", o.Case.ID, o.Case.Utterance, strings.Join(reasons, "; "))
+		case o.SourceReason != "":
+			fmt.Fprintf(&b, "  %s %q: route ok, %s\n", o.Case.ID, o.Case.Utterance, o.SourceReason)
 		}
-		fmt.Fprintf(&b, "  %s %q: %s\n", o.Case.ID, o.Case.Utterance, strings.Join(o.Reasons, "; "))
 	}
 	return b.String()
 }

internal/router/eval/eval_test.go

@@ -266,6 +266,11 @@ func baselineGrammars(acts router.ActMatcher) []router.Grammar {
 	// Same order as buildRouter (voicewire.go). The fixture is only worth
 	// anything while its grammar set is the daemon's grammar set.
 	grammars = append(grammars, router.AgendaQueryGrammars()...)
+	// After the agenda rules and before the feed and list rules, same as
+	// voicewire.go: "что такое лента" is a definition question and the feed
+	// rule would claim it on the noun alone (V-655). Missing here until V-659,
+	// so the fixture was scoring a grammar set the daemon does not run.
+	grammars = append(grammars, router.WorldQueryGrammars()...)
 	grammars = append(grammars, router.FeedQueryGrammar())
 	// The list side of the same exposure: a phrasing with no possessive in it
 	// ("список дел") routed system and never reached queryTasks (Vikunja #467).

internal/router/eval/heads_test.go

@@ -0,0 +1,86 @@
+package eval
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/kami/maven/internal/config"
+	"github.com/kami/maven/internal/router"
+)
+
+// TestONNXRoutingHeads — the cascade with the routing heads wired, which is
+// what V-664 deploys. Opt-in via MAVEN_ONNX_LIB, same as TestONNXBaseline, and
+// one TestONNX* per process.
+//
+// The comparison worth reading is against TestONNXBaseline, which is the same
+// cascade with the same grammars and the same classifier floor and no heads.
+// Only the middle arm varies.
+//
+// It also checks the Go unigram tokenizer against the Python one, because the
+// heads were trained through transformers and are read through a hand-written
+// tokenizer. A mismatch shows up here as a score below what Python measured on
+// the same weights, and nowhere else.
+func TestONNXRoutingHeads(t *testing.T) {
+	lib := os.Getenv("MAVEN_ONNX_LIB")
+	if lib == "" {
+		t.Skip("MAVEN_ONNX_LIB unset — see AGENTS.md § Embedder model for intent routing")
+	}
+	// Absolute, because onnxruntime resolves a graph's external weights file
+	// against the model path it was given, and a relative one lands in the
+	// test's working directory.
+	root, err := filepath.Abs("../../..")
+	if err != nil {
+		t.Fatal(err)
+	}
+	model := filepath.Join(root, "models/embedder/multilingual-e5-small/model_quantized.onnx")
+	tok := filepath.Join(root, "models/embedder/multilingual-e5-small/tokenizer.json")
+	heads := filepath.Join(root, "models/embedder/router-heads/router_heads.onnx")
+	for _, p := range []string{lib, model, tok, heads} {
+		if _, err := os.Stat(p); err != nil {
+			t.Skipf("missing %s: %v", p, err)
+		}
+	}
+	emb, err2 := router.NewONNXEmbedder(model, tok, lib)
+	if err2 != nil {
+		t.Skipf("onnx embedder unavailable: %v", err2)
+	}
+	err = nil
+	defer emb.Close()
+
+	h, err := router.NewRouterHeads(heads, tok)
+	if err != nil {
+		t.Skipf("routing heads unavailable: %v", err)
+	}
+	defer h.Close()
+
+	f, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	rep, err := Score(context.Background(), "heads+classifier", withHeads(t, emb, h), f)
+	if err != nil {
+		t.Fatalf("Score: %v", err)
+	}
+	t.Log("\n" + rep.String() + rep.Failures())
+}
+
+// withHeads mirrors newBaselineRouter and adds the one arm under test. It is a
+// separate function rather than a parameter so the baseline's signature stays
+// the shape every other test calls it with.
+func withHeads(t *testing.T, emb router.Embedder, h *router.RouterHeads) *router.Router {
+	t.Helper()
+	acts := router.DefaultActMatcher{Fns: actFns}
+	return router.New(router.Config{
+		Grammars:   baselineGrammars(acts),
+		Classifier: newBaselineClassifier(t, emb),
+		Extractor: router.Extractor{
+			Time:  router.StubDateTimeParser{},
+			Acts:  acts,
+			Facts: router.DefaultFactParser{},
+		},
+		Threshold: config.DefaultRouterThreshold,
+		Heads:     h,
+	})
+}

internal/router/eval/ru_routing_v1.json

@@ -6,37 +6,41 @@
     "Held-out routing contract. Every utterance here is absent from models/seeds/*.txt (TestFixtureIsHeldOut enforces it verbatim) — scoring a classifier on its own seed phrases measures memorisation, not routing.",
     "This is a CONTRACT, not a snapshot of current behaviour. Cases the classifier cascade fails today are expected to stay in the file and fail loudly; that failure count is the number Vikunja #319 compares against the LLM router before #320 flips the default.",
     "Slot expectations are deployment-independent on purpose. want_fn is a boolean (the act must resolve to SOME allowlisted fn) because the allowlist lives in deploy config, not here. want_fact_key names the loop's rule keys (water/meal/sleep/break/shower) — a fact that lands under the wrong key silently starves the predicate that reads it.",
-    "want_clarify cases carry intent \"\": the contract is that the router refuses rather than guesses. A confident answer there is a worse failure than a miss."
+    "want_clarify cases carry intent \"\": the contract is that the router refuses rather than guesses. A confident answer there is a worse failure than a miss.",
+    "want_source is the second half of a route (V-655). It is present only on query cases, because no other intent reaches queryWalk, and absent there means absent rather than SourceUnknown. Empty is a label and not a gap: it asserts that the decider must name nothing and let the daemon walk the whole chain in order, his data first.",
+    "Seven cases assert that floor and six of them are homelab operations. They cluster because SourceRecall, SourceNetwork and SourceAttention overlap on every question about the box: mavpoll writes its observations into the fact store recall reads. That is a finding about the enum, not a gap in the labelling.",
+    "ru-query-020 and ru-query-024 are the same utterance, as are ru-query-021 and ru-query-025. Both pairs differ in tags and note only, so both pairs are counted twice in every number this fixture reports.",
+    "Every want_source is the destination that SHOULD claim the turn, which on ru-query-026 through 030 is not the one that did. Those five were observed failing on the box on 2026-08-07 (docs/evals/2026-08-07-week-of-usage.md). A fixture that passes on the day it is written measures nothing."
   ],
   "cases": [
-    { "id": "ru-query-001", "utterance": "сколько воды я выпил с утра", "lang": "ru", "intent": "query", "tags": ["aggregate"] },
-    { "id": "ru-query-002", "utterance": "я сегодня вообще пил воду", "lang": "ru", "intent": "query", "tags": ["hard", "fact-shaped"], "note": "past-tense fact lexicon in a question — the classifier's fact centroid pulls this hard" },
-    { "id": "ru-query-003", "utterance": "во сколько я лёг вчера", "lang": "ru", "intent": "query", "tags": ["temporal"] },
-    { "id": "ru-query-004", "utterance": "давно я не тренировался", "lang": "ru", "intent": "query", "tags": ["hard", "no-question-word"] },
-    { "id": "ru-query-005", "utterance": "напоминания на завтра есть", "lang": "ru", "intent": "query", "tags": ["hard", "reminder-shaped"], "note": "asks about reminders, does not create one" },
-    { "id": "ru-query-006", "utterance": "что я записывал про кота", "lang": "ru", "intent": "query", "tags": ["recall"] },
-    { "id": "ru-query-007", "utterance": "сколько раз я ел вчера", "lang": "ru", "intent": "query", "tags": ["aggregate", "hard"] },
-    { "id": "ru-query-008", "utterance": "мой вес за последний месяц", "lang": "ru", "intent": "query", "tags": ["no-verb"] },
-    { "id": "ru-query-009", "utterance": "когда я в последний раз принимал витамины", "lang": "ru", "intent": "query", "tags": ["temporal"] },
-    { "id": "ru-query-010", "utterance": "есть новости по бэкапу базы", "lang": "ru", "intent": "query", "tags": ["homelab"] },
-    { "id": "ru-query-011", "utterance": "почему сервер тормозит", "lang": "ru", "intent": "query", "tags": ["homelab", "hard"], "note": "diagnostic question, not a chat opener" },
-    { "id": "ru-query-012", "utterance": "какие заметки я оставил про полив", "lang": "ru", "intent": "query", "tags": ["recall"] },
-    { "id": "ru-query-013", "utterance": "во сколько у меня встреча", "lang": "ru", "intent": "query", "tags": ["calendar"] },
-    { "id": "ru-query-019", "utterance": "что у меня стоит в календаре на послезавтра", "lang": "ru", "intent": "query", "tags": ["calendar", "hard"], "note": "agenda, not the clock: the daemon answers this from CalendarEvents inside the query branch, so the clock/date system rule must not swallow it" },
-    { "id": "ru-query-022", "utterance": "какие планы на завтра?", "lang": "ru", "intent": "query", "tags": ["calendar"], "note": "the same agenda question as ru-query-019 aimed at another day; it answered \u043f\u043e\u043a\u0430 \u043d\u0435 \u0443\u043c\u0435\u044e on the deployed daemon while the today form worked (Vikunja #471)" },
-    { "id": "ru-query-023", "utterance": "\u043a\u043e\u0433\u0434\u0430 \u043f\u043b\u0430\u043d\u0451\u0440\u043a\u0430?", "lang": "ru", "intent": "query", "tags": ["calendar", "hard"], "note": "a named event with no calendar word — the noun is the only signal that this is a question about his day" },
-    { "id": "ru-query-024", "utterance": "что дальше?", "lang": "ru", "intent": "query", "tags": ["calendar", "no-question-word"], "note": "the rest of the day, with no possessive and no plan word to anchor on; the model called it a fact and the write had to be caught downstream (Vikunja #498)" },
-    { "id": "ru-query-025", "utterance": "расскажи про битву при Ватерлоо", "lang": "ru", "intent": "query", "tags": ["world", "no-question-word"], "note": "a narrative request carries no question mark and no interrogative, so it routed fact; contrast ru-chat-003, where the same verb asks for a joke" },
-    { "id": "ru-query-014", "utterance": "я успеваю до дедлайна", "lang": "ru", "intent": "query", "tags": ["hard", "no-question-word"] },
-    { "id": "ru-query-015", "utterance": "сколько я прошёл шагов", "lang": "ru", "intent": "query", "tags": ["aggregate"] },
-    { "id": "ru-query-016", "utterance": "покажи давление за неделю", "lang": "ru", "intent": "query", "tags": ["hard", "imperative"], "note": "imperative form but a read — must not route to act" },
-    { "id": "ru-query-017", "utterance": "чем я занимался в среду", "lang": "ru", "intent": "query", "tags": ["hard", "chat-shaped"] },
-    { "id": "ru-query-018", "utterance": "хватает ли места под новые бэкапы", "lang": "ru", "intent": "query", "tags": ["homelab"] },
-    { "id": "ru-query-020", "utterance": "что дальше?", "lang": "ru", "intent": "query", "tags": ["agenda", "hard"], "note": "the rest of the day, with no interrogative the model can read as a question — it routed fact until a stage 0 rule claimed it (V-498)" },
-    { "id": "ru-query-021", "utterance": "расскажи про битву при Ватерлоо", "lang": "ru", "intent": "query", "tags": ["world", "hard"], "note": "a world question phrased as an instruction. It routed fact, and the fact gate had to catch the write (V-498)" },
-    { "id": "en-query-001", "utterance": "did I take my vitamins today", "lang": "en", "intent": "query", "tags": ["fact-shaped"] },
-    { "id": "en-query-002", "utterance": "how long since the last backup finished", "lang": "en", "intent": "query", "tags": ["temporal"] },
-    { "id": "en-query-003", "utterance": "show me this week's weight", "lang": "en", "intent": "query", "tags": ["imperative"] },
+    { "id": "ru-query-001", "utterance": "сколько воды я выпил с утра", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["aggregate"] },
+    { "id": "ru-query-002", "utterance": "я сегодня вообще пил воду", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["hard", "fact-shaped"], "note": "past-tense fact lexicon in a question — the classifier's fact centroid pulls this hard" },
+    { "id": "ru-query-003", "utterance": "во сколько я лёг вчера", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["temporal"] },
+    { "id": "ru-query-004", "utterance": "давно я не тренировался", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["hard", "no-question-word"] },
+    { "id": "ru-query-005", "utterance": "напоминания на завтра есть", "lang": "ru", "intent": "query", "want_source": "", "tags": ["hard", "reminder-shaped"], "note": "asks about reminders, does not create one. want_source is the floor on purpose: no query source reads the reminder store, and day-plan is SourceCalendar over a table this box does not write." },
+    { "id": "ru-query-006", "utterance": "что я записывал про кота", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["recall"] },
+    { "id": "ru-query-007", "utterance": "сколько раз я ел вчера", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["aggregate", "hard"] },
+    { "id": "ru-query-008", "utterance": "мой вес за последний месяц", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["no-verb"] },
+    { "id": "ru-query-009", "utterance": "когда я в последний раз принимал витамины", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["temporal"] },
+    { "id": "ru-query-010", "utterance": "есть новости по бэкапу базы", "lang": "ru", "intent": "query", "want_source": "", "tags": ["homelab"], "note": "recall, attention and network can each answer it, because mavpoll writes its netdata and uptime-kuma observations into the fact store recall reads. Naming one takes the other two off the turn." },
+    { "id": "ru-query-011", "utterance": "почему сервер тормозит", "lang": "ru", "intent": "query", "want_source": "", "tags": ["homelab", "hard"], "note": "diagnostic question, not a chat opener. network holds the box and attention holds the alarm about the box. The utterance does not choose, so neither does the label." },
+    { "id": "ru-query-012", "utterance": "какие заметки я оставил про полив", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["recall"] },
+    { "id": "ru-query-013", "utterance": "во сколько у меня встреча", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["calendar"] },
+    { "id": "ru-query-019", "utterance": "что у меня стоит в календаре на послезавтра", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["calendar", "hard"], "note": "agenda, not the clock: the daemon answers this from CalendarEvents inside the query branch, so the clock/date system rule must not swallow it" },
+    { "id": "ru-query-022", "utterance": "какие планы на завтра?", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["calendar"], "note": "the same agenda question as ru-query-019 aimed at another day; it answered \u043f\u043e\u043a\u0430 \u043d\u0435 \u0443\u043c\u0435\u044e on the deployed daemon while the today form worked (Vikunja #471)" },
+    { "id": "ru-query-023", "utterance": "\u043a\u043e\u0433\u0434\u0430 \u043f\u043b\u0430\u043d\u0451\u0440\u043a\u0430?", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["calendar", "hard"], "note": "a named event with no calendar word — the noun is the only signal that this is a question about his day" },
+    { "id": "ru-query-024", "utterance": "что дальше?", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["calendar", "no-question-word"], "note": "the rest of the day, with no possessive and no plan word to anchor on; the model called it a fact and the write had to be caught downstream (Vikunja #498)" },
+    { "id": "ru-query-025", "utterance": "расскажи про битву при Ватерлоо", "lang": "ru", "intent": "query", "want_source": "world", "tags": ["world", "no-question-word"], "note": "a narrative request carries no question mark and no interrogative, so it routed fact; contrast ru-chat-003, where the same verb asks for a joke" },
+    { "id": "ru-query-014", "utterance": "я успеваю до дедлайна", "lang": "ru", "intent": "query", "want_source": "", "tags": ["hard", "no-question-word"], "note": "a deadline lives in the task list, the calendar or Praxis depending on where he put it. The destination depends on his data, not on his words." },
+    { "id": "ru-query-015", "utterance": "сколько я прошёл шагов", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["aggregate"] },
+    { "id": "ru-query-016", "utterance": "покажи давление за неделю", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["hard", "imperative"], "note": "imperative form but a read — must not route to act" },
+    { "id": "ru-query-017", "utterance": "чем я занимался в среду", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["hard", "chat-shaped"] },
+    { "id": "ru-query-018", "utterance": "хватает ли места под новые бэкапы", "lang": "ru", "intent": "query", "want_source": "", "tags": ["homelab"], "note": "disk headroom. network is the only source that reads the box, but the phrasing is a capacity question and not a LAN one." },
+    { "id": "ru-query-020", "utterance": "что дальше?", "lang": "ru", "intent": "query", "want_source": "calendar", "tags": ["agenda", "hard"], "note": "the rest of the day, with no interrogative the model can read as a question — it routed fact until a stage 0 rule claimed it (V-498)" },
+    { "id": "ru-query-021", "utterance": "расскажи про битву при Ватерлоо", "lang": "ru", "intent": "query", "want_source": "world", "tags": ["world", "hard"], "note": "a world question phrased as an instruction. It routed fact, and the fact gate had to catch the write (V-498)" },
+    { "id": "en-query-001", "utterance": "did I take my vitamins today", "lang": "en", "intent": "query", "want_source": "recall", "tags": ["fact-shaped"] },
+    { "id": "en-query-002", "utterance": "how long since the last backup finished", "lang": "en", "intent": "query", "want_source": "", "tags": ["temporal"], "note": "the completion time is a fact the poller wrote, so recall answers it. A person asking this wants the operational answer. Both are true." },
+    { "id": "en-query-003", "utterance": "show me this week's weight", "lang": "en", "intent": "query", "want_source": "recall", "tags": ["imperative"] },
 
