Add a Kiwix search client and score retrieval (Vikunja #403)

Step one of letting Maven read instead of recall. No LLM yet.

internal/kiwix/client.go: search a local Kiwix server, parse the RSS
reply, hand back title + path + plain-text snippet + word count. The
snippet is the unit of context; a full article is ~100KB of HTML and
will not fit a 4096 token window.

internal/kiwix/retrieval_eval.go plus knowledge_v1.json: the 9 knowledge
questions from the phrasing fixture, each with hand-written English
keywords, scored on whether a wanted article comes back in the top 5.
Opt-in via MAVEN_KIWIX_URL, since CI has no Kiwix. No pass bar, the
number is the finding.

Result on the live mirror: 8/8 answerable questions hit, 7 of them at
rank 1. Retrieval works. Keywords are written by hand on purpose, since
Kiwix ranks by keyword and not by meaning, so a natural question fails.
A query-rewrite step is the next piece of work.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

CLAUDE.md

@@ -82,9 +82,23 @@ workspace enforces that the Go and relabelling prompts remain identical.
 
 ## Non-goals (hard constraints)
 
-Never phones home. Not a nag, not autonomous. Maven's persona is **feminine** — Russian
+Not a nag, not autonomous. Maven's persona is **feminine** — Russian
 self-reference must use feminine forms (the user is male; see memory `maven-persona-gender`).
 
+**"Never phones home" is DEPRECATED** (owner's call, 2026-07-31). It used to be a hard
+constraint and it is not one any more: a 0.8B — and a 1.7B — does not know enough to answer
+world questions, so she needs to read external sources. What replaces it:
+
+- **No telemetry, no cloud model, no third-party account.** That part never changes. Nothing
+  about Maven is reported to anyone, and inference stays on the box.
+- **Local sources first.** Kiwix ZIMs on homesrv (Wikipedia, ifixit) before anything on the
+  network. Reading beats recalling for a small model, and a local read costs nothing.
+- **External search is allowed and off unless configured**, like the weather and telegram
+  capabilities.
+- **His notes and facts are never search input.** Looking up why the sky is blue and sending
+  his stored personal notes to an upstream engine are different acts. Only the utterance goes
+  out, never the persona block, history, or matched notes.
+
 ## Web UI conventions
 
 Server-rendered pages share `cmd/mavweb/static/ui.css` (served at `/ui.css`) and the `nav`

DESIGN.md

@@ -15,7 +15,8 @@
 
 **Maven** — self-hosted personal assistant. Manages your day, acts on your
 homelab. One daemon on homesrv (always-on, not the workstation), multiple
-client surfaces. All local, never phones home.
+client surfaces. Inference and data stay on the box; she may READ external
+sources (see Non-goals — "never phones home" is deprecated).
 
 Primary name is "Maven", with feminine-gendered Russian self-reference
 ("она", "меня", "помогла"). Clients may choose their own UI label. Consistent
@@ -35,8 +36,13 @@ Inside boundary — the ones that actually constrain the build:
   she records. A confident wrong fact is worse than a known gap.
 - **Not a nag** — she'd rather miss a nudge than be mutable. Shuts up when
   uncertain. Load-bearing.
-- **Not a stranger** — runs on your stuff, your model, your data. Never
-  phones home.
+- **Not a stranger** — runs on your stuff, your model, your data. No
+  telemetry, no cloud model, no third-party account. She may READ external
+  sources to answer world questions (Kiwix first, then optional search); she
+  never reports anything about you to anyone, and your notes and facts are
+  never used as search input. **"Never phones home" as an absolute is
+  deprecated** — owner's call, 2026-07-31: a small model does not know enough
+  to be useful without reading.
 - **Not a relationship** — mom-tone is a function that makes nudges land, not
   emotional company. Names the drift a warm small model falls into.
 
@@ -458,7 +464,7 @@ decides *insistence*. Both are needed.
 
 sev ≤ 2 drops on away, sev ≥ 3 holds: a missed water nudge is noise, a missed
 backup failure isn't. Away-channels (ntfy/telegram) leave the box — the one
-path that crosses "never phones home," through your own relay. **Minimal
+path that leaves the box for a person to see, through your own relay. **Minimal
 body** — "disk low on homesrv," not detail; don't make notifications a
 shoulder-surf exfil surface.
 

MODEL-BAKEOFF-31-07-2026.md

@@ -0,0 +1,101 @@
+# Resident model bake-off — 31-07-2026
+
+**Recommendation: keep Qwen3.5-0.8B.** LFM2.5-1.2B is worse at routing (52.6% vs 60.5%
+intent accuracy), and the loss is almost entirely Russian (18/61 vs 22/61 RU, while EN is a
+wash). It is also 2.4× slower. The Thinking variant is far worse again.
+
+Settles Vikunja **#278 / #250**.
+
+- Same fixture and scorer as `ROUTING-EVAL-31-07-2026.md`: `internal/router/eval/`
+  (`ru_routing_v1.json`, 76 held-out cases).
+- Reproduce: `MAVEN_LLM_URL=http://127.0.0.1:<port> make eval-router`
+  (`TestLLMRouterBaseline`). Note: there is no `make eval-models` target.
+- All three models served by the same `llama-server` flags — `-c 2048 -ngl 99 -t 6`, only
+  `-m` and `--port` differ. One server at a time on an otherwise idle box, so latencies are
+  real and not contention.
+- Measured on top of the router prompt fix (`origin/overnight/router-prompt` merged in), so
+  the Qwen column is directly comparable to the numbers already recorded.
+
+## Results
+
+`llm-only` — the model alone. This is the column that measures the model.
+
+| | Qwen3.5-0.8B | LFM2.5-1.2B Instruct | LFM2.5-1.2B Thinking |
+|---|---|---|---|
+| **intent-only accuracy** | **60.5%** | 52.6% | 36.8% |
+| full accuracy (intent+slots+gate) | **36.8%** | 32.9% | 21.1% |
+| **RU** | **22/61** | 18/61 | 10/61 |
+| EN | 6/15 | **7/15** | 6/15 |
+| route errors | 0 | 0 | 0 |
+| **p50 / p95 latency** | **1.05s / 1.71s** | 2.47s / 3.62s | 2.42s / 3.24s |
+| missed clarify | 6 / 6 | 6 / 6 | 6 / 6 |
+
+`cascade+llm` — stage-0 → model → classifier floor, what #320 would actually ship. Same
+ordering.
+
+| | Qwen3.5-0.8B | LFM2.5-1.2B Instruct | LFM2.5-1.2B Thinking |
+|---|---|---|---|
+| intent-only accuracy | **61.8%** | 55.3% | 38.2% |
+| full accuracy | **46.1%** | 42.1% | 30.3% |
+| RU / EN | **27/61** / 8/15 | 23/61 / **9/15** | 15/61 / 8/15 |
+| route errors | 0 | 0 | 0 |
+| p50 / p95 latency | **1.28s / 1.94s** | 2.18s / 2.72s | 2.27s / 3.19s |
+
+Full logs: the three runs are archived in the session scratchpad
+(`qwen08.txt`, `lfm-instruct.txt`, `lfm-thinking.txt`).
+
+## Russian-specific failures — the owner's worry is confirmed
+
+LFM2.5's Russian loss is not spread out. It has one large, specific failure: **it hears
+almost any Russian imperative or short phrase as `reminder`.**
+
+- `перезапусти докер` → reminder (want act)
+- `включи вытяжку` → reminder (want act)
+- `закрой жалюзи` → reminder (want act)
+- `заметка: продлить домен в августе` → reminder (want note)
+- `запиши что кран на кухне снова капает` → reminder (want note)
+- `доброе утро` → reminder (want chat)
+- `спасибо тебе` → reminder (want note/chat)
+- `переходи в тихий режим` → reminder (want system)
+
+That is `note→reminder ×4`, `act→reminder ×4`, `chat→reminder ×2` in one run. Qwen's
+equivalent failure axis is `query→fact ×8`, which is a narrower and already-understood bug.
+
+Two more Russian-side problems worth naming:
+
+1. **Fact keys come back empty or wrong in Russian.** `воды попил наконец`, `поужинал`,
+   `поспал часов пять` and `отметь что я позавтракал овсянкой` all returned an empty key.
+   `сходил в душ` and `отдохнул минут двадцать` both returned `water`. Qwen does not do this.
+2. **It leaked German.** `slept about seven hours` produced the fact key
+   `"7 Stunden geschlafen"`. Grammar-valid, semantically garbage — a sign the multilingual
+   mix is not anchored where Maven needs it.
+
+The claimed tool-calling advantage did not show up here. `act` is the closest thing this
+fixture has to a tool call, and LFM2.5 got it wrong more often than Qwen, mostly by calling
+it a reminder. It also produced no `fn` slot on any act, same as Qwen.
+
+## The Thinking variant
+
+Not viable. 36.8% intent accuracy, 10/61 Russian, and no latency saving over Instruct — the
+thinking trace costs time without buying accuracy on a short enum classification. With the
+`enable_thinking=false` diagnostic it collapsed further to 28.9% with 2 route errors
+(`query→reminder ×12`). Do not pursue.
+
+## Notes
+
+- Nothing crashed, nothing ignored the GBNF grammar, and no model produced unparseable JSON
+  in the shippable configurations. Zero route errors for both Instruct and Thinking in
+  `llm-only` and `cascade+llm`. The problem with LFM2.5 is what it decides, not whether it
+  can emit the contract.
+- The `6 / 6` missed clarify is unchanged across all three models. No model fixes the missing
+  refusal lane — that is `Confidence: 1.0` hardcoded in `llmrouter.go` (Vikunja #359), not a
+  model property.
+- The report labels every configuration `(0.8B)`; that string is hardcoded in the test, not a
+  reflection of which gguf was loaded. Model identity was confirmed per run via `/v1/models`.
+- No Go code was changed for this measurement, and no bug was found that needed one.
+
+## What this does not settle
+
+Routing only. LFM2.5 might still phrase better, and phrasing is the resident model's other
+job — that needs its own fixture. But routing is the load-bearing path and Maven is
+Russian-first, so on the evidence here the switch is not worth making.

Makefile

@@ -103,14 +103,17 @@ eval-router:
 eval-recall:
 	MAVEN_ONNX_LIB="$(MAVEN_ONNX_LIB)" $(GO) test -v -count=1 ./internal/memory/recalleval/
 
-# eval-phrasing -- score nudge phrasing (internal/phraser/eval). Verbose so the
+# eval-phrasing -- score nudge phrasing AND the conversational paths (chat,
+# query, general knowledge) in internal/phraser/eval. Verbose so the
 # report and every generated message land in the terminal. With no environment
 # it scores the deterministic Stub only, which is what CI runs. Set
 # MAVEN_LLM_URL to add the resident model:
 #   MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-phrasing
-# The model run is slow (minutes) -- the timeout is raised to match.
+# The model run is slow (minutes) -- the timeout is raised to match. It covers
+# two fixtures now (15 nudges + 27 conversational cases, and the chat replies are
+# the long ones), hence 90m rather than 40m.
 eval-phrasing:
-	$(GO) test -v -count=1 -timeout 40m ./internal/phraser/eval/
+	$(GO) test -v -count=1 -timeout 90m ./internal/phraser/eval/
 
 # eval-models — score ONE llama-server against the same fixture, for the
 # resident-model bake-off (#278, #250). Start a server with the gguf you want,

PHRASING-EVAL-31-07-2026.md

@@ -0,0 +1,169 @@
+# Phrasing evaluation — 31-07-2026
+
+How Maven words a nudge, measured instead of argued. Counterpart to
+`ROUTING-EVAL-31-07-2026.md`.
+
+- Fixture + scorer: `internal/phraser/eval/` (`nudges_v1.json`, 15 cases; `eval.go`, `checks.go`)
+- Reproduce: `MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-phrasing`
+- Model: Qwen3.5-0.8B Q4_K_M, the resident model. Not swapped.
+- Commit: `a40bc55` (prompt fix)
+
+Every check is a string or length test a human can read and disagree with. No model
+grades another model here.
+
+## Result
+
+| | before | after |
+|---|---|---|
+| **cases passing every check** | **0/15** | **13/15** |
+| mood in enum | 6/15 | 15/15 |
+| Russian | 2/15 | 14/15 |
+| length (≤120 chars, ≤16 words) | 13/15 | 15/15 |
+| feminine self-reference | 15/15 | 15/15 |
+| no cringe | 13/15 | 15/15 |
+| on topic | 6/15 | 13/15 |
+| p50 latency | 11.4s | 11.4s |
+
+Latency did not move and is not good. 11s to word one nudge on this box.
+
+## The bug reproduced
+
+Yes, exactly as reported. 7 of 15 messages were the literal string `"..."`, and one was
+`"full voice message"`. Both are text copied straight out of the prompt.
+
+The system prompt said:
+
+```
+Respond ONLY with valid JSON: {"response": "full voice message", "mood": "neutral"}
+```
+
+and the user prompt said:
+
+```
+Respond as JSON: {"response": "...", "mood": "..."}
+```
+
+A 0.8B does not read `"..."` as "put your answer here". It reads it as the answer. The
+prompt was a worked example whose worked part was blank, so the model filled the slot by
+copying. This is the whole of finding 1.
+
+## What else was wrong
+
+Four separate faults, all prompt-side:
+
+1. **Placeholder echo** (7 cases) — above.
+2. **Wrong language** (13/15 failed the language check). The prompt was entirely English
+   and said "in the user's language (Russian or English)". The model picked English. It is
+   never English: the nudge is spoken by a Russian piper voice.
+3. **Rule names are English identifiers.** `netdata_critical`, `service_down`, `break` went
+   into the prompt raw. The model cannot nudge about a topic it has not been told in words,
+   so 9/15 were off topic. The daemon knows what its own rules mean; now it says so.
+4. **Mood invented** (`"warm"`, twice). The enum was listed in a parenthesis at the end of
+   an English sentence. Now it is its own line: "ровно одно из: neutral, happy, thinking,
+   tired, confused."
+
+Plus two non-prompt faults the run exposed:
+
+- **The no-parse fallback was English.** When the model returned nothing usable, the body
+  became `fmt.Sprintf("%s — %s", rule, sev)` — `"water — care"` — and that string went to
+  a Russian TTS. Now it falls back to plain Russian.
+- **Durations were English.** `humanDur` returns "3 hours"; it was landing verbatim inside
+  Russian sentences. Nudges now use a Russian formatter.
+
+## Three iterations, and what each taught
+
+| | score | change |
+|---|---|---|
+| baseline | 0/15 | — |
+| iter 1 | 2/15 | Russian prompt, filled-in examples, Russian durations |
+| iter 2 | 11/15 | required keyword per rule, one example instead of five, Russian fallback |
+| iter 3 | **13/15** | examples moved to topics that are not rules |
+
+The interesting step is 1 → 2. Fixing the placeholder did not fix the disease, it moved it:
+the model stopped copying `"..."` and started copying my first example instead. Five nudges
+in a row came back as `"Ты не пил воду три часа. Налей стакан."` regardless of the rule.
+
+**A small model copies the nearest concrete text in its prompt.** That is one failure mode
+with two symptoms. The fix that stuck was making the examples about laundry and a laptop
+battery — topics no rule ever produces, so copying them is visible in the score rather than
+invisibly passing the water cases.
+
+## Do not oversell 13/15
+
+Seven of the thirteen passes are the **deterministic fallback**, not the model:
+`"Напоминаю: таблетки."`, `"Сервис не отвечает."`, `"Критический алярм: проверь диск."`,
+`"Ты давно не пил воду."`. Those are strings this commit added to Go. The model returned
+nothing parseable and the fallback scored.
+
+So the honest reading is roughly **6/15 from the model, 7/15 from a fallback, 2/15 failing**.
+The prompt fix is real — `"..."` is nearly gone and the language and mood checks are clean —
+but a large part of the jump is that failure now degrades into Russian instead of into
+`"water — care"`. That is a genuine improvement for the operator and a weak one for the model.
+
+The two remaining failures: one `"..."` recurrence (`routine-stretch`) and one meal nudge
+that never says food.
+
+## Tried and reverted: an example-led nudge prompt (#393)
+
+The idea was that a 0.8B copies examples better than it follows rules, so the nudge prompt
+was rewritten to lead with five on-topic examples (water, break, pills, morning, service) and
+the prose rules were compressed to pay for the tokens: 1190 chars down to 986.
+
+It measured **worse**, three runs each side, same llama-server, same fixture:
+
+| run | before | after |
+|---|---|---|
+| 1 | 12/15 (address 14) | 11/15 (address 13) |
+| 2 | 13/15 (address 15) | 12/15 (address 15) |
+| 3 | 14/15 (address 15) | 11/15 (address 12) |
+
+`feminine` and `hisgender` were 15/15 on all six runs, so they measure nothing here. The
+regression is all in `address`: 44/45 before, 40/45 after. Formal "вы"/"ваше" and plural
+imperatives came back, and so did `"..."`.
+
+Two likely causes, both about the same thing — **examples do not carry a prohibition**. The
+old prompt spent a whole sentence on «говоришь на "ты", в единственном числе»; the new one
+demoted that to one item in a long "никогда" list, and the model stopped obeying it. And
+making the examples on-topic let their *wording* leak: a break case came back as
+«Вы давно не пили воду. Выпей стакан.» — the water example, verbatim, in the wrong slot.
+That is exactly the failure the laundry/laptop examples were chosen to avoid.
+
+Change reverted. What survives is the measurement: a rule the model must obey needs its own
+sentence, and examples must stay off-topic. Also note the before side alone spans 12–14 of
+15 — this fixture cannot resolve anything smaller than about three cases.
+
+## Broken, found, not fixed
+
+1. ~~**`checkFeminine` only catches half the constraint.**~~ **Fixed** (#381). It scanned for
+   masculine self-reference only, so three messages that addressed the *owner* in the feminine
+   ("ты давно не отдыхал**а**") scored clean. There is now a second check, `hisgender`: a
+   feminine past-tense verb (-ла/-лась) in a sentence addressed to him ("ты", "тебе", "твой")
+   fails, unless the verb is hers ("я заметила", "напомнила тебе"). It is a suffix rule, not a
+   parser — see the comment in `checks.go` for what it misses. A fresh 15-case run after adding
+   it scored **12/15** with `hisgender` 15/15; the model did not repeat the feminine address in
+   that sample, and the check is pinned by unit tests on the recorded bad strings instead.
+2. **Grammar is not checked at all, and it is bad.** `"Он не ел 11 дней"` (it was 11 hours),
+   `"Сонуждились 7 дней"` (not a word), `"Они забыли воду"` (wrong person entirely). Every
+   one of these passes all six checks. The fixture measures properties, not fluency, and at
+   0.8B fluency is the binding constraint.
+3. **Unit confusion.** The model turns hours into days about a third of the time. The
+   prompt now says "11 ч"; it reads it as days.
+4. **11s p50.** Unchanged and untouched here. A nudge the model takes eleven seconds to
+   word has missed its moment. Worth its own task.
+5. **The keyword hint is close to teaching to the test.** `ruleKeywords` names the word the
+   on-topic check looks for. It is defensible — the daemon genuinely knows its rule topics
+   and the model genuinely cannot infer them from `netdata_critical` — but the on-topic
+   number is softer than the others because of it.
+
+## Next steps
+
+1. ~~**Add a second-person gender check**~~ — done, `hisgender` in `checks.go` (#381).
+2. **Decide whether the fallback should count as a pass.** Right now `Score` cannot tell a
+   model answer from a fallback. Either mark fallback bodies in `PhrasedNudge` or count them
+   in their own column. Without that, any future prompt change can score well by failing
+   more.
+3. **Attack the 11s.** Nudge phrasing is short and non-interactive; thinking off is the first
+   thing to try, as it was for routing (#376).
+4. **Re-measure when #122 lands.** The CPT'd Qwen3-1.7B is the target. 13/15 with seven
+   fallbacks is the floor it has to beat, and the fluency problems above are the ones a
+   bigger, Russian-trained checkpoint should actually fix.

REARCH.md

@@ -90,4 +90,7 @@ later* is the worker + RAG.
 4. **Deferred work** — larger reasoner, custom Piper voice and other expansions.
 
 ## Non-goals (unchanged)
-Never phones home. Not a nag. Not autonomous. Feminine-gendered RU self-ref.
+Not a nag. Not autonomous. Feminine-gendered RU self-ref. No telemetry, no
+cloud model, no third-party account — but she MAY read external sources to
+answer world questions (Kiwix first, search optional). "Never phones home" as
+an absolute is deprecated, owner's call 2026-07-31; see CLAUDE.md § Non-goals.

RECALL-EVAL-31-07-2026.md

@@ -75,6 +75,14 @@ same vector. A note is indexed in both places with the same embedding, so if it
 `QueryNotes` it fails again here — the branch can only ever return a **fact**. Its comment calls it
 "additive"; for notes it is not.
 
+**Fixed (Vikunja #373).** The memory pass now runs *first*, as one search over notes and facts with
+one gate, so whichever memory is clearly the best match answers — note or fact. The notes-only pass
+stays behind it for notes the vector index does not hold. No threshold changed, so the set of
+questions Maven answers is the same; only which memory answers them. The fixture gained two mixed
+note+fact cases (`ru-mixed-031`, `ru-mixed-032`), which is why the counts below are out of 27
+answerable cases and not 25: hash recall@1 36.0% (9/25) → 37.0% (10/27), e5 recall@1 72.0% (18/25) →
+70.4% (19/27) with answered-after-gate 68.0% → 66.7% and false recall unchanged at 1/5.
+
 ### 5. Ranking has no recency or type signal, and the store is not the bottleneck
 
 `internal/store/notes.go:67` sorts by cosine and uses `ts` only to break an exact float tie, which
@@ -143,14 +151,97 @@ the vector memory table was written by the old model, so after this deploy they
 against a new query. A live database needs every note and fact re-embedded before recall works at
 all. Filed as its own task.
 
+## Margin gate — 31-07-2026, third run
+
+Next-steps item 3, done. The absolute gate is replaced by a **margin gate**: answer only when the
+top hit beats the runner-up by more than delta (`top1 − top2 > δ`). Same fixture, same e5 embedder,
+same store as the run above. `internal/memory/gate.go` holds the check; both read paths call it
+(`cmd/mavend/recall.go` and the notes-RAG branch in `voice.go`). New knob `voice.query_min_margin`
+in `deploy/mavend.json`, default 0.008.
+
+### Why the absolute gate could not work, in one line of data
+
+The harness now prints the margin distributions, and they barely overlap where the raw scores
+overlap completely:
+
+| | top-1 score | margin (top1 − top2) |
+|---|---|---|
+| right note first (n=18) | min 0.810, median 0.862, max 0.890 | min 0.001, median 0.029, max 0.053 |
+| must stay silent (n=5) | min 0.795, median 0.815, max 0.835 | min 0.000, median 0.002, **max 0.019** |
+
+Four of the five must-be-silent cases have a margin at or under 0.002 — when there is nothing to
+recall, e5 finds several notes equally close and no clear winner. That is the signal the absolute
+score throws away.
+
+### The delta sweep
+
+Absolute gate held at 0.55 throughout.
+
+```
+delta 0.000: answered 18/25 (72%)  false recall 5/5
+delta 0.002: answered 17/25 (68%)  false recall 3/5
+delta 0.005: answered 17/25 (68%)  false recall 2/5
+delta 0.008: answered 17/25 (68%)  false recall 1/5   <- chosen
+delta 0.010: answered 15/25 (60%)  false recall 1/5
+delta 0.012: answered 14/25 (56%)  false recall 1/5
+delta 0.015: answered 12/25 (48%)  false recall 1/5
+delta 0.020: answered 11/25 (44%)  false recall 0/5
+delta 0.025: answered  9/25 (36%)  false recall 0/5
+delta 0.030: answered  8/25 (32%)  false recall 0/5
+delta 0.040: answered  4/25 (16%)  false recall 0/5
+delta 0.050: answered  2/25 ( 8%)  false recall 0/5
+delta 0.060: answered  0/25 ( 0%)  false recall 0/5
+```
+
+### Chosen: δ = 0.008
+
+It is the best point on the frontier, not a taste call. **0.008 dominates 0.010, 0.012 and 0.015
+outright** — same 1/5 false recall, 8 to 20 points more real recall. Everything below it buys recall
+back only by admitting more false recalls (0.005 → 2/5, 0.002 → 3/5). The next real improvement is
+0.020 at 0/5 false, and it costs 24 points of recall to get there.
+
+The brief's bar was "recall above 60% with false recall at 1/5 or better". 0.008 clears it with room:
+68% and 1/5.
+
+### Before / after
+
+| | absolute gate 0.55 (previous) | margin gate δ=0.008 |
+|---|---|---|
+| recall@1 (ranking, ungated) | 72.0% (18/25) | 72.0% (18/25) — unchanged, the gate does not rank |
+| **answered after the gate** | 72.0% (18/25) | **68.0% (17/25)** |
+| **false recall** | **5/5 (100%)** | **1/5 (20%)** |
+| fixture cases passed | 18/30 | **21/30** |
+
+Four false recalls removed for one real answer. That is the trade the spec asks for — she is not a
+guesser-of-truth. The one survivor is `en-pref-025` ("should i be offered wine"), which recalls a
+filler note at 0.796 with a 0.019 margin: the widest silent-case margin in the fixture, and it sits
+inside the real-recall range, so no delta removes it without taking real answers with it.
+
+### Does the absolute cutoff still earn its keep? Marginally — kept
+
+On this fixture with e5 it is a **no-op**: the lowest right-note score is 0.791, so 0.55 rejects
+nothing the margin does not already reject. It is kept for two reasons, neither glamorous. It still
+does real work for the hash embedder (its own sweep shows answers dropping from 16% to 0% between
+0.30 and 0.50), and it is the only thing standing between the user and a reply built from a store
+where everything is far away but one row happens to be a little less far — a near-empty database, or
+the stale-vector case below. Cheap insurance, no measured cost. If a later embedder makes it bite,
+the sweep is one command.
+
+### Caveat on the numbers
+
+Five must-be-silent cases is a thin basis for a 4-point decision. 1/5 and 2/5 differ by one case.
+The shape of the frontier is trustworthy — margins separate, absolute scores do not — but δ=0.008
+itself should be re-read off a bigger fixture (next-steps item 6) before anyone defends the third
+decimal.
+
 ## Next steps — ordered by value-to-risk; nothing here is a decision
 
 1. **Swap the embedder to `multilingual-e5-small` with `query:`/`passage:` prefixes.** One config
    change plus a prefix in `onnxembedder.go`, re-measurable in one command.
 2. **Re-run `make eval-recall`, then set the gate from the sweep** — not before. Any
    `query_min_score` picked against today's embedder describes a model on its way out.
-3. **Replace the absolute-score gate with a margin gate** (`top1 − top2 > δ`) — as the routing eval
-   concluded, absolute cosine cannot see a flat distribution.
+3. ~~**Replace the absolute-score gate with a margin gate**~~ — done, see the section above.
+   δ=0.008, false recall 5/5 → 1/5.
 4. **Delete or repair the dead `memStore` branch** at `voice.go:776` — search before the gate,
    gate it separately, or restrict it to facts and say so.
 5. **Add a mild time decay to ranking** — the newest statement of a preference is the true one.

ROUTING-EVAL-31-07-2026.md

@@ -66,15 +66,113 @@ Three things this run settles:
    `запиши что…` phrasings toward fact, and that suspicion stands — all five `ru-note-*`
    cases now land on fact. Tracked as Vikunja #375.
 
-**Thinking off is the best configuration measured so far**, on both accuracy and latency
-(Vikunja #376). That is worth understanding before flipping: routing is a short
-classification into a fixed enum with grammar-constrained output, so there is little to
-reason about, and the thinking trace mostly gives a small model room to talk itself out of
-the right answer. Phrasing is a different job and needs measuring separately.
+The `thinking off` column above read as the best configuration measured so far (Vikunja #376).
+**It was wrong** — see the controlled re-run below. Ignore that column.
 