     { "id": "ru-fact-001", "utterance": "только что выпил кружку воды", "lang": "ru", "intent": "fact", "want_fact_key": "water" },
     { "id": "ru-fact-002", "utterance": "воды попил наконец", "lang": "ru", "intent": "fact", "want_fact_key": "water", "tags": ["inverted"] },
@@ -106,6 +110,11 @@
     { "id": "amb-005", "utterance": "потом", "lang": "ru", "want_clarify": true, "tags": ["ambiguous", "filler"] },
     { "id": "amb-006", "utterance": "the thing from earlier", "lang": "en", "want_clarify": true, "tags": ["ambiguous", "anaphora"] },
     { "id": "amb-007", "utterance": "напомни", "lang": "ru", "want_clarify": true, "tags": ["ambiguous", "reminder"], "note": "the reminder verb and nothing else — she knows the shape of the request and not one thing about it. Answered 'не получилось разобрать время напоминания' on the box until V-548: the subjectless-reminder gate tested Slots.Text == \"\", and fillSlots had put the verb in that slot" },
-    { "id": "amb-008", "utterance": "ну напомни же", "lang": "ru", "want_clarify": true, "tags": ["ambiguous", "reminder", "filler"], "note": "the same request wrapped in particles, which is why filler_particles is a lexicon set — without it the particles read as the subject" }
+    { "id": "amb-008", "utterance": "ну напомни же", "lang": "ru", "want_clarify": true, "tags": ["ambiguous", "reminder", "filler"], "note": "the same request wrapped in particles, which is why filler_particles is a lexicon set — without it the particles read as the subject" },
+    { "id": "ru-query-026", "utterance": "что такое TCP?", "lang": "ru", "intent": "query", "want_source": "world", "tags": ["world", "regression"], "note": "weather claimed it on 2026-08-07 and answered \"для какого города?\", because it read one percent closer than the leftover seeds. WorldQueryGrammars claims it at stage 0 now." },
+    { "id": "ru-query-027", "utterance": "сколько будет 17 на 23?", "lang": "ru", "intent": "query", "want_source": "world", "tags": ["world", "arithmetic", "regression"], "note": "same source, same day, same answer about a city. Arithmetic is not a place." },
+    { "id": "ru-query-028", "utterance": "какой у меня любимый язык?", "lang": "ru", "intent": "query", "want_source": "recall", "tags": ["recall", "possessive", "regression"], "note": "the feed answered it with kernel headlines. \"у меня\" is the whole signal and it points inward." },
+    { "id": "ru-query-029", "utterance": "кто такой Линус Торвальдс?", "lang": "ru", "intent": "query", "want_source": "world", "tags": ["world", "person", "regression"], "note": "the personal boundary answered \"не нашла у тебя такой записи\". A named public person is not his data." },
+    { "id": "ru-query-030", "utterance": "что там с бэкапами?", "lang": "ru", "intent": "query", "want_source": "", "tags": ["homelab", "regression"], "note": "search claimed it, which inverts the boundary outward. The fix is the chain order and not a destination: recall, attention and network all answer it, same as ru-query-010." }
   ]
 }