 Still `6 / 6` missed clarify — the router has no way to say "I don't know" (Vikunja #359).
 That is unchanged by anything here.
 
+## Thinking off — 31-07-2026, controlled re-run (Vikunja #376)
+
+The "thinking off wins by 6 points" observation above **does not hold**. It was a measurement
+artefact, and the earlier table's `thinking off` column should be ignored.
+
+The thinking-off variant was scored by a hand-rolled HTTP client living in the test file
+instead of `llm.Client`. That copy did not send `repeat_penalty`, which the real router does
+send (`routeRepeatPenalty = 1.15`). So the two columns differed on two axes at once, and the
+one that mattered was the penalty, not the thinking mode.
+
+Re-measured with everything else held equal — same fixture, same prompt, same grammar, same
+sampling, same idle box, the three configurations run back to back and never concurrently:
+
+| | llm-only, thinking on | llm-only, thinking off | cascade+llm |
+|---|---|---|---|
+| intent-only accuracy | 59.2% (45/76) | 59.2% (45/76) | 61.8% (47/76) |
+| full accuracy (intent+slots+gate) | 38.2% (29/76) | 38.2% (29/76) | 57.9% (44/76) |
+| route errors | 3 | 3 | 0 |
+| grammar violations | 3 (all 3 route errors) | 3 (same 3 cases) | 0 |
+| missed clarify | 5 / 6 | 5 / 6 | 5 / 6 |
+| p50 latency | 836ms | 920ms | 810ms |
+| p95 latency | 1.41s | 2.00s | 1.31s |
+
+Thinking off is not just a tie on the headline numbers — it is identical case for case, with
+the same confusion matrix and the same three unparseable replies. The latency difference is
+run-to-run noise on one box, and it points the wrong way here.
+
+The reason is simpler than any accuracy argument: **this llama-server build ignores the
+request-level thinking switch for this model.** Probed directly against the running server
+with `chat_template_kwargs.enable_thinking = false`, `chat_template_kwargs.thinking = false`
+and top-level `reasoning_budget = 0` — all three return a byte-identical answer with the
+thinking trace still in `reasoning_content`, and the server reports the prompt prefix as
+cached, meaning the rendered template did not change. There was never anything being turned
+off, which is also why the numbers match exactly.
+
+Nothing was defaulted. `internal/llm` still has no `chat_template_kwargs` field, `VoiceConfig`
+has no thinking flag, and `deploy/mavend.json` is unchanged. The misleading third
+configuration is removed from `internal/router/eval` so the table it produced cannot be quoted
+again.
+
+Two caveats worth saying out loud:
+
+- **The fixture is 76 cases.** A 6-point difference on 76 cases is roughly 4-5 cases and would
+  not have been worth trusting even if it had reproduced. This one was exactly 0 cases, which
+  is a much easier call.
+- **This is one server build and one checkpoint** (`b9351`, Qwen3.5-0.8B Q4_K_M). If the
+  #122 checkpoint or a newer llama.cpp does honour the switch, the question reopens — but it
+  reopens as an unmeasured question, not as a 6-point win.
+
+Phrasing was **not** measured. Whether thinking helps there is still open, and now also blocked
+on the same "can we even turn it off" question.
+
+## Clock and calendar rule — 31-07-2026 (Vikunja #374)
+
+`routeSystem` never said whether "который час" or "какое число завтра" are `system` or
+`query`, and `system→query ×4` showed up in every run. The rule added says: the clock and the
+calendar date themselves are `system`; what is *written in* the calendar or in memory
+("что у меня завтра", "какие есть напоминания") stays `query`; and a time named inside a
+request ("напомни завтра…") is just a detail of the request, not a reason for `system`.
+
+That split is not a preference. In `cmd/mavend/voice.go` only `replySystem` owns the clock and
+the date formatter, so a clock question routed to `query` falls into the embedder + note RAG
+and answers "не знаю". The agenda, on the other hand, is answered by `ParseCalendarDate` +
+`CalendarEvents` *inside* the `query` branch, so that side has to stay `query`. The rule sits
+above the question test because every one of these utterances carries a question word and a
+later rule would never be reached.
+
+The fixture is now 77 cases: one calendar-agenda case was added
+(`ru-query-019` "что у меня стоит в календаре на послезавтра", intent `query`) specifically so
+an over-broad system rule cannot pass unnoticed. The clock/date cases (`ru-sys-001/002/005`,
+`en-sys-001`) already existed.
+
+Three runs, same box, back to back, never concurrently:
+
+| | baseline | first rule (too broad) | rule as committed |
+|---|---|---|---|
+| llm-only intent-only | 59.2% (45/76) | 54.5% (42/77) | 59.7% (46/77) |
+| llm-only full | 38.2% | 35.1% | 39.0% |
+| llm-only route errors | 3 | 4 | 5 |
+| llm-only p50 | 1.09s | 0.91s | 0.93s |
+| cascade+llm intent-only | 61.8% (47/76) | 58.4% | 62.3% (48/77) |
+| cascade+llm full | 57.9% | 54.5% | 59.7% |
+| cascade+llm route errors | 0 | 0 | 0 |
+| cascade+llm p50 | 0.91s | 0.80s | 1.04s |
+
+**The targeted bug is fixed and the headline number did not move.** `system→query ×4` is gone
+in both LLM configurations — the `time` and `date` tags go from 0/2 and 0/2 to 2/2 and 2/2 —
+but the model then over-applies the rule, and `query→system ×5` plus `reminder→system ×2`
+appear where they did not exist before. Net accuracy is a wash, inside the noise of a 77-case
+fixture.
+
+The first attempt is shown because it is the honest history: it said "спрашивает время, дату
+или день недели → system" with no scope, which swept up reminders, and it cost 3-5 points. It
+was tightened once, on the reasoning that a rule capturing "напомни завтра в 7" is simply
+wrong, and not tuned further. The remaining `query/reminder → system` over-trigger is a new,
+separate weakness of the sub-1B model and deserves its own task rather than more prompt
+kneading against a held-out fixture.
+
+The rule is kept. It is correct about what the daemon can answer, and the failure it replaces
+was silent ("не знаю" to "который час") while the one it introduces is loud.
+
 ## Findings
 
 ### 1. The resident model does route better — 50.0% vs 36.8%
@@ -145,11 +243,11 @@ Note the grammar's `string ::= "\"" ([^"\\] | "\\" .)* "\""` is unbounded, so no
 
 ### 7. Two hypotheses tested and closed
 
-- **Thinking mode is a non-issue.** Qwen3.5's template defaults `thinking = 1`, so
-  grammar-constrained JSON lands in `reasoning_content` with `content` empty —
-  `llm.Client`'s fallback handles it. A `thinking off` run scored *identically* (18/76,
-  48.7%, same p50). `internal/llm` deliberately does **not** grow a `chat_template_kwargs`
-  field.
+- **Thinking mode is a non-issue.** Confirmed twice now, the second time properly — see the
+  controlled re-run section. Grammar-constrained JSON lands in `reasoning_content` with
+  `content` empty and `llm.Client`'s fallback handles it; the request-level switch does
+  nothing on this build. `internal/llm` deliberately does **not** grow a
+  `chat_template_kwargs` field.
 - **Runaway array repetition does not reproduce.** An isolated smoke test with a stripped
   grammar emitted `{"intent":"reminder"}` until `MaxTokens`; under the real `routeSystem`
   prompt the few-shot examples anchor it to one object. 2 errors in 76, not 76.

cmd/mavend/clarify.go

@@ -40,9 +40,44 @@ var clarifyQuestions = map[dialogue.Slot]string{
 	dialogue.SlotFn:   "Что сделать?",
 }
 
-// clarifyDropped — she asked once, the answer still did not fill the gap, so
-// the request is gone. Said plainly, once, with no second question.
-const clarifyDropped = "Не разобрала — скажи целиком, пожалуйста."
+// clarifyGaveUp — she is out of questions and still does not have the slot. She
+// says so out loud: dropping the request in silence would leave him thinking it
+// landed. Feminine self-reference ("поняла"), as everywhere.
+const clarifyGaveUp = "Прости, я не поняла. Скажи, пожалуйста, по-другому."
+
+// clarifyExpired — his answer came after the TTL, so the parked request is
+// already gone. Same tone as clarifyGaveUp, different reason: too much time
+// passed, not "I did not understand". Feminine self-reference ("ждала",
+// "отпустила"); he is addressed with a plain imperative.
+const clarifyExpired = "Прости, я слишком долго ждала ответа и отпустила прошлую просьбу. Если она ещё нужна, скажи заново."
+
+// clarifyExpiredNotice returns that line when a parked question had just timed
+// out, and "" when nothing was parked. Call it right after
+// resolveClarifyAnswer: a live question is answered there, an expired one is
+// only reported here — the words themselves still go on to be routed fresh.
+func (h *reactiveHandler) clarifyExpiredNotice() string {
+	if h.clarifyStore == nil {
+		return ""
+	}
+	if !h.clarifyStore.TakeExpired(voiceDialogueID, h.now()) {
+		return ""
+	}
+	log.Printf("voice: clarify — parked question expired, telling him and routing the words fresh")
+	return clarifyExpired
+}
+
+// withNotice glues the expiry notice in front of this turn's reply. One turn
+// carries one reply on the wire, so the notice cannot be a message of its own —
+// but neither the notice nor the fresh answer may be dropped.
+func withNotice(notice, reply string) string {
+	if notice == "" {
+		return reply
+	}
+	if reply == "" {
+		return notice
+	}
+	return notice + " " + reply
+}
 
 // missingFor returns the slots a decision still needs, most important first.
 // Empty ⇒ there is nothing identifiable to ask about.
@@ -79,13 +114,14 @@ func (h *reactiveHandler) askClarify(dec router.Decision) (string, bool) {
 		return "", false
 	}
 	h.clarifyStore.Put(voiceDialogueID, &dialogue.PendingQuestion{
-		Intent:    dialogue.Intent(dec.Intent),
-		Slots:     toDialogueSlots(dec.Slots),
-		Missing:   []dialogue.Slot{slot},
-		Utterance: dec.Utterance,
-		Asked:     h.now(),
-		TTL:       clarifyTTL,
-		Attempts:  1, // asked once; MaxAttempts is 1, so there is no second ask
+		Intent:      dialogue.Intent(dec.Intent),
+		Slots:       toDialogueSlots(dec.Slots),
+		Missing:     []dialogue.Slot{slot},
+		Utterance:   dec.Utterance,
+		Asked:       h.now(),
+		TTL:         clarifyTTL,
+		Attempts:    1, // this ask
+		MaxAttempts: h.clarifyMaxAttempts,
 	})
 	log.Printf("voice: clarify — asked about %s for intent=%s", slot, dec.Intent)
 	return question, true
@@ -97,8 +133,9 @@ func (h *reactiveHandler) askClarify(dec router.Decision) (string, bool) {
 // resolveConfirm and checked in the same place.
 //
 // The answer is parsed with the same extractor the router uses, for the intent
-// she parked — no second parser. If it still does not fill the gap the request
-// is dropped: she does not ask again.
+// she parked — no second parser. If it still does not fill the gap she asks
+// again, up to MaxAttempts; after that she says out loud that she did not
+// understand. She never drops the request in silence.
 func (h *reactiveHandler) resolveClarifyAnswer(ctx context.Context, text string) (string, bool) {
 	if h.clarifyStore == nil {
 		return "", false
@@ -107,17 +144,14 @@ func (h *reactiveHandler) resolveClarifyAnswer(ctx context.Context, text string)
 	if q == nil {
 		return "", false
 	}
-	// One shot either way: the question is consumed whether or not the answer
-	// works, so a failed answer can't leave the question armed.
-	h.clarifyStore.Delete(voiceDialogueID)
 
 	intent := router.Intent(q.Intent)
 	answer := h.extractor.Extract(ctx, intent, text, h.now())
 	merged := q.Answer(text, toDialogueSlots(answer))
 	if len(dialogue.StillMissing(q.Missing, merged)) > 0 {
-		log.Printf("voice: clarify — answer %q did not fill %v, dropping", text, q.Missing)
-		return clarifyDropped, true
+		return h.reaskOrGiveUp(q, merged, text), true
 	}
+	h.clarifyStore.Delete(voiceDialogueID)
 
 	// Rebuild the decision as if it had routed cleanly, then run it down the
 	// normal path. Clarify is deliberately false and the intent is unchanged:
@@ -133,6 +167,29 @@ func (h *reactiveHandler) resolveClarifyAnswer(ctx context.Context, text string)
 	return h.finishClarified(ctx, dec), true
 }
 
+// reaskOrGiveUp handles an answer that left the gap open: ask the same question
+// again while she has attempts left, otherwise say she did not understand and
+// let the request go. Never returns "" — a mute give-up reads as "done".
+func (h *reactiveHandler) reaskOrGiveUp(q *dialogue.PendingQuestion, merged dialogue.Slots, text string) string {
+	question := ""
+	if len(q.Missing) > 0 {
+		question = clarifyQuestions[q.Missing[0]]
+	}
+	if question == "" || !q.CanAsk() {
+		h.clarifyStore.Delete(voiceDialogueID)
+		log.Printf("voice: clarify — gave up on %v after %d question(s), answer was %q", q.Missing, q.Attempts, text)
+		return clarifyGaveUp
+	}
+	// Re-park with whatever the answer DID give, the clock restarted and one
+	// more question spent.
+	q.Slots = merged
+	q.Attempts++
+	q.Asked = h.now()
+	h.clarifyStore.Put(voiceDialogueID, q)
+	log.Printf("voice: clarify — answer %q did not fill %v, asking again (attempt %d)", text, q.Missing, q.Attempts)
+	return question
+}
+
 // finishClarified runs a completed decision through the same steps a freshly
 // routed one takes: remember the turn, act, then phrase.
 func (h *reactiveHandler) finishClarified(ctx context.Context, dec router.Decision) string {
@@ -146,6 +203,10 @@ func (h *reactiveHandler) finishClarified(ctx context.Context, dec router.Decisi
 	if reply == "" {
 		reply = h.replier.Reply(dec)
 	}
+	if reply == "" {
+		// Belt: an empty reply here would be a silent drop.
+		reply = clarifyGaveUp
+	}
 	return reply
 }
 

cmd/mavend/clarify_test.go

@@ -86,7 +86,7 @@ func TestClarifyReminderCompletesOnAnswer(t *testing.T) {
 	if !handled {
 		t.Fatal("the answer to an open question must be consumed as an answer")
 	}
-	if reply == clarifyDropped {
+	if reply == clarifyGaveUp {
 		t.Fatalf("a good answer must not drop the request: %q", reply)
 	}
 
@@ -111,7 +111,7 @@ func TestClarifyFactCompletesOnAnswer(t *testing.T) {
 	if _, asked := h.askClarify(clarifyDec(router.IntentFact, router.Slots{Text: "запиши"}, "запиши")); !asked {
 		t.Fatal("a fact with no key should be asked about")
 	}
-	if reply, handled := h.resolveClarifyAnswer(ctx, "пил воду"); !handled || reply == clarifyDropped {
+	if reply, handled := h.resolveClarifyAnswer(ctx, "пил воду"); !handled || reply == clarifyGaveUp {
 		t.Fatalf("answer should complete the fact, handled=%v reply=%q", handled, reply)
 	}
 	if fact, err := st.LatestFact(ctx, "water"); err != nil || fact.Key != "water" {
@@ -137,26 +137,87 @@ func TestClarifyAnswerAfterTTLIsANewRequest(t *testing.T) {
 	}
 }
 
-// TestClarifyUnclearAnswerDropsWithoutAskingAgain — MaxAttempts is 1.
-func TestClarifyUnclearAnswerDropsWithoutAskingAgain(t *testing.T) {
+// TestClarifyAsksThreeTimesThenSaysSo — three questions are allowed, the fourth
+// is not, and running out is SPOKEN. Silence would read as "handled".
+func TestClarifyAsksThreeTimesThenSaysSo(t *testing.T) {
 	ctx := context.Background()
 	h, st, _ := newClarifyHandler(t)
 
 	if _, asked := h.askClarify(clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {
-		t.Fatal("expected a question")
+		t.Fatal("expected a first question")
 	}
+	// Two more unclear answers ⇒ two more questions (3 asks in total).
+	for i := 2; i <= 3; i++ {
+		reply, handled := h.resolveClarifyAnswer(ctx, "ну не знаю")
+		if !handled {
+			t.Fatalf("answer %d must be consumed as an answer", i)
+		}
+		if reply != "На когда напомнить?" {
+			t.Fatalf("attempt %d should ask again, got %q", i, reply)
+		}
+		if h.clarifyStore.Get(voiceDialogueID, h.now()) == nil {
+			t.Fatalf("attempt %d must leave the question armed", i)
+		}
+	}
+
 	reply, handled := h.resolveClarifyAnswer(ctx, "ну не знаю")
-	if !handled || reply != clarifyDropped {
-		t.Fatalf("an unclear answer should drop the request, handled=%v reply=%q", handled, reply)
+	if !handled || reply != clarifyGaveUp {
+		t.Fatalf("the fourth try must give up out loud, handled=%v reply=%q", handled, reply)
 	}
-	if strings.Contains(reply, "?") {
-		t.Fatalf("she must not ask a second question: %q", reply)
+	if reply == "" || strings.Contains(reply, "?") {
+		t.Fatalf("giving up must be spoken and must not be another question: %q", reply)
 	}
 	if h.clarifyStore.Get(voiceDialogueID, h.now()) != nil {
-		t.Fatal("a dropped request must leave no armed question")
+		t.Fatal("a given-up request must leave no armed question")
 	}
 	if reminders, err := st.DueReminders(ctx, h.now().Add(48*time.Hour)); err != nil || len(reminders) != 0 {
-		t.Fatalf("a dropped request must not create anything: reminders=%v err=%v", reminders, err)
+		t.Fatalf("a given-up request must not create anything: reminders=%v err=%v", reminders, err)
+	}
+}
+
+// TestClarifyMaxAttemptsIsConfigurable — one question when the config says one.
+func TestClarifyMaxAttemptsIsConfigurable(t *testing.T) {
+	ctx := context.Background()
+	h, _, _ := newClarifyHandler(t)
+	h.clarifyMaxAttempts = 1
+
+	if _, asked := h.askClarify(clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {
+		t.Fatal("expected a question")
+	}
+	if reply, handled := h.resolveClarifyAnswer(ctx, "ну не знаю"); !handled || reply != clarifyGaveUp {
+		t.Fatalf("with max 1 she must give up at once, handled=%v reply=%q", handled, reply)
+	}
+}
+
+// TestClarifyRestatedAnswerWins — «в 11:00», then «нет, в 15:00». The second
+// value is the one that lands.
+func TestClarifyRestatedAnswerWins(t *testing.T) {
+	ctx := context.Background()
+	h, st, _ := newClarifyHandler(t)
+
+	if _, asked := h.askClarify(clarifyDec(router.IntentReminder, router.Slots{Text: "напомни позвонить маме"}, "напомни позвонить маме")); !asked {
+		t.Fatal("expected a question")
+	}
+	// First answer parses, but re-park it by hand as if she had asked again:
+	// what matters here is that Answer prefers the newer value over the parked
+	// one, which is the case the daemon hits on a re-ask.
+	q := h.clarifyStore.Get(voiceDialogueID, h.now())
+	if q == nil {
+		t.Fatal("expected an armed question")
+	}
+	first := h.extractor.Extract(ctx, router.IntentReminder, "в 11:00", h.now())
+	q.Slots = q.Answer("в 11:00", toDialogueSlots(first))
+
+	if reply, handled := h.resolveClarifyAnswer(ctx, "нет, в 15:00"); !handled || reply == clarifyGaveUp {
+		t.Fatalf("the restated answer should complete the request, handled=%v reply=%q", handled, reply)
+	}
+	reminders, err := st.DueReminders(ctx, h.now().Add(48*time.Hour))
+	if err != nil || len(reminders) != 1 {
+		t.Fatalf("expected one reminder: %v err=%v", reminders, err)
+	}
+	want := h.extractor.Extract(ctx, router.IntentReminder, "в 15:00", h.now())
+	if !reminders[0].FireTs.Equal(want.Time) {
+		t.Fatalf("reminder at %v, want the restated %v", reminders[0].FireTs, want.Time)
 	}
 }
 
@@ -228,6 +289,36 @@ func TestNoQuestionWhenNothingIsMissing(t *testing.T) {
 	}
 }
 
+// TestClarifyExpiryIsAnnouncedAndWordsStillRoute — his answer lands after the
+// TTL: she must say the old request is gone AND still answer the new words.
+func TestClarifyExpiryIsAnnouncedAndWordsStillRoute(t *testing.T) {
+	ctx := context.Background()
+	h, _, now := newClarifyHandler(t)
+	emb := router.NewHashEmbedder(1024)
+	h.embedder = emb
+	h.router = buildRouter(emb, h.matcher, 0.55, nil)
+
+	if _, asked := h.askClarify(clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {
+		t.Fatal("expected a question")
+	}
+	*now = now.Add(clarifyTTL + time.Second)
+
+	reply := h.handleText(ctx, "как дела")
+	if !strings.HasPrefix(reply, clarifyExpired) {
+		t.Fatalf("expired question must be announced first, got %q", reply)
+	}
+	if strings.TrimSpace(strings.TrimPrefix(reply, clarifyExpired)) == "" {
+		t.Fatalf("the new words must still be answered, got only the notice: %q", reply)
+	}
+	if h.clarifyStore.Get(voiceDialogueID, h.now()) != nil {
+		t.Fatal("the expired question must be gone")
+	}
+	// The notice is said once, not on every later utterance.
+	if reply := h.handleText(ctx, "как дела"); strings.Contains(reply, clarifyExpired) {
+		t.Fatalf("notice repeated on a later turn: %q", reply)
+	}
+}
+
 // TestNoPendingQuestionFallsThrough — with nothing parked, an utterance routes
 // normally.
 func TestNoPendingQuestionFallsThrough(t *testing.T) {

cmd/mavend/main.go

@@ -58,6 +58,7 @@ import (
 	"github.com/kami/maven/internal/delivery/telegramsink"
 	"github.com/kami/maven/internal/ipc"
 	"github.com/kami/maven/internal/loop"
+	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/phraser"
 	"github.com/kami/maven/internal/store"
 	"github.com/kami/maven/internal/webauthn"
@@ -189,7 +190,9 @@ func (l *lockedAPI) MorningStatus(ctx context.Context) ([]ipc.MorningRoutineStat
 func run(args []string) error {
 	cfgPath := flag.String("config", defaultConfigPath(), "path to mavend JSON config")
 	wrappedKeyPath := flag.String("wrapped-key-file", "", "path to wrapped encryption key blob (enables cold-start unlock)")
+	reembed := flag.Bool("reembed", false, "re-embed every stored note and fact with the configured embedder, then serve normally (run once after an embedder swap)")
 	flag.CommandLine.Parse(args)
+	reembedOnStart = *reembed
 	cfg, err := config.Load(*cfgPath)
 	if err != nil {
 		return err
@@ -269,13 +272,13 @@ func run(args []string) error {
 		phr = phraser.NewStub()
 		if cfg.Phraser != nil {
 			pc := phraser.Config{
-				ModelPath:  cfg.Phraser.ModelPath,
-				BinPath:    cfg.Phraser.BinPath,
-				Listen:     cfg.Phraser.Listen,
-				NGpuLayers: cfg.Phraser.NGpuLayers,
-				NCtx:       cfg.Phraser.NCtx,
-				Timeout:    time.Duration(cfg.Phraser.Timeout),
-				Persona:    personaFromCfg(cfg),
+				ModelPath:    cfg.Phraser.ModelPath,
+				BinPath:      cfg.Phraser.BinPath,
+				Listen:       cfg.Phraser.Listen,
+				NGpuLayers:   cfg.Phraser.NGpuLayers,
+				NCtx:         cfg.Phraser.NCtx,
+				Timeout:      time.Duration(cfg.Phraser.Timeout),
+				ContextBlock: contextBlockFn(cfg, time.Now),
 			}
 			if pc.BinPath == "" {
 				pc.BinPath = "llama-server"
@@ -439,13 +442,13 @@ func run(args []string) error {
 			phr = phraser.NewStub()
 			if cfg.Phraser != nil {
 				pc := phraser.Config{
-					ModelPath:  cfg.Phraser.ModelPath,
-					BinPath:    cfg.Phraser.BinPath,
-					Listen:     cfg.Phraser.Listen,
-					NGpuLayers: cfg.Phraser.NGpuLayers,
-					NCtx:       cfg.Phraser.NCtx,
-					Timeout:    time.Duration(cfg.Phraser.Timeout),
-					Persona:    personaFromCfg(cfg),
+					ModelPath:    cfg.Phraser.ModelPath,
+					BinPath:      cfg.Phraser.BinPath,
+					Listen:       cfg.Phraser.Listen,
+					NGpuLayers:   cfg.Phraser.NGpuLayers,
+					NCtx:         cfg.Phraser.NCtx,
+					Timeout:      time.Duration(cfg.Phraser.Timeout),
+					ContextBlock: contextBlockFn(cfg, time.Now),
 				}
 				if pc.BinPath == "" {
 					pc.BinPath = "llama-server"
@@ -599,12 +602,29 @@ func run(args []string) error {
 	return nil
 }
 
-// personaFromCfg extracts the voice persona from the config, or returns ""
-// when voice isn't configured. Used to pass a character prompt into the
-// LLM phraser without requiring voice to be enabled.
-func personaFromCfg(cfg *config.Config) string {
-	if cfg.Voice != nil {
-		return cfg.Voice.Persona
+// personaFacts reads the optional, deployment-specific facts (his name, his
+// city, the free-text persona string) out of the config. Everything here may
+// be empty — the context block is correct without any of it.
+func personaFacts(cfg *config.Config) persona.Facts {
+	f := persona.Facts{
+		// Telegram lives outside the voice block, so it counts either way.
+		Telegram: cfg.Telegram != nil && cfg.Telegram.BotToken != "" && cfg.Telegram.ChatID != "",
 	}
-	return ""
+	if cfg.Voice == nil {
+		return f
+	}
+	f.OwnerName = cfg.Voice.OwnerName
+	f.City = cfg.Voice.City
+	f.Static = cfg.Voice.Persona
+	// Same test wireVoice uses to pick the real provider over the stub.
+	f.Weather = cfg.Voice.Weather != nil && cfg.Voice.Weather.Provider == "open-meteo"
+	f.Tools = len(cfg.Voice.Tools) > 0
+	return f
+}
+
+// contextBlockFn returns the per-turn renderer of the shared context block.
+// Per turn, not once at startup, because the block states the current time.
+func contextBlockFn(cfg *config.Config, now func() time.Time) func() string {
+	f := personaFacts(cfg)
+	return func() string { return f.Block(now()) }
 }