internal/router/heads.go

@@ -0,0 +1,243 @@
+package router
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	ort "github.com/yalue/onnxruntime_go"
+)
+
+// The routing heads (V-546, V-661, V-664). Four linear heads over one masked
+// mean pool of a fine-tuned copy of multilingual-e5-small: intent,
+// destination, BIO slot tags and clarify. Trained on workpc, exported to ONNX,
+// and read here.
+//
+// Why this is not the classifier. The classifier compares one utterance to
+// frozen seed phrases by cosine. A head is a softmax over the label set, so it
+// cannot name a value that does not exist, and its max is a calibratable
+// confidence where Confidence: 1.0 was a hardcode.
+//
+// Why it is not the resident model either. It answers in single-digit
+// milliseconds against the model's p50 of 1.19s, and it names a destination
+// the classifier arm never names at all.
+//
+// The body is a COPY of the embedder weights, fine-tuned. It must never
+// replace models/embedder/multilingual-e5-small — memory recall depends on
+// that file scoring what it scored.
+//
+// The slot head is exported and deliberately not read. Slots already come from
+// the stage-2 extractor, and mapping BIO tags back to text needs character
+// offsets the unigram tokenizer does not keep. Reading it is separate work.
+const (
+	// headsSeq — the sequence length the heads were trained at. Padding is
+	// masked out of both attention and the pool, so this changes nothing but
+	// truncation, and truncation is what training did at 64.
+	headsSeq = 64
+
+	// headsThreshold — max softmax over the intent head, below which the heads
+	// decline and the cascade carries on to the resident model.
+	//
+	// 0.6 is the knee measured on the 88-case intent fixture
+	// (docs/evals/2026-08-08-routing-heads-in-go.md). It keeps 81 of 88 cases
+	// at 97.5% accuracy. Every higher value up to 0.9 drops right answers and
+	// keeps the same two wrong ones, so it buys nothing.
+	headsThreshold = 0.6
+)
+
+// RouterHeads runs the exported graph. Nil is a working value everywhere: a
+// deployment with no weights file routes exactly as it did before this
+// existed.
+type RouterHeads struct {
+	tokenizer *unigramTokenizer
+	session   *ort.DynamicSession[int64, float32]
+	intents   []Intent
+	sources   []Source
+	threshold float64
+}
+
+// headsMeta — router_heads.json, written beside the weights by the exporter.
+// The label order is the head's output order and cannot be inferred from Go.
+type headsMeta struct {
+	Intents []string `json:"intents"`
+	Sources []string `json:"sources"`
+	Prefix  string   `json:"prefix"`
+}
+
+// NewRouterHeads loads the graph and its label order. modelPath points at the
+// .onnx; the external weights and router_heads.json sit beside it.
+//
+// It assumes the ONNX environment is already initialised, because the embedder
+// does that at startup and the runtime allows it once.
+func NewRouterHeads(modelPath, tokenizerPath string) (*RouterHeads, error) {
+	metaPath := filepath.Join(filepath.Dir(modelPath), "router_heads.json")
+	raw, err := os.ReadFile(metaPath)
+	if err != nil {
+		return nil, fmt.Errorf("heads: read %s: %w", metaPath, err)
+	}
+	var meta headsMeta
+	if err := json.Unmarshal(raw, &meta); err != nil {
+		return nil, fmt.Errorf("heads: parse %s: %w", metaPath, err)
+	}
+	if meta.Prefix != queryPrefix {
+		return nil, fmt.Errorf("heads: trained with prefix %q, this build uses %q",
+			meta.Prefix, queryPrefix)
+	}
+
+	intents := make([]Intent, len(meta.Intents))
+	for i, s := range meta.Intents {
+		intents[i] = Intent(s)
+	}
+	sources := make([]Source, len(meta.Sources))
+	for i, s := range meta.Sources {
+		// SourceUnknown is not in Sources, because it is the absence of a
+		// choice. It is a class the head can emit, and the one it should emit
+		// often, so it is allowed here and nowhere else.
+		if s != string(SourceUnknown) && !ValidSource(Source(s)) {
+			return nil, fmt.Errorf("heads: unknown destination %q in %s", s, metaPath)
+		}
+		sources[i] = Source(s)
+	}
+
+	tok, err := newUnigramTokenizer(tokenizerPath)
+	if err != nil {
+		return nil, fmt.Errorf("heads: tokenizer: %w", err)
+	}
+	session, err := ort.NewDynamicSession[int64, float32](
+		modelPath,
+		[]string{"input_ids", "attention_mask"},
+		[]string{"intent", "source", "slots", "clarify"},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("heads: create session: %w", err)
+	}
+
+	return &RouterHeads{
+		tokenizer: tok,
+		session:   session,
+		intents:   intents,
+		sources:   sources,
+		threshold: headsThreshold,
+	}, nil
+}
+
+func (h *RouterHeads) Close() error {
+	if h == nil {
+		return nil
+	}
+	h.session.Destroy()
+	return nil
+}
+
+// headsResult — one forward pass, read back.
+type headsResult struct {
+	Intent     Intent
+	Source     Source
+	Confidence float64
+	Clarify    bool
+}
+
+// Route runs the heads and reports whether they are confident enough to answer.
+// A false second return is a decline, not an error: the cascade goes on to the
+// resident model, which is what happens today.
+func (h *RouterHeads) Route(ctx context.Context, utterance string) (headsResult, bool, error) {
+	if h == nil {
+		return headsResult{}, false, nil
+	}
+	ids, mask, _ := h.tokenizer.Encode(queryPrefix + utterance)
+	ids, mask = ids[:headsSeq], mask[:headsSeq]
+	// The tokenizer pads and truncates to its own length, which is longer than
+	// this one. Cutting the tail can cut the separator with it, so put it back.
+	if mask[headsSeq-1] == 1 {
+		ids[headsSeq-1] = sepTokenID
+	}
+
+	shape := ort.NewShape(1, headsSeq)
+	idsT, err := ort.NewTensor(shape, ids)
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: ids tensor: %w", err)
+	}
+	defer idsT.Destroy()
+	maskT, err := ort.NewTensor(shape, mask)
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: mask tensor: %w", err)
+	}
+	defer maskT.Destroy()
+
+	intentT, err := ort.NewEmptyTensor[float32](ort.NewShape(1, int64(len(h.intents))))
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: intent tensor: %w", err)
+	}
+	defer intentT.Destroy()
+	sourceT, err := ort.NewEmptyTensor[float32](ort.NewShape(1, int64(len(h.sources))))
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: source tensor: %w", err)
+	}
+	defer sourceT.Destroy()
+	slotsT, err := ort.NewEmptyTensor[float32](ort.NewShape(1, headsSeq, int64(numBIOTags)))
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: slots tensor: %w", err)
+	}
+	defer slotsT.Destroy()
+	clarifyT, err := ort.NewEmptyTensor[float32](ort.NewShape(1, 2))
+	if err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: clarify tensor: %w", err)
+	}
+	defer clarifyT.Destroy()
+
+	if err := h.session.Run(
+		[]*ort.Tensor[int64]{idsT, maskT},
+		[]*ort.Tensor[float32]{intentT, sourceT, slotsT, clarifyT},
+	); err != nil {
+		return headsResult{}, false, fmt.Errorf("heads: run: %w", err)
+	}
+
+	// The graph applies its own softmax, so these are probabilities and the max
+	// is the same number the eval calibrated the threshold against.
+	i, conf := argmax(intentT.GetData())
+	res := headsResult{
+		Intent:     h.intents[i],
+		Confidence: conf,
+	}
+	cl := clarifyT.GetData()
+	res.Clarify = len(cl) == 2 && cl[1] > cl[0]
+
+	// The destination head is trained on query rows and is meaningless on any
+	// other intent, the same way queryWalk is never reached by one.
+	if res.Intent == IntentQuery {
+		s, _ := argmax(sourceT.GetData())
+		res.Source = h.sources[s]
+	}
+
+	// The clarify head decides on its own, and it decides first. It answers a
+	// different question from the intent head — not which intent, but whether
+	// there is enough here to act on at all — so a low intent confidence is no
+	// reason to discard it. It is usually the same turns: "вода" reads as
+	// intent act at 0.23 and clarify at 0.98, and letting the intent threshold
+	// bury that hands the turn to the classifier, which routes it confidently
+	// and never asks.
+	if res.Clarify {
+		return res, true, nil
+	}
+	if conf < h.threshold {
+		return res, false, nil
+	}
+	return res, true, nil
+}
+
+// numBIOTags — O plus B- and I- for each of Maven's five slots. The head is not
+// read, but the graph writes it and the output tensor has to be the right size.
+const numBIOTags = 11
+
+func argmax(v []float32) (int, float64) {
+	best, bestV := 0, math.Inf(-1)
+	for i, x := range v {
+		if float64(x) > bestV {
+			best, bestV = i, float64(x)
+		}
+	}
+	return best, bestV
+}

internal/router/intent.go

@@ -105,6 +105,27 @@ type Decision struct {
 	Slots      Slots
 	Clarify    bool // stage 3: below threshold — ask, don't guess
 
+	// Source — where the answer lives, for a query. The second half of the
+	// route, and empty on every other intent. SourceUnknown means no decider
+	// named one and the daemon walks its whole chain, which is what shipped
+	// before this field existed. See source.go for why it is twelve values.
+	Source Source
+
+	// SourceAnchored — a stage 0 grammar named that destination, matching a
+	// literal pattern to do it. Only the router sets this, and only there.
+	//
+	// It exists because one thing downstream is not reversible by evidence
+	// (V-666). Naming a destination normally takes guessing sources off a turn,
+	// and one of those is the personal boundary, which is what stops a question
+	// about him from reaching the world. A grammar that read "что такое X" may
+	// take it off. A model or a softmax may not, because a wrong destination
+	// there widens what leaves the box rather than costing an answer.
+	//
+	// Read Stage instead and the two decisions get coupled: stage 0 also means
+	// confidence 1.0 and an anchored claim band, and a later cascade change
+	// could make one true where the other is not.
+	SourceAnchored bool
+
 	// Continued — this decision was rebuilt from the previous turn rather
 	// than routed, because the utterance was an ellipsis ("а завтра?").
 	// Handlers use it to know that Slots.Text is the PREVIOUS turn's topic

internal/router/llmrouter.go

@@ -45,12 +45,19 @@ const routeGrammar = `
 root   ::= "[" ws action ("," ws action)* ws "]"
 action ::= "{" ws "\"intent\"" ws ":" ws intent ("," ws field)* ws "}"
 intent ::= "\"fact\"" | "\"reminder\"" | "\"note\"" | "\"query\"" | "\"act\"" | "\"chat\"" | "\"system\"" | "\"unknown\""
-field  ::= key ws ":" ws string
+field  ::= (key ws ":" ws string) | ("\"source\"" ws ":" ws source)
 key    ::= "\"key\"" | "\"value\"" | "\"text\"" | "\"verb\""
+source ::= "\"recall\"" | "\"calendar\"" | "\"tasks\"" | "\"list\"" | "\"money\"" | "\"weather\"" | "\"home\"" | "\"network\"" | "\"feeds\"" | "\"attention\"" | "\"self\"" | "\"world\"" | "\"\""
 string ::= "\"" ([^"\\\x00-\x1F] | "\\" ["\\/bfnrt] | "\\u" [0-9a-fA-F]{4}){0,120} "\""
 ws     ::= [ \t\n]{0,4}
 `
 
+// TestRouteGrammarCoversSources holds the source rule above to router.Sources.
+// The enum is the point: a grammar cannot emit a destination that does not
+// exist, which is the guarantee V-546 wants from a softmax and gets here for
+// free. Empty is the thirteenth alternative and it is not an oversight — it is
+// the SourceUnknown floor, and the model must be able to decline.
+
 // routeSystem — the router prompt. Changed 31-07-2026: the query test now sits
 // above the fact test and there is an explicit question test. Before that, a
 // question naming a fact key ("сколько воды я выпил с утра") matched the fact
@@ -121,6 +128,31 @@ const routeSystem = `Классифицируй ровно одно сообще
 "что такое кватернион?" → {"intent":"query","text":"что такое кватернион"}
 "ага" → {"intent":"chat","text":"ага"}
 
+Только для query добавь поле source — где лежит ответ:
+- recall — его заметки, факты и то, что он раньше говорил
+- calendar — встречи и события
+- tasks — список задач
+- list — списки покупок и другие именованные списки
+- money — траты
+- weather — погода
+- home — свет, устройства, дом
+- network — локальная сеть, сервер, диски
+- feeds — новостные ленты
+- attention — что требует внимания сейчас
+- self — вопрос про самого ассистента
+- world — всё остальное: определения, счёт, люди, факты о мире
+
+Пустое значение "" — нормальный ответ и его надо ставить часто. Ставь "", если ответ могут дать сразу два источника или если не уверен: тогда проверяются все по порядку, и это правильно. Никогда не угадывай.
+
+"сколько воды я выпил с утра" → {"intent":"query","text":"сколько воды я выпил с утра","source":"recall"}
+"что я записывал про кота" → {"intent":"query","text":"что я записывал про кота","source":"recall"}
+"во сколько у меня встреча" → {"intent":"query","text":"во сколько у меня встреча","source":"calendar"}
+"что такое docker?" → {"intent":"query","text":"что такое docker","source":"world"}
+"кто такой Линус Торвальдс?" → {"intent":"query","text":"кто такой Линус Торвальдс","source":"world"}
+"сколько будет 17 на 23?" → {"intent":"query","text":"сколько будет 17 на 23","source":"world"}
+"почему сервер тормозит" → {"intent":"query","text":"почему сервер тормозит","source":""}
+"есть новости по бэкапу базы" → {"intent":"query","text":"есть новости по бэкапу базы","source":""}
+
 Ответ — JSON-массив: по одному объекту на каждую просьбу. Обычно один. Если в реплике несколько просьб — по объекту на каждую. "напомни купить молоко, и запиши что кофе кончился" → [{"intent":"reminder","text":"купить молоко"},{"intent":"note","text":"кофе кончился"}]. Только JSON, без пояснений.`
 