cmd/mavend/query_recall_test.go

@@ -0,0 +1,158 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"math"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/ipc"
+	"github.com/kami/maven/internal/memory"
+	"github.com/kami/maven/internal/phraser"
+	"github.com/kami/maven/internal/router"
+	"github.com/kami/maven/internal/voice"
+)
+
+// fixedEmbedder hands back a vector chosen per text, so a test can say exactly
+// how close each stored memory is to the question. The real embedders make
+// scores that are realistic but not controllable, and this test is about the
+// gate, not about the embedder.
+type fixedEmbedder struct{ vecs map[string][]float32 }
+
+func (f *fixedEmbedder) Dim() int     { return 4 }
+func (f *fixedEmbedder) Close() error { return nil }
+
+func (f *fixedEmbedder) Embed(_ context.Context, text string) ([]float32, error) {
+	v, ok := f.vecs[text]
+	if !ok {
+		return nil, fmt.Errorf("fixedEmbedder: no vector for %q", text)
+	}
+	return v, nil
+}
+
+// scoreVec builds a unit vector whose cosine against the query vector
+// (1,0,0,0) is exactly score.
+func scoreVec(score float64) []float32 {
+	rest := math.Sqrt(1 - score*score)
+	return []float32{float32(score), float32(rest), 0, 0}
+}
+
+// recordingPhraser remembers what the query path handed it to phrase, which is
+// how the test can tell which pass produced the answer.
+type recordingPhraser struct {
+	*phraser.Stub
+	notes []string
+}
+
+func (r *recordingPhraser) PhraseQuery(ctx context.Context, utterance string, notes []string) (string, error) {
+	r.notes = notes
+	return r.Stub.PhraseQuery(ctx, utterance, notes)
+}
+
+// recallCase — one stored memory: its text, how close it is to the question,
+// whether it is a note or a fact, and whether the notes table holds it too.
+type recallCase struct {
+	text  string
+	score float64
+	kind  string
+}
+
+// buildRecallHandler stores the given memories and returns a handler whose
+// query path can be run directly. Notes go into BOTH the notes table and the
+// vector index, which is what the daemon does (voice.go's IntentNote).
+func buildRecallHandler(t *testing.T, question string, mems []recallCase) (*reactiveHandler, *recordingPhraser) {
+	t.Helper()
+	ctx := context.Background()
+	st := newTestStore(t)
+	emb := &fixedEmbedder{vecs: map[string][]float32{question: {1, 0, 0, 0}}}
+	mem := memory.NewInMemoryStore()
+	now := time.Now()
+
+	for i, m := range mems {
+		vec := scoreVec(m.score)
+		emb.vecs[m.text] = vec
+		id := fmt.Sprintf("%s:%d", m.kind, i)
+		if m.kind == "note" {
+			if _, err := st.WriteNote(ctx, now, m.text, vec, "tap:voice"); err != nil {
+				t.Fatalf("WriteNote: %v", err)
+			}
+		}
+		if err := mem.Insert(ctx, id, vec, map[string]string{"text": m.text, "type": m.kind}); err != nil {
+			t.Fatalf("memory insert: %v", err)
+		}
+	}
+
+	phr := &recordingPhraser{Stub: phraser.NewStub()}
+	h := &reactiveHandler{
+		api:             ipc.NewStoreAPI(st),
+		embedder:        emb,
+		replier:         voice.NewStubReplier(),
+		phraser:         phr,
+		now:             func() time.Time { return now },
+		memStore:        mem,
+		dataStore:       st,
+		queryMinScore:   0.55,
+		queryMinMargin:  0.008,
+		weatherProvider: nil,
+	}
+	return h, phr
+}
+
+func askQuery(t *testing.T, h *reactiveHandler, question string) string {
+	t.Helper()
+	return h.applyAction(context.Background(), router.Decision{
+		Intent:    router.IntentQuery,
+		Utterance: question,
+	})
+}
+
+// TestQueryRecallNoteCanWin — the note-recall regression (Vikunja #373). Notes
+// and facts share one vector index, and a note that clearly beats everything
+// else must be the answer. Before the fix the memory pass only ran after the
+// notes-only gate had already rejected the same note at the same score, so only
+// a fact could ever come back from it.
+func TestQueryRecallNoteCanWin(t *testing.T) {
+	const q = "где молоко"
+
+	t.Run("a clearly best note answers", func(t *testing.T) {
+		h, phr := buildRecallHandler(t, q, []recallCase{
+			{text: "молоко стоит в холодильнике", score: 0.90, kind: "note"},
+			{text: "выучил пару аккордов", score: 0.50, kind: "note"},
+		})
+		reply := askQuery(t, h, q)
+		if want := "вот что я нашла: молоко стоит в холодильнике"; reply != want {
+			t.Errorf("reply %q, want %q", reply, want)
+		}
+		// One text, the winning memory's — the answer came from the memory
+		// pass, not from handing the phraser every note in the table.
+		if len(phr.notes) != 1 || phr.notes[0] != "молоко стоит в холодильнике" {
+			t.Errorf("phraser got %q, want just the recalled note", phr.notes)
+		}
+	})
+
+	// The other half of "one gate over everything": a fact that matches better
+	// than the best note now answers, instead of losing to a note that only had
+	// to beat other notes.
+	t.Run("the better-matching fact answers", func(t *testing.T) {
+		h, _ := buildRecallHandler(t, q, []recallCase{
+			{text: "молоко стоит в холодильнике", score: 0.80, kind: "note"},
+			{text: "купил молоко в среду", score: 0.95, kind: "fact"},
+		})
+		if reply := askQuery(t, h, q); reply != "купил молоко в среду" {
+			t.Errorf("reply %q, want the fact read back", reply)
+		}
+	})
+
+	// The gate is untouched: two memories this close mean the embedder cannot
+	// tell them apart, and silence still beats a coin flip.
+	t.Run("no clear best stays silent", func(t *testing.T) {
+		h, _ := buildRecallHandler(t, q, []recallCase{
+			{text: "молоко стоит в холодильнике", score: 0.860, kind: "note"},
+			{text: "молоко закончилось", score: 0.858, kind: "note"},
+		})
+		if reply := askQuery(t, h, q); reply != "не знаю." {
+			t.Errorf("reply %q, want silence", reply)
+		}
+	})
+}

cmd/mavend/recall.go

@@ -2,24 +2,24 @@ package main
 
 import "github.com/kami/maven/internal/memory"
 
-// bestRecall is the read side of the long-term memory store: the top hit's
-// stored text when it clears the confidence gate. This recalls across BOTH
-// notes and facts (facts aren't in the notes table, so this is the only path
-// that can answer "when did I last …?" from a captured fact). A note hit here
-// is redundant with the notes-RAG path — by design; the two indexes can diverge
-// once the backend is swapped for a persistent/external store. ok=false when
-// there's no hit above the threshold or the hit carries no text.
-func bestRecall(results []memory.Result, min float64) (string, bool) {
-	if len(results) == 0 {
-		return "", false
+// bestRecall is the read side of the long-term memory store: the top hit when
+// it clears the confidence gate. The index holds BOTH notes and facts, and
+// either can win — the caller looks at the returned hit's meta["type"] to see
+// which. Facts aren't in the notes table, so this is the only path that can
+// answer "when did I last …?" from a captured fact.
+//
+// The whole hit is returned, not just its text, because "which memory answered"
+// decides how the answer is said: a note gets phrased in Maven's voice, a fact
+// is read back as stored.
+//
+// ok=false when the hit fails the confidence gate (see memory.Confident: an
+// absolute floor plus a margin over the runner-up) or carries no text.
+func bestRecall(results []memory.Result, minScore, minMargin float64) (memory.Result, bool) {
+	if !memory.Confident(results, minScore, minMargin) {
+		return memory.Result{}, false
 	}
-	top := results[0]
-	if top.Score < min {
-		return "", false
+	if results[0].Meta["text"] == "" {
+		return memory.Result{}, false
 	}
-	text := top.Meta["text"]
-	if text == "" {
-		return "", false
-	}
-	return text, true
+	return results[0], true
 }

cmd/mavend/recall_test.go

@@ -8,23 +8,24 @@ import (
 
 func TestBestRecall(t *testing.T) {
 	const min = 0.55
+	const margin = 0.008
 
 	t.Run("empty results", func(t *testing.T) {
-		if _, ok := bestRecall(nil, min); ok {
+		if _, ok := bestRecall(nil, min, margin); ok {
 			t.Error("empty results returned ok")
 		}
 	})
 
 	t.Run("top below threshold", func(t *testing.T) {
 		res := []memory.Result{{Score: 0.4, Meta: map[string]string{"text": "выпил воды"}}}
-		if _, ok := bestRecall(res, min); ok {
+		if _, ok := bestRecall(res, min, margin); ok {
 			t.Error("below-threshold hit returned ok")
 		}
 	})
 
 	t.Run("hit without text meta", func(t *testing.T) {
 		res := []memory.Result{{Score: 0.9, Meta: map[string]string{"type": "fact"}}}
-		if _, ok := bestRecall(res, min); ok {
+		if _, ok := bestRecall(res, min, margin); ok {
 			t.Error("textless hit returned ok")
 		}
 	})
@@ -34,12 +35,43 @@ func TestBestRecall(t *testing.T) {
 			{Score: 0.82, Meta: map[string]string{"text": "выпил воды в три часа", "type": "fact"}},
 			{Score: 0.60, Meta: map[string]string{"text": "другое"}},
 		}
-		got, ok := bestRecall(res, min)
+		got, ok := bestRecall(res, min, margin)
 		if !ok {
 			t.Fatal("clearing hit not returned")
 		}
-		if got != "выпил воды в три часа" {
-			t.Errorf("wrong text: %q", got)
+		if got.Meta["text"] != "выпил воды в три часа" {
+			t.Errorf("wrong text: %q", got.Meta["text"])
+		}
+		if got.Meta["type"] != "fact" {
+			t.Errorf("kind lost: %q", got.Meta["type"])
+		}
+	})
+
+	// The index holds notes and facts together, so a note has to be able to win
+	// it — for a long time it could not (Vikunja #373).
+	t.Run("a note can win", func(t *testing.T) {
+		res := []memory.Result{
+			{Score: 0.86, Meta: map[string]string{"text": "молоко в холодильнике", "type": "note"}},
+			{Score: 0.61, Meta: map[string]string{"text": "выпил воды", "type": "fact"}},
+		}
+		got, ok := bestRecall(res, min, margin)
+		if !ok {
+			t.Fatal("clearly-best note not returned")
+		}
+		if got.Meta["type"] != "note" || got.Meta["text"] != "молоко в холодильнике" {
+			t.Errorf("got %v, want the note", got.Meta)
+		}
+	})
+
+	// The runner-up is almost as close, so the embedder cannot tell the two
+	// notes apart. Silence beats reading back a coin flip.
+	t.Run("runner-up too close", func(t *testing.T) {
+		res := []memory.Result{
+			{Score: 0.860, Meta: map[string]string{"text": "выпил воды в три часа"}},
+			{Score: 0.858, Meta: map[string]string{"text": "другое"}},
+		}
+		if _, ok := bestRecall(res, min, margin); ok {
+			t.Error("thin-margin hit returned ok")
 		}
 	})
 }

cmd/mavend/replier_llm.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/router"
 	"github.com/kami/maven/internal/voice"
 )
@@ -23,13 +24,19 @@ type completer interface {
 type llmReplier struct {
 	c    completer
 	stub *voice.StubReplier
+
+	// block renders the shared context block per turn (who he is, the time).
+	// nil ⇒ the prompt stands alone.
+	block func() string
 }
 
-func newLLMReplier(c completer) *llmReplier {
-	return &llmReplier{c: c, stub: voice.NewStubReplier()}
+func newLLMReplier(c completer, block func() string) *llmReplier {
+	return &llmReplier{c: c, stub: voice.NewStubReplier(), block: block}
 }
 
-const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}.`
+const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Владелец — мужчина, говоришь с ним на "ты", в единственном числе; никогда не "вы"/"ваш" и не "он"/"его". Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
+Пример: {"response": "Записала, что ты выпил стакан воды.", "mood": "neutral"}
+Никогда не пиши "..." в поле response.`
 
 func (r *llmReplier) Reply(d router.Decision) string {
 	if d.Clarify {
@@ -38,7 +45,7 @@ func (r *llmReplier) Reply(d router.Decision) string {
 	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
 	defer cancel()
 	out, err := r.c.Complete(ctx, llm.Req{
-		System:        replySystem,
+		System:        persona.Prepend(r.block, replySystem),
 		User:          replyContext(d),
 		MaxTokens:     512,
 		RepeatPenalty: 1.3,

cmd/mavend/replier_llm_test.go

@@ -17,7 +17,7 @@ type mockCompleter struct {
 func (m mockCompleter) Complete(_ context.Context, _ llm.Req) (string, error) { return m.out, m.err }
 
 func TestLLMReplierReturnsLLMReply(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: `{"response":"записала, кофе закончился","mood":"neutral"}`})
+	r := newLLMReplier(mockCompleter{out: `{"response":"записала, кофе закончился","mood":"neutral"}`}, nil)
 	got := r.Reply(router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}})
 	if got != "записала, кофе закончился" {
 		t.Errorf("got %q, want %q", got, "записала, кофе закончился")
@@ -25,7 +25,7 @@ func TestLLMReplierReturnsLLMReply(t *testing.T) {
 }
 
 func TestLLMReplierFallsBackToPlainText(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: "записала, кофе закончился"})
+	r := newLLMReplier(mockCompleter{out: "записала, кофе закончился"}, nil)
 	got := r.Reply(router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}})
 	if got != "записала, кофе закончился" {
 		t.Errorf("got %q, want %q", got, "записала, кофе закончился")
@@ -33,7 +33,7 @@ func TestLLMReplierFallsBackToPlainText(t *testing.T) {
 }
 
 func TestLLMReplierFallsBackToStubOnError(t *testing.T) {
-	r := newLLMReplier(mockCompleter{err: errTestLLMDown})
+	r := newLLMReplier(mockCompleter{err: errTestLLMDown}, nil)
 	noteDec := router.Decision{Intent: router.IntentNote}
 	got := r.Reply(noteDec)
 	want := voice.NewStubReplier().Reply(noteDec)
@@ -43,7 +43,7 @@ func TestLLMReplierFallsBackToStubOnError(t *testing.T) {
 }
 
 func TestLLMReplierFallsBackToStubOnEmpty(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: ""})
+	r := newLLMReplier(mockCompleter{out: ""}, nil)
 	noteDec := router.Decision{Intent: router.IntentNote}
 	got := r.Reply(noteDec)
 	want := voice.NewStubReplier().Reply(noteDec)
@@ -53,7 +53,7 @@ func TestLLMReplierFallsBackToStubOnEmpty(t *testing.T) {
 }
 
 func TestLLMReplierClarifyUsesStub(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: "я всё поняла"})
+	r := newLLMReplier(mockCompleter{out: "я всё поняла"}, nil)
 	clarifyDec := router.Decision{Clarify: true}
 	got := r.Reply(clarifyDec)
 	want := voice.NewStubReplier().Reply(clarifyDec)