 // routeRepeatPenalty — the sub-1B model loops one sentence inside the text field
@@ -172,6 +204,7 @@ type routeAction struct {
 	Value  string `json:"value"`
 	Text   string `json:"text"`
 	Verb   string `json:"verb"`
+	Source string `json:"source"`
 }
 
 // Route asks the model for one decision. The bool is false when there is no
@@ -240,6 +273,14 @@ func (lr *LLMRouter) Route(ctx context.Context, utterance string, now time.Time)
 	case IntentQuery:
 		d.Intent = IntentQuery
 		d.Slots.Text = firstNonEmpty(a.Text, utterance)
+		// Through ValidSource, and on query alone. The grammar already bounds
+		// the enum, but the grammar is a request to a server that may be
+		// running a different build, and a destination this binary does not
+		// know would take real query sources off the turn. Anything unknown
+		// drops to SourceUnknown, which is the floor and costs nothing.
+		if ValidSource(Source(a.Source)) {
+			d.Source = Source(a.Source)
+		}
 	case IntentAct:
 		d.Intent = IntentAct
 		d.Slots.Text = firstNonEmpty(a.Verb, utterance)

internal/router/llmrouter_test.go

@@ -398,3 +398,70 @@ func TestLLMReminderWithSubjectIsNotGated(t *testing.T) {
 		t.Fatalf("a complete reminder was sent back as a question: %+v", d.Slots)
 	}
 }
+
+// TestRouteGrammarCoversSources — the grammar enum and router.Sources are two
+// hand-written lists of the same twelve destinations, and nothing else notices
+// when one grows. A destination missing from the grammar is a destination the
+// model is structurally unable to name, which is the exact defect V-517
+// measured for Praxis: not a weak model, an absent string.
+func TestRouteGrammarCoversSources(t *testing.T) {
+	for _, s := range Sources {
+		if !strings.Contains(routeGrammar, `"\"`+string(s)+`\""`) {
+			t.Errorf("routeGrammar cannot emit %q — the model can never name it", s)
+		}
+	}
+	// The floor has to be reachable too, or the model is forced to pick one.
+	if !strings.Contains(routeGrammar, `"\"\""`) {
+		t.Error(`routeGrammar cannot emit "" — the model cannot decline a destination`)
+	}
+	// Count the alternatives on the source rule: an extra one is a destination
+	// the daemon would drop to SourceUnknown after the model spent tokens on it.
+	for _, line := range strings.Split(routeGrammar, "\n") {
+		if !strings.HasPrefix(line, "source ") {
+			continue
+		}
+		if got, want := strings.Count(line, "|")+1, len(Sources)+1; got != want {
+			t.Errorf("source rule has %d alternatives, want %d (Sources plus the floor)", got, want)
+		}
+	}
+}
+
+// The destination is read back only through ValidSource. A model on an older or
+// newer build can write a string this binary does not know, and trusting it
+// would take real query sources off the turn for a name nothing answers.
+func TestLLMUnknownSourceFallsToTheFloor(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"query","text":"что там с бэкапами","source":"praxis"}`)
+	d, err := r.Route(context.Background(), "что там с бэкапами", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Source != SourceUnknown {
+		t.Fatalf("invented destination %q was trusted, want the floor", d.Source)
+	}
+}
+
+// And a known one survives, or the read-back is just a filter.
+func TestLLMNamedSourceSurvives(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"query","text":"кто такой Линус Торвальдс","source":"world"}`)
+	d, err := r.Route(context.Background(), "кто такой Линус Торвальдс?", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Source != SourceWorld {
+		t.Fatalf("source %q, want %q", d.Source, SourceWorld)
+	}
+}
+
+// A destination on anything but a query is dropped. Only IntentQuery reaches
+// queryWalk, so a source elsewhere is a field nobody reads and a claim nobody
+// checks.
+func TestLLMSourceIsQueryOnly(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"note","text":"кофе кончился","source":"recall"}`)
+	d, err := r.Route(context.Background(), "запиши что кофе кончился", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Source != SourceUnknown {
+		t.Fatalf("a note carried destination %q", d.Source)
+	}
+}

internal/router/onnxembedder.go

@@ -66,12 +66,19 @@ func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, e
 func (e *onnxEmbedder) Dim() int { return embedDim }
 
 // ID names the loaded model for the DB marker (Vikunja #378): the model file's
-// own name plus the dimension, so pointing the config at another model changes
-// the string on its own.
+// own name, the dimension, and the tokenizer revision, so pointing the config
+// at another model changes the string on its own.
 func (e *onnxEmbedder) ID() string { return e.id }
 
+// tokenizerRev — bumped whenever the tokenizer changes what it emits for the
+// same text, because that changes every vector while the model file's name
+// stays put. Rev 2 is the fix for the reversed word pieces (V-664): stored
+// passages embedded under rev 1 no longer sit in the same space as a query
+// embedded now, and ReembedAll rewrites them because this string moved.
+const tokenizerRev = 2
+
 // modelIDFromPath turns /opt/.../multilingual-e5-small.onnx into
-// "multilingual-e5-small@384".
+// "multilingual-e5-small@384/tok2".
 func modelIDFromPath(modelPath string) string {
 	name := modelPath
 	if i := strings.LastIndexAny(name, "/\\"); i >= 0 {
@@ -81,7 +88,7 @@ func modelIDFromPath(modelPath string) string {
 	if name == "" {
 		name = "onnx"
 	}
-	return fmt.Sprintf("%s@%d", name, embedDim)
+	return fmt.Sprintf("%s@%d/tok%d", name, embedDim, tokenizerRev)
 }
 
 // Embed treats the text as a query. The classifier compares one short
@@ -339,14 +346,17 @@ func (t *unigramTokenizer) encodeWord(word string) []int64 {
 		}
 	}
 
+	// Backtracking walks the word from its end, and prepending each piece puts
+	// it back in reading order. There used to be a second reverse after this
+	// loop, which undid it: every multi-piece word came out backwards, and
+	// "query: вода" tokenized to [0 12 1294 41 12489 2] where the reference
+	// tokenizer gives [0 41 1294 12 12489 2] (V-664). A transformer reads
+	// position, so the pieces of a long Russian word were being read in the
+	// wrong order on every turn.
 	var result []int64
 	for i := n; i > 0; i = prev[i] {
 		result = append([]int64{bestID[i]}, result...)
 	}
-	// Reverse
-	for l, r := 0, len(result)-1; l < r; l, r = l+1, r-1 {
-		result[l], result[r] = result[r], result[l]
-	}
 	return result
 }
 

internal/router/router.go

@@ -31,6 +31,12 @@ type Config struct {
 	// error/parse failure, falls through to the classifier (never fails the
 	// turn on the model).
 	LLM *LLMRouter
+	// Heads — optional routing heads over the fine-tuned embedder copy. When
+	// set, Route consults them after stage 0 and before the LLM router. They
+	// decline below their own confidence threshold, so a low-confidence turn
+	// reaches the model exactly as it does today. Nil is the shipped-before
+	// behaviour and costs nothing.
+	Heads *RouterHeads
 }
 
 // Router — the deterministic cascade. Route never guesses: stage 0 wins
@@ -42,6 +48,7 @@ type Router struct {
 	extractor  Extractor
 	threshold  float64
 	llm        *LLMRouter
+	heads      *RouterHeads
 }
 
 func New(cfg Config) *Router {
@@ -51,6 +58,7 @@ func New(cfg Config) *Router {
 		extractor:  cfg.Extractor,
 		threshold:  cfg.Threshold,
 		llm:        cfg.LLM,
+		heads:      cfg.Heads,
 	}
 }
 
@@ -90,6 +98,10 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
 			continue // grammar matched shape but not content → fall through
 		}
 		d.Utterance = utterance
+		// A literal pattern named that destination, which is the one provenance
+		// allowed to take the personal boundary off a turn (V-666). Set here and
+		// nowhere else, so no other arm of the cascade can claim it.
+		d.SourceAnchored = d.Source != SourceUnknown
 		// The grammar decided the intent; the extractor fills the slots it did
 		// not match (V-572). See fillMatchedSlots for why every grammar gets it.
 		r.fillMatchedSlots(ctx, &d, now)
@@ -98,6 +110,60 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
 	}
 	r.noteGrammarOutcomes(ctx, len(r.grammars), declinedBuild, "", "")
 
+	// stage 0b — routing heads (when wired). A softmax over the label set, so
+	// it cannot name an intent or a destination that does not exist, and its
+	// max is a real confidence. It runs before the model because it is three
+	// orders of magnitude faster and scores better on both halves of the route.
+	//
+	// It declines below its threshold rather than clarifying. A declined turn
+	// carries on to the model and then the classifier, which is what a box with
+	// no weights file does on every turn.
+	if r.heads != nil {
+		res, ok, err := r.heads.Route(ctx, utterance)
+		switch {
+		case err != nil:
+			log.Printf("router: heads fell through to the rest of the cascade: %v", err)
+			decision.Note(ctx, decision.Claim{
+				Stage: decision.StageRoute, Claimant: claimantHeads,
+				Outcome: decision.Declined, Reason: "error: " + err.Error(),
+			})
+		case !ok:
+			decision.Note(ctx, decision.Scored(decision.StageRoute, claimantHeads,
+				string(res.Intent), res.Confidence, decision.Declined,
+				"below the heads confidence threshold"))
+		default:
+			d := Decision{
+				Utterance:  utterance,
+				Stage:      2,
+				Intent:     res.Intent,
+				Confidence: res.Confidence,
+				Source:     res.Source,
+				Clarify:    res.Clarify,
+			}
+			r.fillSlots(ctx, &d, now)
+			decision.Note(ctx, decision.Claim{
+				Stage: decision.StageRoute, Claimant: claimantLLM,
+				Outcome: decision.NeverAsked, Reason: "the routing heads answered",
+			})
+			decision.Note(ctx, decision.Claim{
+				Stage: decision.StageRoute, Claimant: claimantClassifier,
+				Outcome: decision.NeverAsked, Reason: "the routing heads answered",
+			})
+			outcome, reason := decision.Won, ""
+			if d.Clarify {
+				outcome, reason = decision.Thinned, "the clarify head says there is too little here to act on"
+			}
+			decision.Note(ctx, decision.Scored(decision.StageRoute, claimantHeads,
+				string(d.Intent), d.Confidence, outcome, reason))
+			return d, nil
+		}
+	} else {
+		decision.Note(ctx, decision.Claim{
+			Stage: decision.StageRoute, Claimant: claimantHeads,
+			Outcome: decision.NeverAsked, Reason: "no routing heads are wired",
+		})
+	}
+
 	// stage 1a — LLM router (when wired). It reasons over the utterance instead
 	// of nearest-centroid guessing. On any error/parse-fail, fall through to the
 	// classifier cascade (never fail the turn on the model).

internal/router/source.go

@@ -0,0 +1,78 @@
+package router
+
+// Source — where the answer to a query lives. It is the second half of a
+// routing decision and it used to be made outside the router entirely (V-655).
+//
+// The cascade sorted an utterance into one of seven intents with stage 0 rules,
+// the resident model and the classifier behind it, a fixture measuring it and
+// the decision trace recording it. Then IntentQuery handed the turn to
+// querySources in the daemon, a chain of twenty-two branches deciding by seed
+// similarity in a fixed order, with none of that. So the careful sorter did the
+// easy half and the sloppy one did the hard half: on 2026-08-07 weather claimed
+// "что такое TCP?" and answered "для какого города?", because weather read one
+// percent closer to the turn than the pile of leftover seeds did, and one
+// percent was enough. Search would have answered it and search was never asked.
+//
+// "query" is not a destination. It is a shrug. This is the field that says
+// where to look.
+//
+// # Why twelve and not twenty-two
+//
+// A destination is what a decider can plausibly name from the utterance alone,
+// not one entry per source. Three of the daemon's sources are successive passes
+// over his own words and a fourth reads the facts by key: which of them lands
+// the hit is an ordering detail inside the chain, and no utterance says. They
+// are SourceRecall together. The same goes for the metasearch, the offline
+// encyclopedia and a page he named by URL, which are SourceWorld.
+//
+// # Empty is a real value and it is the floor
+//
+// SourceUnknown means nobody decided. The daemon then walks the whole chain in
+// its original order, which is the behaviour that shipped before this field
+// existed. So the classifier arm sets nothing and costs nothing, and a box
+// whose model is down routes queries exactly as it did.
+type Source string
+
+const (
+	// SourceUnknown — no decider named a destination. Walk the chain.
+	SourceUnknown Source = ""
+
+	// His own data.
+	SourceRecall   Source = "recall"   // notes, facts and what he has said before
+	SourceCalendar Source = "calendar" // events, and the only date-aware destination
+	SourceTasks    Source = "tasks"    // the task list
+	SourceList     Source = "list"     // the shopping and other named lists
+	SourceMoney    Source = "money"    // the spending facts the poller writes
+
+	// The surroundings.
+	SourceWeather   Source = "weather"   // the forecast for a place
+	SourceHome      Source = "home"      // lights, devices, the house
+	SourceNetwork   Source = "network"   // the LAN and what is on it
+	SourceFeeds     Source = "feeds"     // the RSS she reads
+	SourceAttention Source = "attention" // what Praxis says needs looking at
+
+	// Everything else.
+	SourceSelf  Source = "self"  // a question about Maven herself
+	SourceWorld Source = "world" // search, the ZIMs, a page he named
+)
+
+// Sources — every destination a decider may name, in a fixed order so a prompt,
+// a grammar table and a test all read the same list. SourceUnknown is not a
+// member: it is the absence of a choice, not one of the choices.
+var Sources = []Source{
+	SourceRecall, SourceCalendar, SourceTasks, SourceList, SourceMoney,
+	SourceWeather, SourceHome, SourceNetwork, SourceFeeds, SourceAttention,
+	SourceSelf, SourceWorld,
+}
+
+// ValidSource reports whether s is one a decider may name. Anything else,
+// including a destination invented by a model, is dropped back to
+// SourceUnknown by the caller rather than trusted.
+func ValidSource(s Source) bool {
+	for _, known := range Sources {
+		if s == known {
+			return true
+		}
+	}
+	return false
+}

internal/router/stage0.go

@@ -187,9 +187,11 @@ func SystemTimeDateGrammars() []Grammar {
 // written ("the clock/date system rule must not swallow it"); the daemon
 // disagreed with the fixture and the daemon was wrong.
 //
-// Routing, not answering. These set the intent and nothing else — which source
-// in the query chain claims the turn stays the chain's decision, and a
-// question with no date still falls through queryCalendar to recall.
+// Routing, not answering. Two of the five also name the calendar as the
+// destination (V-655), which narrows who may GUESS their way onto the turn and
+// claims nothing. Every source that looks something up still runs, in the order
+// it always did, so a question with no date still falls through queryCalendar
+// to recall.
 //
 // Deliberately not folded into SystemTimeDateGrammars: those exist to send
 // utterances TO system, these exist to keep utterances OUT of it, and one
@@ -199,9 +201,15 @@ func AgendaQueryGrammars() []Grammar {
 		{
 			// An explicit calendar noun is unambiguous wherever it appears:
 			// "что в календаре на завтра", "покажи расписание на среду".
+			//
+			// The one agenda rule that names its destination, because an
+			// explicit calendar noun leaves nothing to weigh (V-655). The
+			// possessive rules below deliberately do not: "что у меня в списке
+			// покупок" matches agenda-query, and naming the calendar there
+			// would take the list source off the turn.
 			Name:    "calendar-query",
 			Pattern: regexp.MustCompile(`(?i)(календар|расписани|повестк)`),
-			Build:   agendaQueryBuild,
+			Build:   queryTo(SourceCalendar),
 		},
 		{
 			// The agenda phrasing with no calendar noun. Anchored at the start
@@ -251,9 +259,11 @@ func AgendaQueryGrammars() []Grammar {
 			// "во сколько созвон". He is asking when something on his calendar
 			// happens, and the noun is the only signal. Closed list, so "когда
 			// битва при Ватерлоо" is still a world question.
+			// Names the calendar (V-655): the noun list is closed and every
+			// member of it is an event, so there is nothing else to weigh.
 			Name:    "event-time-query",
 			Pattern: regexp.MustCompile(`(?i)^\s*(когда|во\s+сколько|в\s+котором\s+часу)\s+(будет\s+|у\s+нас\s+)?(планёрк|планерк|встреч|созвон|митинг|совещани|звонок|созвон|приём|прием|интервью|собеседовани|тренировк|урок|занятие|пара)[а-я]*(\s|[?!.]|$)`),
-			Build:   agendaQueryBuild,
+			Build:   queryTo(SourceCalendar),
 		},
 	}
 }
@@ -340,6 +350,11 @@ func narrativeQueryBuild(m []string) (Decision, bool) {
 		Intent:     IntentQuery,
 		Confidence: 1.0,
 		Slots:      Slots{Text: topic},
+		// The world, because that is the shape this asks for and the rule has
+		// already declined the two cases where it is not: entertainment, and
+		// questions about her (V-655). His own notes are still read first — a
+		// destination narrows who may guess and reorders nothing.
+		Source: SourceWorld,
 	}, true
 }
 

internal/router/worldquery.go

@@ -0,0 +1,86 @@
+package router
+
+import "regexp"
+
+// WorldQueryGrammars — stage-0 rules for the two question shapes that name the
+// world in their own words, and say so plainly enough that no scorer is needed
+// (V-655).
+//
+// They exist because of what happens when nothing deterministic claims these.
+// Measured on the box on 2026-08-07 (docs/evals/2026-08-07-week-of-usage.md,
+// section 4): "что такое TCP?" and "сколько будет 17 на 23?" were both answered
+// "для какого города?", and "кто такой Линус Торвальдс?" was answered "не знаю —
+// не нашла у тебя такой записи". None of those three is about him, about the
+// weather, or about anything on this box.
+//
+// The mechanism is the destination, not the answer. Naming SourceWorld does not
+// send the turn outside and does not skip a single source that looks something
+// up: his notes, his facts and the personal boundary all still run first, in the
+// order they always did. What it does is stop the sources that claim on seed
+// similarity from taking the turn on the way past. Weather cannot claim a
+// question about a protocol once the utterance has said which side it is on.
+//
+// Both patterns are spelled out here rather than drawn from internal/lexicon,
+// which is the same call the agenda rules made: these are interrogative FRAMES
+// of two words, not a closed class of single words, and the lexicon holds
+// classes. Nothing here is a stem pattern over open vocabulary — the variable
+// part of each rule is the topic, and the rule reads none of it.
+func WorldQueryGrammars() []Grammar {
+	return []Grammar{
+		{
+			// "что такое X", "кто такой X". A request for what a thing or a
+			// person IS, which his own data can answer and usually cannot.
+			//
+			// The topic is deliberately not captured into Slots.Text. Every
+			// source below reads the utterance, "что такое TCP?" is already the
+			// best query string for it, and the agenda rules make the same call
+			// for the same reason.
+			Name:    "definition-query",
+			Pattern: definitionQueryPattern,
+			Build:   queryTo(SourceWorld),
+		},
+		{
+			// "сколько будет 17 на 23", "сколько будет 2+2". Arithmetic, which
+			// the metasearch answers and no local source holds. The digits are
+			// what make it arithmetic: "сколько будет гостей" names no number
+			// and is a question about his evening.
+			Name:    "arithmetic-query",
+			Pattern: arithmeticQueryPattern,
+			Build:   queryTo(SourceWorld),
+		},
+	}
+}
+
+// definitionQueryPattern — anchored at the start, because "напомни узнать что
+// такое TCP" is a reminder that happens to contain the frame.
+//
+// (\s|[?!.]|$) and not \b: Go's \b is ASCII-only and never fires after a
+// Cyrillic letter, so the ASCII form silently matches nothing. The agenda rules
+// carry the same note.
+var definitionQueryPattern = regexp.MustCompile(
+	`(?i)^\s*(что\s+так(ое|ая)|кто\s+так(ой|ая|ие)|what\s+is|who\s+is)(\s|[?!.]|$)`)
+
+// arithmeticQueryPattern — the ask, then a digit somewhere after it. Loose on
+// what sits between them on purpose: the operator is spoken half a dozen ways
+// ("на", "умножить на", "плюс", "+") and reading them is the calculator's job,
+// not this rule's. All this decides is which side of the boundary the turn is
+// on.
+var arithmeticQueryPattern = regexp.MustCompile(
+	`(?i)^\s*(сколько\s+будет|посчитай|вычисли|how\s+much\s+is)\s.*\d`)
+
+// queryTo builds a stage-0 query Decision that names where the answer lives.
+//
+// The utterance travels intact and no slot is filled, which is the same
+// contract agendaQueryBuild has: confidence 1.0 on the intent and the
+// destination, and every source below still decides for itself whether it has
+// an answer. Naming a destination narrows who may guess. It promises nothing.
+func queryTo(dest Source) func([]string) (Decision, bool) {
+	return func([]string) (Decision, bool) {
+		return Decision{
+			Stage:      0,
+			Intent:     IntentQuery,
+			Confidence: 1.0,
+			Source:     dest,
+		}, true
+	}
+}