cmd/mavend/system_reply_test.go

@@ -0,0 +1,78 @@
+package main
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// systemHandler — a handler with nothing but a fixed clock, which is all
+// replySystem needs.
+func systemHandler(now time.Time) *reactiveHandler {
+	return &reactiveHandler{now: func() time.Time { return now }}
+}
+
+// TestReplySystemDateOffset — "какое число завтра" must answer tomorrow's
+// date, not today's (Vikunja #388).
+func TestReplySystemDateOffset(t *testing.T) {
+	// Thursday, 30 July 2026.
+	now := time.Date(2026, 7, 30, 14, 5, 0, 0, time.UTC)
+	h := systemHandler(now)
+	cases := []struct{ utterance, want string }{
+		{"какое сегодня число", "сегодня четверг, 30 июля 2026 года"},
+		{"какое число", "сегодня четверг, 30 июля 2026 года"},
+		{"какое число завтра", "завтра пятница, 31 июля 2026 года"},
+		{"какое число послезавтра", "послезавтра суббота, 1 августа 2026 года"},
+		{"какое было число вчера", "вчера среда, 29 июля 2026 года"},
+	}
+	for _, c := range cases {
+		got := h.replySystem(context.Background(), router.Decision{Utterance: c.utterance})
+		if got != c.want {
+			t.Errorf("replySystem(%q) = %q, want %q", c.utterance, got, c.want)
+		}
+	}
+}
+
+// A day she cannot work out must not come back as today's date — that is the
+// same silent wrong answer #388 was about, one step further out.
+func TestReplySystemUnknownDayIsHonest(t *testing.T) {
+	now := time.Date(2026, 7, 30, 14, 5, 0, 0, time.UTC)
+	h := systemHandler(now)
+	for _, u := range []string{
+		"какое число в пятницу",
+		"какое число через неделю",
+		"какое число в понедельник",
+	} {
+		got := h.replySystem(context.Background(), router.Decision{Utterance: u})
+		if got != onlyNearDaysReply {
+			t.Errorf("replySystem(%q) = %q, want the honest reply", u, got)
+		}
+	}
+	// The days she does know must not be caught by the same guard.
+	if got := h.replySystem(context.Background(), router.Decision{Utterance: "какое число завтра"}); got == onlyNearDaysReply {
+		t.Error("завтра was treated as an unknown day")
+	}
+}
+
+// TestReplySystemClockCity — the clock arm must not answer local time for a
+// question about another city (Vikunja #388). She keeps one clock, so every
+// named place gets the honest "local time only" answer.
+func TestReplySystemClockCity(t *testing.T) {
+	now := time.Date(2026, 7, 30, 12, 0, 0, 0, time.UTC)
+	h := systemHandler(now)
+	cases := []struct{ utterance, want string }{
+		{"который час", "сейчас 12 часов ровно"},
+		{"который час в киеве", onlyLocalTimeReply},
+		{"сколько времени в москве", onlyLocalTimeReply},
+		{"который час в лондоне", onlyLocalTimeReply},
+		{"который час в бишкеке", onlyLocalTimeReply},
+	}
+	for _, c := range cases {
+		got := h.replySystem(context.Background(), router.Decision{Utterance: c.utterance})
+		if got != c.want {
+			t.Errorf("replySystem(%q) = %q, want %q", c.utterance, got, c.want)
+		}
+	}
+}

cmd/mavend/voice.go

@@ -171,6 +171,7 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		emb = router.NewHashEmbedder(1024)
 	}
 	w.embedder = emb
+	checkStoredEmbedder(dataStore, emb)
 
 	// ----- tool executor (the enabled act allowlist, store-backed) -----
 	// Config tools are the declarative bootstrap: seed them into the store as
@@ -206,11 +207,11 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	if threshold <= 0 {
 		threshold = config.DefaultRouterThreshold
 	}
-	// Both routing paths are weak on held-out utterances — the classifier gets
-	// 36.8% of intents right, the resident model 50.0% and much slower. Off by
-	// default (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.LLMRouter, llmClient))
+	// The resident model routes by default: 63.2% of held-out intents right
+	// 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(), llmClient))
 
 	// ----- sessions registry (shared with voicesink) -----
 	sessions := voice.NewSessions()
@@ -227,14 +228,25 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	}
 
 	// ----- dialogue (multi-turn slot carry-over; 2-min follow-up window) -----
-	dialogueSessions := dialogue.NewSessionStore(2 * time.Minute)
+	// Store-backed when the daemon passes a store, so a restart mid-conversation
+	// keeps the thread (Vikunja #363). Sessions past their TTL are dropped on
+	// load, never revived. Clarify's parked question stays in memory only.
+	var dialogueSessions *dialogue.SessionStore
+	if dataStore != nil {
+		dialogueSessions = dialogue.NewPersistentSessionStore(2*time.Minute, dataStore)
+		if err := dialogueSessions.Load(context.Background(), time.Now()); err != nil {
+			log.Printf("dialogue: load saved sessions: %v", err)
+		}
+	} else {
+		dialogueSessions = dialogue.NewSessionStore(2 * time.Minute)
+	}
 	clarifyStore := dialogue.NewClarifyStore(clarifyTTL)
 	timeParser := router.NewPythonDateParser()
 
 	// ----- replier (LLM-backed when the engine is on, Stub floor otherwise) -----
 	replier := voice.Replier(voice.NewStubReplier())
 	if llmClient != nil {
-		replier = newLLMReplier(llmClient)
+		replier = newLLMReplier(llmClient, contextBlockFn(cfg, time.Now))
 	}
 
 	// ----- the handler (the reactive path; closes over stt / tts / router / coreAPI / memory) -----
@@ -255,10 +267,13 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		dataStore:        dataStore,
 		dialogueSessions: dialogueSessions,
 		clarifyStore:     clarifyStore,
-		extractor:        router.Extractor{Time: timeParser, Acts: matcher, Facts: router.DefaultFactParser{}},
-		queryMinScore:    cfg.Voice.QueryMinScore,
-		timeParser:       timeParser,
-		ecosystem:        eco,
+		// 0 here (unset config) ⇒ the dialogue default.
+		clarifyMaxAttempts: cfg.Voice.ClarifyMaxAttempts,
+		extractor:          router.Extractor{Time: timeParser, Acts: matcher, Facts: router.DefaultFactParser{}},
+		queryMinScore:      cfg.Voice.QueryMinScore,
+		queryMinMargin:     cfg.Voice.QueryMinMargin,
+		timeParser:         timeParser,
+		ecosystem:          eco,
 	}
 
 	// ----- the server (TCP listener) -----
@@ -300,6 +315,9 @@ type reactiveHandler struct {
 	// load-bearing math (same posture as the presence thresholds). Set by
 	// wireVoice from VoiceConfig; default 0.55.
 	queryMinScore float64
+	// queryMinMargin — the second half of that gate: how far the top hit must
+	// beat the runner-up. 0 ⇒ margin off.
+	queryMinMargin float64
 
 	// timeParser — used as a fallback for stage-0 reminder grammar matches
 	// (where the extractor didn't run). Shared with the router's extractor.
@@ -314,6 +332,10 @@ type reactiveHandler struct {
 	// clarify.go). nil ⇒ she falls back to the canned "не поняла" reply.
 	clarifyStore *dialogue.ClarifyStore
 
+	// clarifyMaxAttempts — questions per request before she gives up out loud.
+	// 0 ⇒ dialogue.DefaultMaxAttempts (3). Set from VoiceConfig.
+	clarifyMaxAttempts int
+
 	// extractor parses the answer to an open question, with the same parsers
 	// the router's own stage-2 uses.
 	extractor router.Extractor
@@ -392,9 +414,17 @@ func (h *reactiveHandler) HandlePushToTalk(ctx context.Context, req voice.PushTo
 		return h.reply(ctx, reply, nil)
 	}
 
-	// 1b2. clarify answer — if she asked a question last turn, this utterance is
-	// its answer, not a fresh command. After the confirm check: a y/n gate is
-	// armed by her own prompt and is the narrower claim on the utterance.
+	// 1b2. expired clarify — a question was parked but its TTL ran out, so the
+	// request behind it is gone. Say that out loud (see clarify.go) and carry
+	// on: these words are still routed as a fresh utterance below, with the
+	// notice glued in front of whatever the fresh routing answers. Checked
+	// BEFORE the answer path: reading a parked question drops an expired one.
+	expiredNotice := h.clarifyExpiredNotice()
+
+	// 1b3. clarify answer — if she asked a live question last turn, this
+	// utterance is its answer, not a fresh command. After the confirm check: a
+	// y/n gate is armed by her own prompt and is the narrower claim on the
+	// utterance.
 	if reply, handled := h.resolveClarifyAnswer(ctx, text); handled {
 		return h.reply(ctx, reply, nil)
 	}
@@ -404,7 +434,7 @@ func (h *reactiveHandler) HandlePushToTalk(ctx context.Context, req voice.PushTo
 	// unreliably (it's a command, not a free-form query), so we match it
 	// before routing. Same pattern as the confirm turn above.
 	if reply, handled := h.resolveQuietToggle(ctx, text); handled {
-		return h.reply(ctx, reply, nil)
+		return h.reply(ctx, withNotice(expiredNotice, reply), nil)
 	}
 
 	// 2. router — classify the utterance.
@@ -413,10 +443,10 @@ func (h *reactiveHandler) HandlePushToTalk(ctx context.Context, req voice.PushTo
 		// ErrNoIntents ⇒ classifier unseeded (cold boot). reply with a
 		// "still warming up" rather than a wire error.
 		if errors.Is(err, router.ErrNoIntents) {
-			return h.reply(ctx, "я ещё не понимаю свободную речь — скоро научусь.", nil)
+			return h.reply(ctx, withNotice(expiredNotice, "я ещё не понимаю свободную речь — скоро научусь."), nil)
 		}
 		log.Printf("voice: router error: %v", err)
-		return h.reply(ctx, "не получилось разобрать команду.", nil)
+		return h.reply(ctx, withNotice(expiredNotice, "не получилось разобрать команду."), nil)
 	}
 
 	// 2b. dialogue — fill this turn's missing slots from a prior same-intent
@@ -437,7 +467,7 @@ func (h *reactiveHandler) HandlePushToTalk(ctx context.Context, req voice.PushTo
 	// stands.
 	if dec.Clarify {
 		if question, asked := h.askClarify(dec); asked {
-			return h.reply(ctx, question, nil)
+			return h.reply(ctx, withNotice(expiredNotice, question), nil)
 		}
 	}
 
@@ -453,7 +483,7 @@ func (h *reactiveHandler) HandlePushToTalk(ctx context.Context, req voice.PushTo
 
 	// 5. tts — synthesise the reply text; return to the voice server which
 	// ships it back on the conn.
-	return h.reply(ctx, replyText, nil)
+	return h.reply(ctx, withNotice(expiredNotice, replyText), nil)
 }
 
 // handleText — the core reactive path without stt/tts: confirm check →
@@ -468,7 +498,9 @@ func (h *reactiveHandler) handleText(ctx context.Context, text string) string {
 		return reply
 	}
 
-	// 1b2. clarify answer — same check as HandlePushToTalk.
+	// 1b2/1b3. expired clarify then clarify answer — same order and reasons as
+	// HandlePushToTalk.
+	expiredNotice := h.clarifyExpiredNotice()
 	if reply, handled := h.resolveClarifyAnswer(ctx, text); handled {
 		return reply
 	}
@@ -477,10 +509,10 @@ func (h *reactiveHandler) handleText(ctx context.Context, text string) string {
 	dec, err := h.router.Route(ctx, text, h.now())
 	if err != nil {
 		if errors.Is(err, router.ErrNoIntents) {
-			return "я ещё не понимаю свободную речь — скоро научусь."
+			return withNotice(expiredNotice, "я ещё не понимаю свободную речь — скоро научусь.")
 		}
 		log.Printf("voice: handleText router error: %v", err)
-		return "не получилось разобрать команду."
+		return withNotice(expiredNotice, "не получилось разобрать команду.")
 	}
 	log.Printf("voice: route result: intent=%s slots=%+v", dec.Intent, dec.Slots)
 
@@ -497,7 +529,7 @@ func (h *reactiveHandler) handleText(ctx context.Context, text string) string {
 	// 2c. clarify — same as HandlePushToTalk: ask about the one missing thing.
 	if dec.Clarify {
 		if question, asked := h.askClarify(dec); asked {
-			return question
+			return withNotice(expiredNotice, question)
 		}
 	}
 
@@ -509,7 +541,7 @@ func (h *reactiveHandler) handleText(ctx context.Context, text string) string {
 	if replyText == "" {
 		replyText = h.replier.Reply(dec)
 	}
-	return replyText
+	return withNotice(expiredNotice, replyText)
 }
 
 // applyAction — executes the router's Decision. Intent-by-intent:
@@ -720,7 +752,9 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 		}
 