internal/router/worldquery_test.go

@@ -0,0 +1,88 @@
+package router
+
+import "testing"
+
+// The three utterances from the 2026-08-07 week on the box that no local source
+// could answer and three different local sources claimed anyway. Stage 0 has to
+// say which side of the boundary they are on, because by the time the chain is
+// walking, the only thing separating them from a weather forecast is a cosine.
+func TestAWorldQuestionNamesTheWorld(t *testing.T) {
+	cases := []struct {
+		utterance string
+		rule      string
+	}{
+		{"что такое TCP?", "definition-query"},
+		{"кто такой Линус Торвальдс?", "definition-query"},
+		{"что такая мембрана", "definition-query"},
+		{"кто такая Ада Лавлейс?", "definition-query"},
+		{"what is TCP?", "definition-query"},
+		{"сколько будет 17 на 23?", "arithmetic-query"},
+		{"посчитай 2+2", "arithmetic-query"},
+		{"сколько будет 5 умножить на 6", "arithmetic-query"},
+	}
+	for _, c := range cases {
+		dec, rule, ok := matchWorldQuery(c.utterance)
+		if !ok {
+			t.Errorf("%q: no world rule claimed it", c.utterance)
+			continue
+		}
+		if rule != c.rule {
+			t.Errorf("%q: claimed by %q, want %q", c.utterance, rule, c.rule)
+		}
+		if dec.Intent != IntentQuery {
+			t.Errorf("%q: intent %q, want query", c.utterance, dec.Intent)
+		}
+		if dec.Source != SourceWorld {
+			t.Errorf("%q: source %q, want %q", c.utterance, dec.Source, SourceWorld)
+		}
+	}
+}
+
+// The frame has to be the whole opening or the rule is reading somebody else's
+// sentence. Every case here contains a world-question shape and is not one.
+func TestAWorldRuleDeclinesWhatIsNotItsShape(t *testing.T) {
+	cases := []struct {
+		utterance string
+		why       string
+	}{
+		{"напомни узнать что такое TCP", "a reminder that happens to quote the frame"},
+		{"запиши что такое TCP", "a capture that happens to quote the frame"},
+		{"сколько будет гостей", "an ask with no number is not arithmetic"},
+		{"что у меня сегодня?", "his agenda, and the agenda rules own it"},
+		{"кто там?", "not the frame"},
+		{"посчитай расходы", "no number, so the money source keeps it"},
+	}
+	for _, c := range cases {
+		if _, rule, ok := matchWorldQuery(c.utterance); ok {
+			t.Errorf("%q: claimed by %q, want no claim — %s", c.utterance, rule, c.why)
+		}
+	}
+}
+
+// The destination is advice about who may guess, never a filled slot. A rule
+// that quietly captured the topic would change what every source below reads.
+func TestNamingTheWorldFillsNoSlot(t *testing.T) {
+	dec, _, ok := matchWorldQuery("что такое TCP?")
+	if !ok {
+		t.Fatal("definition-query did not claim it")
+	}
+	if dec.Slots.Text != "" || dec.Slots.HasTime || dec.Slots.HasFn || dec.Slots.HasKey {
+		t.Errorf("slots = %+v, want none filled", dec.Slots)
+	}
+	if dec.Confidence != 1.0 {
+		t.Errorf("confidence = %v, want 1.0 for a stage-0 match", dec.Confidence)
+	}
+}
+
+func matchWorldQuery(utterance string) (Decision, string, bool) {
+	for _, g := range WorldQueryGrammars() {
+		m := g.Pattern.FindStringSubmatch(utterance)
+		if m == nil {
+			continue
+		}
+		if dec, ok := g.Build(m); ok {
+			return dec, g.Name, true
+		}
+	}
+	return Decision{}, "", false
+}

internal/rss/poller.go

@@ -89,8 +89,8 @@ type Poller struct {
 	ranker  Ranker
 	cfg     Config
 	nextDue map[string]time.Time
-	seen    map[string]map[string]bool // feed → item ID, for items with no date
-	polled  map[string]bool            // feed → polled at least once in THIS process
+	seen    map[string]*seenIDs // feed → item IDs, for items with no date
+	polled  map[string]bool     // feed → polled at least once in THIS process
 }
 
 // NewPoller wires a poller. Returns nil when there is nothing to poll — a
@@ -122,7 +122,7 @@ func NewPoller(feeds []FeedConfig, fetch Fetcher, notes Notes, marks Marks, embe
 		feeds: valid, fetch: fetch, notes: notes, marks: marks,
 		embed: embed, ranker: ranker, cfg: cfg,
 		nextDue: map[string]time.Time{},
-		seen:    map[string]map[string]bool{},
+		seen:    map[string]*seenIDs{},
 		polled:  map[string]bool{},
 	}
 }
@@ -283,6 +283,38 @@ func (p *Poller) mark(ctx context.Context, feed string, now time.Time) (time.Tim
 	return at, true
 }
 
+// maxSeenPerFeed bounds the undated-item set. It has to stay comfortably above
+// any one feed's front page, or an item still listed there would fall out of the
+// set and be written a second time. A few hundred entries covers the largest
+// page anyone publishes, and the set only has to span one poll window plus the
+// resync guard, not all of history.
+const maxSeenPerFeed = 512
+
+// seenIDs is a bounded insertion-ordered set. The map answers the lookup, the
+// slice remembers what to drop first, so an undated feed cannot grow the poller
+// for as long as mavend runs.
+type seenIDs struct {
+	ids   map[string]bool
+	order []string
+}
+
+// add records id and reports whether it was new.
+func (s *seenIDs) add(id string) bool {
+	if s.ids == nil {
+		s.ids = make(map[string]bool, maxSeenPerFeed)
+	}
+	if s.ids[id] {
+		return false
+	}
+	s.ids[id] = true
+	s.order = append(s.order, id)
+	if len(s.order) > maxSeenPerFeed {
+		delete(s.ids, s.order[0])
+		s.order = s.order[1:]
+	}
+	return true
+}
+
 // fresh — two dedup rules, because feeds are inconsistent about dates. A dated
 // item must be newer than the mark; an undated one is kept once per process by
 // ID.
@@ -308,12 +340,11 @@ func (p *Poller) fresh(f FeedConfig, it Item, mark, now time.Time, resync bool)
 		id = it.Title
 	}
 	if p.seen[f.Name] == nil {
-		p.seen[f.Name] = map[string]bool{}
+		p.seen[f.Name] = &seenIDs{}
 	}
-	if p.seen[f.Name][id] {
+	if !p.seen[f.Name].add(id) {
 		return false
 	}
-	p.seen[f.Name][id] = true
 	return !resync
 }
 

internal/rss/poller_test.go

@@ -3,6 +3,7 @@ package rss
 import (
 	"context"
 	"errors"
+	"fmt"
 	"strings"
 	"testing"
 	"time"
@@ -208,3 +209,27 @@ func TestNoFeedsMeansNoPoller(t *testing.T) {
 		t.Fatal("a feed with no name or url is not a configuration")
 	}
 }
+
+// An undated feed used to grow p.seen for as long as mavend ran. The set is
+// bounded now, and the bound must not cost the dedupe an item still on the
+// front page — only ids far older than any page fall out.
+func TestSeenIDsBounded(t *testing.T) {
+	var s seenIDs
+	for i := 0; i < maxSeenPerFeed*3; i++ {
+		if !s.add(fmt.Sprintf("item-%d", i)) {
+			t.Fatalf("item-%d read as already seen", i)
+		}
+		if len(s.ids) > maxSeenPerFeed || len(s.order) > maxSeenPerFeed {
+			t.Fatalf("after %d inserts: ids=%d order=%d, cap is %d",
+				i+1, len(s.ids), len(s.order), maxSeenPerFeed)
+		}
+	}
+	// The newest insert is still deduped; the oldest was evicted.
+	last := fmt.Sprintf("item-%d", maxSeenPerFeed*3-1)
+	if s.add(last) {
+		t.Fatalf("%s read as new, so the most recent id was dropped", last)
+	}
+	if !s.add("item-0") {
+		t.Fatal("item-0 survived, so nothing was evicted")
+	}
+}

internal/store/conncap_test.go

@@ -0,0 +1,154 @@
+package store
+
+import (
+	"context"
+	"database/sql"
+	"fmt"
+	"path/filepath"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// conncap_test.go measures whether a read queues behind a write at
+// SetMaxOpenConns(1), which is what openAt sets (V-642). It is a measurement
+// harness, not an assertion: the numbers it prints are the evidence, and the
+// decision to move the cap or leave it belongs in docs/evals.
+//
+// Run it with -v, and note that it is skipped under -short because it spends
+// seconds on purpose.
+
+// openCapped opens a plaintext store at the given connection cap. In-package,
+// so it can reach the handle openAt caps at 1.
+func openCapped(t *testing.T, cap int) *Store {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "cap.db")
+	db, err := openAt(context.Background(), path)
+	if err != nil {
+		t.Fatalf("openAt: %v", err)
+	}
+	db.SetMaxOpenConns(cap)
+	s := &Store{db: db}
+	t.Cleanup(func() { _ = s.Close() })
+	return s
+}
+
+func percentile(d []time.Duration, p float64) time.Duration {
+	if len(d) == 0 {
+		return 0
+	}
+	i := int(float64(len(d)-1) * p)
+	return d[i]
+}
+
+// seedFacts writes n facts so a read has rows to decode.
+func seedFacts(t *testing.T, s *Store, n int) {
+	t.Helper()
+	ctx := context.Background()
+	now := time.Now().UTC()
+	for i := 0; i < n; i++ {
+		key := fmt.Sprintf("seed_%d", i)
+		if _, err := s.SetValue(ctx, KindSelf, key, "tap:test",
+			map[string]int{"ml": i}, now.Add(time.Duration(i)*time.Millisecond)); err != nil {
+			t.Fatalf("seed %d: %v", i, err)
+		}
+	}
+}
+
+// measureReadsUnderWrites reports read latency percentiles while a writer
+// writes at a fixed pace. The pace matters: an unpaced writer completes a
+// different number of writes at each cap, because at a higher cap it competes
+// with the readers for the write lock instead of taking turns on one
+// connection. Two runs that did different work cannot be compared.
+// It runs for a fixed wall-clock window rather than a fixed read count, so the
+// paced writer does the same work at every cap. Tying the window to a read
+// count made the faster configuration receive fewer writes.
+func measureReadsUnderWrites(t *testing.T, s *Store, window, pace time.Duration) []time.Duration {
+	t.Helper()
+	ctx := context.Background()
+
+	var stop atomic.Bool
+	var writes atomic.Int64
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
+		now := time.Now().UTC()
+		for i := 0; !stop.Load(); i++ {
+			key := fmt.Sprintf("hot_%d", i%16)
+			if _, err := s.SetValue(ctx, KindSelf, key, "tap:test",
+				map[string]int{"n": i}, now.Add(time.Duration(i)*time.Millisecond)); err != nil {
+				t.Errorf("write: %v", err)
+				return
+			}
+			writes.Add(1)
+			time.Sleep(pace)
+		}
+	}()
+
+	var lat []time.Duration
+	deadline := time.Now().Add(window)
+	for time.Now().Before(deadline) {
+		start := time.Now()
+		if _, err := s.RecentFacts(ctx, 50); err != nil {
+			t.Fatalf("RecentFacts: %v", err)
+		}
+		lat = append(lat, time.Since(start))
+	}
+	stop.Store(true)
+	wg.Wait()
+	t.Logf("in %v: %d reads, %d writes", window, len(lat), writes.Load())
+
+	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
+	return lat
+}
+
+// TestConnCap_ReadLatencyUnderWrites is the V-642 measurement: read latency at
+// cap 1 against cap 4, same workload, same schema, same driver.
+func TestConnCap_ReadLatencyUnderWrites(t *testing.T) {
+	if testing.Short() {
+		t.Skip("measurement harness; runs for seconds")
+	}
+	for _, cap := range []int{1, 4} {
+		t.Run(fmt.Sprintf("cap=%d", cap), func(t *testing.T) {
+			s := openCapped(t, cap)
+			seedFacts(t, s, 500)
+			lat := measureReadsUnderWrites(t, s, 2*time.Second, 2*time.Millisecond)
+			t.Logf("cap=%d reads=%d p50=%v p95=%v max=%v",
+				cap, len(lat), percentile(lat, 0.50), percentile(lat, 0.95), lat[len(lat)-1])
+		})
+	}
+}
+
+// TestConnCap_ReadBlocksBehindOpenSnapshot is the sharper claim: at cap 1 an
+// open read-only transaction holds the only connection, so an unrelated read
+// cannot proceed until it commits. This is why the store exposes no way to
+// begin one — `Store.DB` used to, and was deleted in V-642 with no caller. The
+// test stays as the reason, so re-adding that seam fails a measurement rather
+// than shipping a stall.
+func TestConnCap_ReadBlocksBehindOpenSnapshot(t *testing.T) {
+	if testing.Short() {
+		t.Skip("measurement harness; waits on a timeout")
+	}
+	for _, cap := range []int{1, 4} {
+		t.Run(fmt.Sprintf("cap=%d", cap), func(t *testing.T) {
+			s := openCapped(t, cap)
+			seedFacts(t, s, 50)
+
+			tx, err := s.db.BeginTx(context.Background(), &sql.TxOptions{ReadOnly: true})
+			if err != nil {
+				t.Fatalf("BeginTx: %v", err)
+			}
+			defer func() { _ = tx.Rollback() }()
+
+			ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+			defer cancel()
+			start := time.Now()
+			_, err = s.RecentFacts(ctx, 10)
+			t.Logf("cap=%d read alongside an open snapshot: waited %v, err=%v",
+				cap, time.Since(start).Round(time.Millisecond), err)
+		})
+	}
+}

internal/store/memory.go

@@ -68,6 +68,13 @@ func (m *MemoryStore) Insert(ctx context.Context, id string, vec []float32, meta
 // Rows under memory.NonRecallPrefix are excluded in SQL. They are speaker
 // voiceprints sharing this table, and note recall must not rank them; see that
 // constant for why the previous arrangement only appeared to do this.
+//
+// Every row is still scored, because a full scan is what picks the winners.
+// What the scan does NOT do is pay for a row it is about to discard: the score
+// is read straight off the stored bytes without materializing a []float32, and
+// the meta blob is copied and unmarshalled only for a row that has entered the
+// topK. Losers cost one dot product and nothing else. Ranking is unchanged —
+// same scores, same order, same ties.
 func (m *MemoryStore) Search(ctx context.Context, vec []float32, topK int) ([]memory.Result, error) {
 	if topK <= 0 {
 		topK = 10
@@ -80,30 +87,127 @@ func (m *MemoryStore) Search(ctx context.Context, vec []float32, topK int) ([]me
 	}
 	defer rows.Close()
 