 		// Calendar questions: "что у меня сегодня?", "планы на завтра?"
-		if date, ok := router.ParseCalendarDate(dec.Utterance, time.Now()); ok {
+		// h.now(), not time.Now(): the handler's clock is the injected one, so
+		// this arm can be tested at a fixed time like the rest.
+		if date, ok := router.ParseCalendarDate(dec.Utterance, h.now()); ok {
 			events, err := h.api.CalendarEvents(ctx, date, date.Add(24*time.Hour))
 			if err != nil {
 				log.Printf("voice: calendar events: %v", err)
@@ -755,27 +789,54 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 			log.Printf("voice: embed query: %v", err)
 			return "не получилось найти ответ."
 		}
+		// Long-term memory first: ONE search over everything Maven remembers
+		// (notes and facts share this index) and ONE confidence gate, so the
+		// memory that is clearly the best match answers — a note just as much
+		// as a fact.
+		//
+		// This used to run only after the notes-only gate below had already
+		// rejected the same note at the same score, which no note could ever
+		// survive a second time: the branch could only return a fact (#373).
+		// Order, not the gate, was the bug — the set of questions Maven answers
+		// is unchanged, only which memory gets to answer them.
+		if h.memStore != nil {
+			if hits, herr := h.memStore.Search(ctx, vec, 3); herr == nil {
+				if hit, ok := bestRecall(hits, h.queryMinScore, h.queryMinMargin); ok {
+					text := hit.Meta["text"]
+					// A note is phrased in Maven's voice; a fact is read back
+					// as it was stored.
+					if hit.Meta["type"] == "note" {
+						if reply, perr := h.phraser.PhraseQuery(ctx, dec.Utterance, []string{text}); perr == nil && reply != "" {
+							return reply
+						}
+					}
+					return text
+				}
+			} else {
+				log.Printf("voice: memory search: %v", herr)
+			}
+		}
+
 		notes, err := h.api.QueryNotes(ctx, vec, 5)
 		if err != nil {
 			log.Printf("voice: query notes: %v", err)
 			return "не получилось найти ответ."
 		}
-		// Confidence gate: below threshold, say "I don't know" rather than read
-		// back the least-unrelated note — a confident wrong recall is worse than
-		// a gap (spec's "not a guesser-of-truth"). Same instinct as the loop's
-		// since(key)==null → don't fire. Tuned for the ONNX embedder; the Hash
-		// floor scores lexically and may rarely clear it.
-		if len(notes) == 0 || notes[0].Score < h.queryMinScore {
-			// Long-term memory recall (notes + facts) before general knowledge:
-			// the notes table can't answer fact questions, but the memory store
-			// indexes both. Only runs when notes-RAG already gave up → additive.
-			if h.memStore != nil {
-				if hits, herr := h.memStore.Search(ctx, vec, 3); herr == nil {
-					if text, ok := bestRecall(hits, h.queryMinScore); ok {
-						return text
-					}
-				}
-			}
+		// Notes-only pass, for notes the vector index above does not hold (an
+		// older note written before it existed). Same gate, notes-only
+		// candidates.
+		//
+		// Confidence gate: below it, say "I don't know" rather than read back
+		// the least-unrelated note — a confident wrong recall is worse than a
+		// gap (spec's "not a guesser-of-truth"). Same instinct as the loop's
+		// since(key)==null → don't fire. Two parts: an absolute cosine floor,
+		// and a margin over the runner-up, which is the part that works with
+		// the e5 embedder's narrow score band. See memory.Confident.
+		noteScores := make([]float64, len(notes))
+		for i, n := range notes {
+			noteScores[i] = n.Score
+		}
+		if !memory.ConfidentScores(noteScores, h.queryMinScore, h.queryMinMargin) {
 			// Try general knowledge from the phraser before giving up
 			reply, err := h.phraser.PhraseQuery(ctx, dec.Utterance, nil)
 			if err != nil || reply == "" {
@@ -877,6 +938,101 @@ var ruMonths = []string{
 	"июля", "августа", "сентября", "октября", "ноября", "декабря",
 }
 
+// onlyLocalTimeReply — the honest answer when the user asks the time somewhere
+// other than here. She only keeps one clock, and saying so is better than
+// naming the wrong city's time.
+//
+// There used to be a city→time-zone table here. It was removed on purpose: the
+// user only ever asks for local time, so the table was a second list of cities
+// to keep in step with the weather one for no gain.
+const onlyLocalTimeReply = "я знаю только местное время, про другие города пока не скажу."
+
+// notPlaceAfterV — words that follow "в" without naming a place, so
+// mentionsUnknownPlace does not mistake them for a city.
+var notPlaceAfterV = map[string]bool{
+	"данный": true, "данную": true, "этот": true, "эту": true,
+	"котором": true, "какое": true, "какой": true, "который": true,
+	"общем": true, "точности": true, "курсе": true, "сутках": true,
+	"часах": true, "минутах": true, "секундах": true, "неделе": true,
+}
+
+// mentionsUnknownPlace reports whether the question has a "в <слово>" phrase
+// that looks like a place we do not know ("который час в киеве"). Used only to
+// pick the honest "local time only" reply instead of answering local time as
+// if it were the city's.
+func mentionsUnknownPlace(u string) bool {
+	toks := strings.Fields(u)
+	for i := 0; i+1 < len(toks); i++ {
+		if toks[i] != "в" && toks[i] != "во" {
+			continue
+		}
+		next := strings.Trim(toks[i+1], ".,?!")
+		if next == "" || notPlaceAfterV[next] {
+			continue
+		}
+		// A number after "в" is a clock ("в 5 часов"), not a place.
+		if _, err := strconv.Atoi(strings.SplitN(next, ":", 2)[0]); err == nil {
+			continue
+		}
+		return true
+	}
+	return false
+}
+
+// onlyNearDaysReply — she can work out today, tomorrow, the day after and
+// yesterday, and nothing further. Said out loud instead of answering today's
+// date for a day she did not understand.
+const onlyNearDaysReply = "я считаю только сегодня, завтра, послезавтра и вчера — про другие дни пока не скажу."
+
+// dayWords — day references the calendar parser cannot resolve. A weekday name
+// or a "через …" phrase means he asked about a specific other day.
+var dayWords = []string{
+	"понедельник", "вторник", "сред", "четверг", "пятниц", "суббот", "воскресен",
+	"через", "monday", "tuesday", "wednesday", "thursday", "friday", "saturday", "sunday",
+}
+
+// mentionsUnknownDay reports whether the question names a day the calendar
+// parser could not resolve. Mirror of mentionsUnknownPlace: it exists only to
+// pick an honest reply over a confidently wrong one.
+//
+// Only called after ParseCalendarDate has already failed, so "завтра" and the
+// other words it does know never reach here.
+func mentionsUnknownDay(u string) bool {
+	for _, w := range dayWords {
+		if strings.Contains(u, w) {
+			return true
+		}
+	}
+	return false
+}
+
+// ruClock renders the clock part of the time reply: "15 часов 4 минуты".
+func ruClock(t time.Time) string {
+	h, m := t.Hour(), t.Minute()
+	hourWord := ruPlural(h, "час", "часа", "часов")
+	if m == 0 {
+		return fmt.Sprintf("%d %s ровно", h, hourWord)
+	}
+	return fmt.Sprintf("%d %s %d %s", h, hourWord, m, ruPlural(m, "минута", "минуты", "минут"))
+}
+
+// dayPrefix names the day relative to now ("завтра", "вчера", …) so the date
+// reply opens the way a person would say it.
+func dayPrefix(now, day time.Time) string {
+	base := time.Date(now.Year(), now.Month(), now.Day(), 0, 0, 0, 0, now.Location())
+	switch int(day.Sub(base).Hours() / 24) {
+	case -1:
+		return "вчера"
+	case 0:
+		return "сегодня"
+	case 1:
+		return "завтра"
+	case 2:
+		return "послезавтра"
+	}
+	return "это"
+}
+
 func ruPlural(n int, one, two, many string) string {
 	n = n % 100
 	if n > 10 && n < 20 {
@@ -955,18 +1111,29 @@ func (h *reactiveHandler) replySystem(ctx context.Context, dec router.Decision)
 
 	switch {
 	case strings.Contains(u, "час") || strings.Contains(u, "врем"):
-		h := now.Hour()
-		m := now.Minute()
-		hourWord := ruPlural(h, "час", "часа", "часов")
-		if m == 0 {
-			return fmt.Sprintf("сейчас %d %s ровно", h, hourWord)
+		// "который час в киеве" — she keeps one clock, so any named place gets
+		// the honest answer. Never local time dressed up as the city's.
+		if mentionsUnknownPlace(u) {
+			return onlyLocalTimeReply
 		}
-		minWord := ruPlural(m, "минута", "минуты", "минут")
-		return fmt.Sprintf("сейчас %d %s %d %s", h, hourWord, m, minWord)
+		return "сейчас " + ruClock(now)
 	case strings.Contains(u, "день") || strings.Contains(u, "числ"):
-		dow := ruWeekdays[now.Weekday()]
-		month := ruMonths[now.Month()-1]
-		return fmt.Sprintf("сегодня %s, %d %s %d года", dow, now.Day(), month, now.Year())
+		// "какое число завтра" — answer for the day the user asked about,
+		// not today. Reuses the router's calendar day-word parser.
+		day := now
+		prefix := "сегодня"
+		if d, ok := router.ParseCalendarDate(u, now); ok {
+			day = d
+			prefix = dayPrefix(now, d)
+		} else if mentionsUnknownDay(u) {
+			// He named a day she cannot work out ("в пятницу", "через неделю").
+			// Answering today's date here would be the same silent wrong answer
+			// this arm was fixed for, so say what she can do instead.
+			return onlyNearDaysReply
+		}
+		dow := ruWeekdays[day.Weekday()]
+		month := ruMonths[day.Month()-1]
+		return fmt.Sprintf("%s %s, %d %s %d года", prefix, dow, day.Day(), month, day.Year())
 	case strings.Contains(u, "кто дома") || strings.Contains(u, "человек дома"):
 		return "присутствие пока не подключено к голосовому запросу."
 	case strings.Contains(u, "памят") || strings.Contains(u, "процессор") || strings.Contains(u, "загрузк") || strings.Contains(u, "статус") || strings.Contains(u, "работа") || strings.Contains(u, "сервис") || strings.Contains(u, "диск") || strings.Contains(u, "ip") || strings.Contains(u, "аптайм") || strings.Contains(u, "трафик") || strings.Contains(u, "интернет"):
@@ -1709,3 +1876,71 @@ func jsonStringImpl(s string) string {
 	b = append(b, '"')
 	return string(b)
 }
+
+// reembedOnStart is the -reembed flag (set in run()). Opt-in on purpose: see
+// runReembed.
+var reembedOnStart bool
+
+// checkStoredEmbedder compares the embedder we just loaded with the one that
+// wrote the vectors already in the DB (Vikunja #378).
+//
+// The two models we have both make 384-dim vectors, so a size check catches
+// nothing: after a swap, recall silently compares vectors from different
+// spaces and the scores are noise. So we say it out loud. Recall itself is not
+// changed here — the fix is `mavend -reembed`.
+func checkStoredEmbedder(dataStore *store.Store, emb router.Embedder) {
+	if dataStore == nil {
+		return
+	}
+	current := router.EmbedderID(emb)
+	if reembedOnStart {
+		runReembed(dataStore, emb, current)
+		return
+	}
+	stored, mismatch, err := dataStore.CheckEmbedder(context.Background(), current)
+	if err != nil {
+		log.Printf("voice: embedder marker check failed: %v", err)
+		return
+	}
+	if mismatch {
+		log.Printf("voice: WARNING embedder MISMATCH — stored vectors were written by %q but the configured embedder is %q; recall scores are noise until the notes and facts are re-embedded — run `mavend -reembed` once (Vikunja #378)", stored, current)
+		return
+	}
+	log.Printf("voice: embedder marker ok (%s)", current)
+}
+
+// runReembed is the one-shot backfill behind -reembed.
+//
+// Why a flag and not automatic on mismatch: the embedder is ONNX on the
+// laptop's CPU, so a few thousand notes is minutes of work. Doing that silently
+// inside a normal start would look like the daemon hanging on boot. So the user
+// runs it once, deliberately, after an embedder swap; the mismatch warning
+// above tells them to. It re-embeds, logs what it did, and then the daemon
+// carries on serving as usual — no separate binary, no second start needed.
+func runReembed(dataStore *store.Store, emb router.Embedder, current string) {
+	log.Printf("voice: re-embedding stored notes and facts with %s — this can take a few minutes, do not interrupt", current)
+	res, err := dataStore.ReembedAll(context.Background(), current,
+		// EmbedPassage, not EmbedQuery: these are stored texts being searched
+		// FOR, which is the side they were written with.
+		func(ctx context.Context, text string) ([]float32, error) {
+			return router.EmbedPassage(ctx, emb, text)
+		})
+	if err != nil {
+		log.Printf("voice: re-embed FAILED, nothing was changed and no marker was written — safe to run again: %v", err)
+		return
+	}
+	if res.Skipped {
+		log.Printf("voice: re-embed skipped — the stored vectors were already written by %s", current)
+		return
+	}
+	log.Printf("voice: re-embed done — %d notes in the notes table, %d notes and %d facts in the memory index, took %s; stored vectors now belong to %s",
+		res.Notes, res.MemNotes, res.Facts, res.Took.Round(time.Second), current)
+
+	// A row with no text cannot be re-embedded, so its vector is still the old
+	// model's noise while the marker now says everything is current. Both write
+	// paths always store the text, so this should be zero — say it loudly
+	// rather than bury it in the line above if it ever isn't.
+	if res.NoText > 0 {
+		log.Printf("voice: WARNING %d stored rows had no text, so their vectors could not be re-embedded and are still noise; they will never match anything useful (Vikunja #378)", res.NoText)
+	}
+}

deploy/mavend.json

@@ -40,7 +40,10 @@
       "tokenizer_path": "/opt/maven/models/embedder/multilingual-e5-small/tokenizer.json",
       "lib_path": "/opt/maven/lib/libonnxruntime.so"
     },
-    "llm_router": false,
+    "llm_router": true,
+    "query_min_score": 0.55,
+    "query_min_margin": 0.008,
+    "clarify_max_attempts": 3,
     "tool_timeout": "30s",
     "tools": [
       { "name": "status",     "cmd": ["systemctl", "status"],         "scope": "homelab", "destructive": false },

internal/config/config.go

@@ -258,18 +258,25 @@ type VoiceConfig struct {
 	RouterThreshold float64 `json:"router_threshold,omitempty"`
 
 	// LLMRouter — route with the resident model instead of the embedding
-	// classifier. Measured on the held-out fixture (ROUTING-EVAL-31-07-2026.md)
-	// the model gets 50.0% of intents right against the classifier's 36.8%, but
-	// it costs about 800ms per turn instead of 30ms.
+	// classifier. On by default since Vikunja #320.
 	//
-	// TODO: the default stays false until this lands.
-	// Extractor.Extract never runs on an LLM decision, so acts arrive with no
-	// Fn and reminders with no Time. Turning this on today makes routing more
-	// accurate and less safe.
+	// Measured on the held-out fixture (ROUTING-EVAL-31-07-2026.md): 63.2% of
+	// intents right against the classifier's 50.0%, and no route errors. It
+	// costs about 1s per turn instead of 30ms.
 	//
-	// The router can now refuse: it answers "unknown" when it cannot route, and
-	// the turn drops to the classifier and its clarify gate (Vikunja #359).
-	LLMRouter bool `json:"llm_router,omitempty"`
+	// It is safe to leave on. The model can refuse — it answers "unknown" when
+	// it cannot route, and the turn drops to the classifier and its clarify
+	// gate. Any LLM error does the same, so a turn never breaks on the model.
+	// Slot extraction runs on LLM decisions too, so acts get their Fn and
+	// reminders their Time.
+	//
+	// Set it false to go back to the classifier, e.g. on a box with no
+	// llama-server or when 1s a turn is too slow.
+	//
+	// It is a pointer so that "missing from the file" and "explicitly false"
+	// are different things: missing means on, false means off. Read it with
+	// UseLLMRouter(), not directly.
+	LLMRouter *bool `json:"llm_router,omitempty"`
 