-	var out []memory.Result
+	// sql.RawBytes hands us the driver's own buffer, valid only until the next
+	// Next(). Nothing here outlives the row except what topK.offer copies on a
+	// survivor, so the three columns cost no allocation per row.
+	var id, blob, metaJSON sql.RawBytes
+	top := newTopK(topK)
 	for rows.Next() {
-		var id, metaJSON string
-		var blob []byte
 		if err := rows.Scan(&id, &blob, &metaJSON); err != nil {
 			return nil, fmt.Errorf("memory: row: %w", err)
 		}
-		meta := map[string]string{}
-		if err := json.Unmarshal([]byte(metaJSON), &meta); err != nil {
-			return nil, fmt.Errorf("memory: unmarshal meta for %q: %w", id, err)
-		}
-		out = append(out, memory.Result{ID: id, Score: dot(vec, decodeVec(blob)), Meta: meta})
+		top.offer(dotBlob(vec, blob), id, metaJSON)
 	}
 	if err := rows.Err(); err != nil {
 		return nil, fmt.Errorf("memory: rows: %w", err)
 	}
 
-	sort.Slice(out, func(i, j int) bool { return out[i].Score > out[j].Score })
-	if topK < len(out) {
-		out = out[:topK]
+	survivors := top.sorted()
+	out := make([]memory.Result, 0, len(survivors))
+	for _, c := range survivors {
+		meta := map[string]string{}
+		if err := json.Unmarshal(c.meta, &meta); err != nil {
+			return nil, fmt.Errorf("memory: unmarshal meta for %q: %w", c.id, err)
+		}
+		out = append(out, memory.Result{ID: c.id, Score: c.score, Meta: meta})
 	}
 	return out, nil
 }
 
+// candidate is one row that is currently in the topK: its score, its id, and
+// its meta blob copied out of the driver's buffer. The copy is the price of
+// surviving, and only survivors pay it.
+type candidate struct {
+	score float64
+	id    string
+	meta  []byte
+}
+
+// topK keeps the k highest-scoring candidates seen so far as a min-heap, so the
+// weakest survivor is always heap[0] and one comparison decides whether a new
+// row is worth copying. k is 10 in practice, so the heap is tiny and the whole
+// structure fits in cache.
+//
+// It is a plain slice with hand-written sift operations rather than
+// container/heap, because that interface boxes every element into an `any` on
+// Push and costs an allocation per surviving row.
+type topK struct {
+	k    int
+	heap []candidate
+}
+
+func newTopK(k int) *topK {
+	return &topK{k: k, heap: make([]candidate, 0, k)}
+}
+
+// offer admits a row if it beats the weakest survivor, or if the heap is not
+// full yet. id and meta are the driver's buffers and are copied here, never
+// retained.
+//
+// A row that only ties the weakest survivor does not displace it, so among
+// equal scores the earliest k rows are kept. The full sort this replaced used
+// sort.Slice, which is not stable, so it broke such a tie arbitrarily. That is
+// the ONE observable difference between the two, and it is deliberate:
+// deterministic beats arbitrary.
+//
+// It is not academic. Under the real embedder an exact tie means duplicate
+// vectors and nothing in the recall eval moved (V-643). Under the hash
+// embedder the eval's deterministic floor uses, ties are everywhere — it is
+// bag-of-words, so every note sharing no word with the query scores exactly 0
+// — and recall@3 on that run moved 74.1% to 81.5% purely because the zeros now
+// come out in a fixed order. Neither number measures retrieval. recall@1 and
+// false recall, which the eval actually asserts, are unchanged on both runs.
+func (t *topK) offer(score float64, id, meta []byte) {
+	if t.k == 0 {
+		return
+	}
+	if len(t.heap) < t.k {
+		t.heap = append(t.heap, candidate{score: score, id: string(id), meta: append([]byte(nil), meta...)})
+		t.up(len(t.heap) - 1)
+		return
+	}
+	if score <= t.heap[0].score {
+		return
+	}
+	t.heap[0] = candidate{score: score, id: string(id), meta: append([]byte(nil), meta...)}
+	t.down(0)
+}
+
+func (t *topK) up(i int) {
+	for i > 0 {
+		parent := (i - 1) / 2
+		if t.heap[parent].score <= t.heap[i].score {
+			return
+		}
+		t.heap[parent], t.heap[i] = t.heap[i], t.heap[parent]
+		i = parent
+	}
+}
+
+func (t *topK) down(i int) {
+	for {
+		l, r, small := 2*i+1, 2*i+2, i
+		if l < len(t.heap) && t.heap[l].score < t.heap[small].score {
+			small = l
+		}
+		if r < len(t.heap) && t.heap[r].score < t.heap[small].score {
+			small = r
+		}
+		if small == i {
+			return
+		}
+		t.heap[small], t.heap[i] = t.heap[i], t.heap[small]
+		i = small
+	}
+}
+
+// sorted drains the heap into descending score order — what Search returns.
+func (t *topK) sorted() []candidate {
+	out := t.heap
+	sort.Slice(out, func(i, j int) bool { return out[i].score > out[j].score })
+	return out
+}
+
 // ByPrefix returns every row whose id starts with prefix, vectors included.
 //
 // This is not a similarity query and deliberately does not score anything:
@@ -240,6 +344,26 @@ func decodeVec(b []byte) []float32 {
 	return v
 }
 
+// dotBlob is dot against a vector still in its stored encoding, so scoring a
+// row the query is about to discard does not allocate the []float32 that
+// decodeVec would build. Same arithmetic, same order of operations, so it
+// returns bit-identical scores to dot(a, decodeVec(b)).
+//
+// A blob whose length isn't a multiple of 4 is truncated to the whole-element
+// prefix, matching decodeVec, and a length mismatch is 0, matching dot.
+func dotBlob(a []float32, b []byte) float64 {
+	n := len(b) / 4
+	if len(a) != n || n == 0 {
+		return 0
+	}
+	var sum float64
+	for i := 0; i < n; i++ {
+		f := math.Float32frombits(binary.LittleEndian.Uint32(b[4*i:]))
+		sum += float64(a[i]) * float64(f)
+	}
+	return sum
+}
+
 // dot is the cosine similarity for L2-normalized vectors (mismatched lengths ⇒
 // 0, matching internal/memory's cosine).
 func dot(a, b []float32) float64 {

internal/store/memory_bench_test.go

@@ -0,0 +1,77 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"math"
+	"math/rand"
+	"path/filepath"
+	"testing"
+)
+
+// benchDim is the resident embedder's width (multilingual-e5-small, 384), so
+// the per-row decode cost the benchmark measures is the real one.
+const benchDim = 384
+
+// seedMemVectors fills a fresh store with n L2-normalized rows carrying a meta
+// blob the size recall actually stores — the note text plus its type — because
+// the cost this benchmark exists to measure is unmarshalling that blob for
+// every row when only topK survivors need it.
+func seedMemVectors(tb testing.TB, n int) *MemoryStore {
+	tb.Helper()
+	path := filepath.Join(tb.TempDir(), "mem_bench.db")
+	st, err := Open(context.Background(), path)
+	if err != nil {
+		tb.Fatalf("Open: %v", err)
+	}
+	tb.Cleanup(func() { _ = st.Close() })
+	m := st.VectorMemory()
+
+	rng := rand.New(rand.NewSource(1))
+	ctx := context.Background()
+	for i := 0; i < n; i++ {
+		if err := m.Insert(ctx, fmt.Sprintf("note:%d", i), randUnitVec(rng, benchDim), map[string]string{
+			"type": "note",
+			"text": fmt.Sprintf("заметка номер %d о том, что надо не забыть сделать на неделе", i),
+		}); err != nil {
+			tb.Fatalf("Insert %d: %v", i, err)
+		}
+	}
+	return m
+}
+
+func randUnitVec(rng *rand.Rand, dim int) []float32 {
+	v := make([]float32, dim)
+	var norm float64
+	for i := range v {
+		f := rng.NormFloat64()
+		v[i] = float32(f)
+		norm += f * f
+	}
+	norm = math.Sqrt(norm)
+	for i := range v {
+		v[i] = float32(float64(v[i]) / norm)
+	}
+	return v
+}
+
+// BenchmarkMemoryStoreSearch measures one recall query against a store of n
+// rows. Row counts bracket the documented scale: 1000 is a plausible today,
+// 10000 is the "thousands, not millions" ceiling the type doc claims a full
+// scan is fine at.
+func BenchmarkMemoryStoreSearch(b *testing.B) {
+	for _, n := range []int{1000, 10000} {
+		b.Run(fmt.Sprintf("rows=%d", n), func(b *testing.B) {
+			m := seedMemVectors(b, n)
+			q := randUnitVec(rand.New(rand.NewSource(2)), benchDim)
+			ctx := context.Background()
+			b.ReportAllocs()
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				if _, err := m.Search(ctx, q, 10); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}

internal/store/memory_search_test.go

@@ -0,0 +1,107 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"math/rand"
+	"sort"
+	"testing"
+
+	"github.com/kami/maven/internal/memory"
+)
+
+// naiveSearch is the implementation Search replaced: score every row into a
+// slice, sort the whole slice, truncate. It stays in the test file as the
+// reference the bounded-heap version is judged against, because "recall must
+// not change" is a claim about output, not about the code that produces it.
+func naiveSearch(t *testing.T, m *MemoryStore, vec []float32, topK int) []memory.Result {
+	t.Helper()
+	rows, err := m.db.QueryContext(context.Background(),
+		`SELECT id, vec FROM memory_vectors WHERE id NOT LIKE ? ESCAPE '\'`,
+		escapeLike(memory.NonRecallPrefix)+"%")
+	if err != nil {
+		t.Fatalf("naive scan: %v", err)
+	}
+	defer rows.Close()
+	var out []memory.Result
+	for rows.Next() {
+		var id string
+		var blob []byte
+		if err := rows.Scan(&id, &blob); err != nil {
+			t.Fatalf("naive row: %v", err)
+		}
+		out = append(out, memory.Result{ID: id, Score: dot(vec, decodeVec(blob))})
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("naive rows: %v", err)
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Score > out[j].Score })
+	if topK < len(out) {
+		out = out[:topK]
+	}
+	return out
+}
+
+// TestMemoryStoreSearchMatchesNaive is the constraint on V-643: the bounded
+// heap must return exactly what a full scan and sort returned. Distinct random
+// vectors, so no two scores tie and the ranking is total — a mismatch here is
+// arithmetic or heap logic, not a tie-break difference.
+func TestMemoryStoreSearchMatchesNaive(t *testing.T) {
+	ctx := context.Background()
+	m := newMemTestStore(t).VectorMemory()
+
+	rng := rand.New(rand.NewSource(7))
+	const rows, dim = 500, 64
+	for i := 0; i < rows; i++ {
+		if err := m.Insert(ctx, fmt.Sprintf("n%d", i), randUnitVec(rng, dim), map[string]string{
+			"text": fmt.Sprintf("note %d", i),
+		}); err != nil {
+			t.Fatalf("Insert %d: %v", i, err)
+		}
+	}
+
+	for _, topK := range []int{1, 3, 10, 50, rows, rows + 100} {
+		q := randUnitVec(rng, dim)
+		got, err := m.Search(ctx, q, topK)
+		if err != nil {
+			t.Fatalf("Search topK=%d: %v", topK, err)
+		}
+		want := naiveSearch(t, m, q, topK)
+		if len(got) != len(want) {
+			t.Fatalf("topK=%d: got %d results, naive returned %d", topK, len(got), len(want))
+		}
+		for i := range want {
+			if got[i].ID != want[i].ID {
+				t.Errorf("topK=%d rank %d: got %q, naive says %q", topK, i, got[i].ID, want[i].ID)
+			}
+			if got[i].Score != want[i].Score {
+				t.Errorf("topK=%d rank %d (%s): score %v, naive says %v",
+					topK, i, got[i].ID, got[i].Score, want[i].Score)
+			}
+		}
+		if len(got) > 0 && got[0].Meta["text"] == "" {
+			t.Errorf("topK=%d: survivor %s has no meta — it was never unmarshalled", topK, got[0].ID)
+		}
+	}
+}
+
+// TestDotBlobMatchesDot pins the claim in dotBlob's doc comment: reading the
+// vector out of its stored bytes is bit-identical to decoding it first. Scores
+// feed a gate with a 0.008 margin, so "close enough" is not the bar.
+func TestDotBlobMatchesDot(t *testing.T) {
+	rng := rand.New(rand.NewSource(11))
+	for i := 0; i < 200; i++ {
+		a := randUnitVec(rng, 384)
+		b := randUnitVec(rng, 384)
+		if got, want := dotBlob(a, encodeVec(b)), dot(a, b); got != want {
+			t.Fatalf("dotBlob = %v, dot = %v", got, want)
+		}
+	}
+	// Length mismatch is 0 in both, and so is an empty vector.
+	if got := dotBlob([]float32{1, 0}, encodeVec([]float32{1, 0, 0})); got != 0 {
+		t.Errorf("mismatched lengths scored %v, want 0", got)
+	}
+	if got := dotBlob(nil, nil); got != 0 {
+		t.Errorf("empty scored %v, want 0", got)
+	}
+}

internal/store/store.go

@@ -95,7 +95,20 @@ func openAt(ctx context.Context, path string) (*sql.DB, error) {
 	if err != nil {
 		return nil, fmt.Errorf("open %s: %w", path, err)
 	}
-	// single writer expected; the daemon is the only process touching the db.
+	// One connection, so every statement is serialised at the database and no
+	// caller above needs a lock of its own. internal/ipc's Server relies on
+	// exactly this, which is why the cap is an invariant rather than a tuning
+	// knob: raising it moves the serialisation guarantee somewhere it is not
+	// written down.
+	//
+	// Measured on 07-08-2026 (V-642, docs/evals/2026-08-07-store-connection-cap.md).
+	// WAL exists to let readers run beside one writer, and the cap gives that
+	// up, but reads do not queue: p50 594µs against 525µs at a cap of four,
+	// while write throughput more than halves. The one thing the cap cannot
+	// survive is a long-lived transaction, which holds the only connection and
+	// stalls every read for its lifetime. So the store begins none, and
+	// TestConnCap_ReadBlocksBehindOpenSnapshot is the standing measurement of
+	// what re-adding one would cost.
 	db.SetMaxOpenConns(1)
 	if _, err := db.ExecContext(ctx, schemaSQL); err != nil {
 		if closeErr := db.Close(); closeErr != nil {
@@ -133,13 +146,6 @@ func (s *Store) Close() error {
 	return s.enc.closeAndSeal(s.db)
 }
 
-// DB exposes the underlying handle for internal read-only snapshots.
-// Used by the loop to take a consistent read under a single transaction.
-// Modules never receive this handle — core mediates.
-func (s *Store) DB(ctx context.Context) (*sql.Tx, error) {
-	return s.db.BeginTx(ctx, &sql.TxOptions{ReadOnly: true})
-}
-
 var (
 	// ErrNoFact — no non-voided row exists for this key.
 	ErrNoFact = errors.New("store: no fact for key")

internal/stt/http.go

@@ -0,0 +1,92 @@
+package stt
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"strconv"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// HTTPTranscriber — speech-to-text on another host, over HTTP.
+//
+// mavsttd is whisper.cpp linked into a Go daemon and reached over a unix
+// socket. CrisperWhisper 2.0 cannot be reached that way: whisper.cpp derives
+// its language count from the vocabulary size, and CW2's 51897 tokens shift
+// seven special token ids. It runs under transformers instead, as a service
+// beside the model on workpc. See docs/evals/2026-08-09-crisperwhisper2-russian-wer.md.
+//
+// So this is the second transport for the same seam, not a second seam. The
+// caller still sees stt.Transcriber and one method.
+type HTTPTranscriber struct {
+	url   string
+	token string
+	lang  string
+	http  *http.Client
+}
+
+// NewHTTPTranscriber builds the remote client. token may be empty for a
+// service on a trusted socket, but audio is the most sensitive thing that
+// crosses this seam, so a LAN deployment should always set one.
+func NewHTTPTranscriber(url, token, lang string, timeout time.Duration) *HTTPTranscriber {
+	return &HTTPTranscriber{
+		url:   url,
+		token: token,
+		lang:  lang,
+		http:  &http.Client{Timeout: timeout},
+	}
+}
+
+// ErrFormat — the audio is not the one canonical shape. Refused at the seam
+// rather than sent to a model that expects something else.
+var ErrFormat = errors.New("stt: audio is not 16kHz mono pcm_s16le")
+
+type httpTranscript struct {
+	Text       string  `json:"text"`
+	Confidence float64 `json:"confidence"`
+}
+
+// Transcribe posts the raw PCM and reads back the text.
+//
+// The body is the PCM bytes themselves rather than JSON. A minute of 16kHz
+// mono is under 2MB raw and about 2.6MB base64, and the format is fixed by
+// audio.PCM16kMono, so a header carries it more cheaply than an envelope.
+func (t *HTTPTranscriber) Transcribe(ctx context.Context, a audio.Audio) (string, float64, error) {
+	if !a.Format.IsValid() {
+		return "", 0, ErrFormat
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, t.url, bytes.NewReader(a.Bytes))
+	if err != nil {
+		return "", 0, fmt.Errorf("stt: build request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/octet-stream")
+	req.Header.Set("X-Sample-Rate", strconv.Itoa(a.Format.SampleRate))
+	req.Header.Set("X-Channels", strconv.Itoa(a.Format.Channels))
+	req.Header.Set("X-Sample-Bits", strconv.Itoa(a.Format.SampleBits))
+	req.Header.Set("X-Language", t.lang)
+	if t.token != "" {
+		req.Header.Set("Authorization", "Bearer "+t.token)
+	}
+
+	resp, err := t.http.Do(req)
+	if err != nil {
+		return "", 0, fmt.Errorf("stt: post audio: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", 0, fmt.Errorf("stt: remote returned %d", resp.StatusCode)
+	}
+
+	var out httpTranscript
+	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
+		return "", 0, fmt.Errorf("stt: decode transcript: %w", err)
+	}
+	return out.Text, out.Confidence, nil
+}
+
+var _ Transcriber = (*HTTPTranscriber)(nil)

internal/stt/http_test.go

@@ -0,0 +1,89 @@
+package stt
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strconv"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+func TestHTTPTranscriberSendsRawPCM(t *testing.T) {
+	t.Parallel()
+	var gotBody []byte
+	var gotHeader http.Header
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotBody, _ = io.ReadAll(r.Body)
+		gotHeader = r.Header.Clone()
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = io.WriteString(w, `{"text":"привет","confidence":0.82}`)
+	}))
+	defer srv.Close()
+
+	a := audio.Audio{Format: audio.PCM16kMono, Bytes: []byte("pcm-bytes")}
+	tr := NewHTTPTranscriber(srv.URL, "s3cret", "ru", 2*time.Second)
+	text, conf, err := tr.Transcribe(context.Background(), a)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if text != "привет" || conf != 0.82 {
+		t.Fatalf("got %q %v", text, conf)
+	}
+	if string(gotBody) != "pcm-bytes" {
+		t.Fatalf("body should be the PCM itself, got %q", gotBody)
+	}
+	if got := gotHeader.Get("X-Sample-Rate"); got != strconv.Itoa(audio.PCM16kMono.SampleRate) {
+		t.Fatalf("X-Sample-Rate = %q", got)
+	}
+	if got := gotHeader.Get("X-Language"); got != "ru" {
+		t.Fatalf("X-Language = %q", got)
+	}
+	// Audio is the most sensitive thing crossing this seam.
+	if got := gotHeader.Get("Authorization"); got != "Bearer s3cret" {
+		t.Fatalf("Authorization = %q", got)
+	}
+}
+
+func TestHTTPTranscriberOmitsEmptyToken(t *testing.T) {
+	t.Parallel()
+	var auth string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		auth = r.Header.Get("Authorization")
+		_, _ = io.WriteString(w, `{"text":"x"}`)
+	}))
+	defer srv.Close()
+	a := audio.Audio{Format: audio.PCM16kMono, Bytes: []byte("x")}
+	if _, _, err := NewHTTPTranscriber(srv.URL, "", "ru", time.Second).Transcribe(context.Background(), a); err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if auth != "" {
+		t.Fatalf("Authorization should be absent, got %q", auth)
+	}
+}
+
+func TestHTTPTranscriberRefusesWrongFormat(t *testing.T) {
+	t.Parallel()
+	a := audio.Audio{Format: audio.Format{SampleRate: 44100, Channels: 2, SampleBits: 16, Encoding: "pcm_s16le"}}
+	_, _, err := NewHTTPTranscriber("http://example.invalid", "", "ru", time.Second).Transcribe(context.Background(), a)
+	if !errors.Is(err, ErrFormat) {
+		t.Fatalf("want ErrFormat, got %v", err)
+	}
+}
+
+func TestHTTPTranscriberErrorsOnBadStatus(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusUnauthorized)
+	}))
+	defer srv.Close()
+	a := audio.Audio{Format: audio.PCM16kMono, Bytes: []byte("x")}
+	_, _, err := NewHTTPTranscriber(srv.URL, "", "ru", time.Second).Transcribe(context.Background(), a)
+	if err == nil {
+		t.Fatal("a 401 must be an error, so the Pair falls back")
+	}
+}