 	// QueryMinScore — the note-recall confidence gate. Top cosine below this
 	// ⇒ "I don't know" instead of a guess. Tuned for the ONNX embedder (0.55);
@@ -277,12 +284,31 @@ type VoiceConfig struct {
 	// default if unset.
 	QueryMinScore float64 `json:"query_min_score,omitempty"`
 
+	// QueryMinMargin — the second half of the recall gate: the top hit must
+	// beat the runner-up by more than this. The absolute score above cannot do
+	// the job on its own, because the e5 embedder puts every cosine in one
+	// narrow high band, so a made-up question scores as high as a real one.
+	// The margin asks whether one note is clearly the best instead.
+	// Negative ⇒ off. 0 ⇒ the default below.
+	QueryMinMargin float64 `json:"query_min_margin,omitempty"`
+
+	// ClarifyMaxAttempts — how many clarifying questions she may ask about one
+	// request before she gives up and says she did not understand. Default 3.
+	ClarifyMaxAttempts int `json:"clarify_max_attempts,omitempty"`
+
 	// Persona — optional prompt prefix that tunes maven's character. Prepended
 	// to every LLM system prompt (nudge phrasing, note queries, general
 	// knowledge). Empty string ⇒ current hardcoded persona (feminine-gendered
 	// Russian self-reference). Example: "Be formal and answer in English only."
 	Persona string `json:"persona,omitempty"`
 
+	// OwnerName / City — optional facts about the owner, added to the shared
+	// context block (internal/persona). Empty is fine: the block still states
+	// who he is grammatically (a man, addressed as "ты") and the current time.
+	// Nothing about correct behaviour may depend on these being filled in.
+	OwnerName string `json:"owner_name,omitempty"`
+	City      string `json:"city,omitempty"`
+
 	// Weather — the weather provider config. nil ⇒ the daemon wires
 	// the stub provider (returns ErrNotConfigured — "погода не настроена").
 	// Set provider to "open-meteo" to use the keyless Open-Meteo API.
@@ -404,7 +430,16 @@ const (
 	DefaultAutotuneInterval = 10 * time.Minute
 	DefaultRouterThreshold  = 0.55
 	DefaultQueryMinScore    = 0.55
-	DefaultToolTimeout      = 30 * time.Second
+	// Read off the margin sweep in internal/memory/recalleval on the e5
+	// embedder: 0.008 answers 68% of real questions (down from 72%) and cuts
+	// false recall from 5/5 to 1/5. Every larger delta costs real recall
+	// without removing that last one until 0.020, which drops recall to 44%.
+	DefaultQueryMinMargin = 0.008
+	// DefaultClarifyMaxAttempts — see dialogue.DefaultMaxAttempts.
+	DefaultClarifyMaxAttempts = 3
+	DefaultToolTimeout        = 30 * time.Second
+	// DefaultLLMRouter — route with the resident model unless told otherwise.
+	DefaultLLMRouter = true
 
 	DefaultFactEnrichmentInterval = 30 * time.Second
 )
@@ -486,9 +521,24 @@ func (c *Config) applyDefaults() {
 		if c.Voice.QueryMinScore <= 0 {
 			c.Voice.QueryMinScore = DefaultQueryMinScore
 		}
+		// Unset ⇒ default. Negative is how you turn the margin off on purpose,
+		// so it is clamped to 0 rather than replaced by the default.
+		switch {
+		case c.Voice.QueryMinMargin == 0:
+			c.Voice.QueryMinMargin = DefaultQueryMinMargin
+		case c.Voice.QueryMinMargin < 0:
+			c.Voice.QueryMinMargin = 0
+		}
+		if c.Voice.ClarifyMaxAttempts <= 0 {
+			c.Voice.ClarifyMaxAttempts = DefaultClarifyMaxAttempts
+		}
 		if c.Voice.ToolTimeout <= 0 {
 			c.Voice.ToolTimeout = Duration(DefaultToolTimeout)
 		}
+		if c.Voice.LLMRouter == nil {
+			on := DefaultLLMRouter
+			c.Voice.LLMRouter = &on
+		}
 	}
 
 	// routines: default severity to care-class (1) — the safe floor: a
@@ -507,6 +557,15 @@ func (c *Config) applyDefaults() {
 	}
 }
 
+// UseLLMRouter reports whether to route with the resident model. Unset means
+// on; only an explicit false in the config turns it off.
+func (v *VoiceConfig) UseLLMRouter() bool {
+	if v == nil || v.LLMRouter == nil {
+		return DefaultLLMRouter
+	}
+	return *v.LLMRouter
+}
+
 func (c *Config) validate() error {
 	if c.Phraser != nil {
 		if c.Phraser.ModelPath == "" {

internal/config/config_test.go

@@ -171,14 +171,26 @@ func TestWeatherConfigNilOK(t *testing.T) {
 	}
 }
 
-func TestLLMRouterDefaultsOff(t *testing.T) {
+func TestLLMRouterDefaultsOn(t *testing.T) {
 	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100"}}`)
 	c, err := Load(p)
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	if c.Voice.LLMRouter {
-		t.Error("voice.llm_router absent should mean false")
+	if !c.Voice.UseLLMRouter() {
+		t.Error("voice.llm_router absent should mean on")
+	}
+}
+
+// Missing and explicitly false must not mean the same thing.
+func TestLLMRouterExplicitFalseTurnsItOff(t *testing.T) {
+	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100","llm_router":false}}`)
+	c, err := Load(p)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.Voice.UseLLMRouter() {
+		t.Error("voice.llm_router false should turn it off")
 	}
 }
 
@@ -188,7 +200,7 @@ func TestLLMRouterRead(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	if !c.Voice.LLMRouter {
+	if !c.Voice.UseLLMRouter() {
 		t.Error("voice.llm_router true was not read")
 	}
 }

internal/delivery/dispatcher.go

@@ -137,9 +137,10 @@ func NewDispatcher(cfg Config) *Dispatcher {
 // picks for (severity, presence), sends via the matching sink, and records
 // one nudge row per successful send. returns the dispatches (one per channel).
 //
-// a Drop channel = no send, no record (the nudge was suppressed by routing,
-// not by a failure — "a missed water nudge is noise"). a nil sink = channel
-// not wired, skip silently. a send error stops the dispatch and returns what
+// a Drop channel = no send (the nudge was suppressed by routing, not by a
+// failure — "a missed water nudge is noise"), but it does leave a 'dropped'
+// outbox row so the suppression is visible. a nil sink = channel not wired,
+// skip silently. a send error stops the dispatch and returns what
 // got through — the daemon decides whether to retry.
 func (d *Dispatcher) DispatchNudge(ctx context.Context, pn PhrasedNudge, now time.Time) ([]Dispatch, error) {
 	c := pn.Candidate
@@ -148,6 +149,16 @@ func (d *Dispatcher) DispatchNudge(ctx context.Context, pn PhrasedNudge, now tim
 	for i := 0; i < len(channels); i++ {
 		ch := channels[i]
 		if ch == ChannelDrop {
+			// the routing table suppressed this nudge on purpose (a care nudge
+			// while you're away is noise). that stays — but it must not be
+			// invisible, or "she dropped it" and "the rule never fired" look
+			// the same afterwards. no nudges row: that table feeds the
+			// ignored_rate signal, and a nudge nobody could see must not
+			// count as ignored.
+			id := d.beginOutbox(ctx, "nudge", c.Rule.Name, 0, ch, pn.Summary, now)
+			d.completeOutbox(ctx, id, store.DeliveryDropped, now)
+			log.Printf("dispatcher: dropped %s (sev%d, presence=%s) — routing table suppressed it",
+				c.Rule.Name, c.Severity, c.State.Presence)
 			continue
 		}
 		s := Sendable{
@@ -396,6 +407,13 @@ func messageForChannel(s Sendable) string {
 	if !isAway(s.Channel) {
 		return s.Body
 	}
+	return AwayMessage(s)
+}
+
+// AwayMessage — the only text an off-box channel may ever carry. Exported so
+// the away sinks share this one rule instead of each inventing a fallback: the
+// summary if we have one, otherwise a fixed generic line. Never the body.
+func AwayMessage(s Sendable) string {
 	if s.Summary != "" {
 		return s.Summary
 	}

internal/delivery/durability_test.go

@@ -37,10 +37,12 @@ func TestVoiceNoSessionFallthroughLeavesOutboxTrail(t *testing.T) {
 			[]string{"voice", "ntfy"}, []string{store.DeliveryFailed, store.DeliverySent}},
 		{"sev4 falls through to telegram", loop.Sev4,
 			[]string{"voice", "telegram"}, []string{store.DeliveryFailed, store.DeliverySent}},
-		{"sev1 does not fall through", loop.Sev1,
-			[]string{"voice"}, []string{store.DeliveryFailed}},
-		{"sev2 does not fall through", loop.Sev2,
-			[]string{"voice"}, []string{store.DeliveryFailed}},
+		// care severities still don't reach an away channel; since #370 the
+		// drop itself is a visible row instead of nothing.
+		{"sev1 drops instead of falling through", loop.Sev1,
+			[]string{"voice", "drop"}, []string{store.DeliveryFailed, store.DeliveryDropped}},
+		{"sev2 drops instead of falling through", loop.Sev2,
+			[]string{"voice", "drop"}, []string{store.DeliveryFailed, store.DeliveryDropped}},
 	}
 	for _, c := range cases {
 		t.Run(c.name, func(t *testing.T) {

internal/delivery/ntfysink/ntfysink.go

@@ -2,10 +2,10 @@
 //
 // ntfy is the away-channel for sev3 (ops soft) nudges, sev4 (ops hard)
 // nudges when present (alongside voice), and reminders when away. the
-// message body is the Sendable's Summary — the minimal-body rule from the
+// message body is delivery.AwayMessage — the minimal-body rule from the
 // spec ("disk low on homesrv," not detail; no shoulder-surf exfil through
-// the relay). voice gets Body; away channels get Summary, enforced at the
-// sink so a phraser bug can't exfil.
+// 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
@@ -69,18 +69,14 @@ func New(cfg Config) (*Sink, error) {
 	}, nil
 }
 
-// Send publishes one notification to ntfy. the body is the Sendable's Summary
-// (minimal body); Title is "maven" (consistent sender identity on the lock
-// screen — the content is in the body). Priority maps from severity/kind so
+// Send publishes one notification to ntfy. the body is the minimal away
+// message (never the full body); Title is "maven" (consistent sender identity
+// on the lock screen — the content is in the body). Priority maps from severity/kind so
 // the phone client can ring differently for an alarm vs a soft ops nudge.
 func (s *Sink) Send(ctx context.Context, d delivery.Sendable) error {
-	body := d.Summary
-	if body == "" {
-		body = d.Body // terse full message beats no message
-	}
-	if body == "" {
-		return fmt.Errorf("ntfysink: empty message for %s", d.Channel)
-	}
+	// never fall back to d.Body: ntfy leaves the box, so an empty summary gets
+	// a generic line instead of the full detail.
+	body := delivery.AwayMessage(d)
 
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, s.topicURL(), strings.NewReader(body))
 	if err != nil {

internal/delivery/ntfysink/ntfysink_test.go

@@ -147,9 +147,9 @@ func TestSendBodyIsSummaryNotFullBody(t *testing.T) {
 	}
 }
 
-func TestSendFallsBackToBodyWhenSummaryEmpty(t *testing.T) {
-	// a terse full message is better than no message; the phraser should
-	// produce a summary for away-bound severities, but don't silently drop.
+func TestSendNeverSendsTheBodyWhenSummaryEmpty(t *testing.T) {
+	// #368: this used to fall back to the full body. ntfy leaves the box, so
+	// an empty summary gets a fixed generic line plus the rule name instead.
 	rs := newRecordingServer(t, 200, "")
 	srv := httptest.NewServer(rs.handler())
 	defer srv.Close()
@@ -160,12 +160,15 @@ func TestSendFallsBackToBodyWhenSummaryEmpty(t *testing.T) {
 		t.Fatalf("Send: %v", err)
 	}
 	_, _, body, _, _, _ := rs.snapshot()
-	if body != s.Body {
-		t.Fatalf("fallback body: want %q, got %q", s.Body, body)
+	want := delivery.GenericAwayMessage + ": service_down"
+	if body != want {
+		t.Fatalf("body: want %q, got %q", want, body)
 	}
 }
 
-func TestSendRejectsEmptyMessage(t *testing.T) {
+func TestSendNeverSendsAnEmptyMessage(t *testing.T) {
+	// with nothing at all to say we still send the generic line — an away
+	// channel can never carry detail, but it also never goes out blank.
 	rs := newRecordingServer(t, 200, "")
 	srv := httptest.NewServer(rs.handler())
 	defer srv.Close()
@@ -173,9 +176,13 @@ func TestSendRejectsEmptyMessage(t *testing.T) {
 	sink, _ := New(Config{BaseURL: srv.URL, Topic: "maven"})
 	s := nudgeSendable(loop.Sev3, "")
 	s.Body = ""
-	err := sink.Send(context.Background(), s)
-	if err == nil {
-		t.Fatal("want error for empty message")
+	s.RuleName = ""
+	if err := sink.Send(context.Background(), s); err != nil {
+		t.Fatalf("Send: %v", err)
+	}
+	_, _, body, _, _, _ := rs.snapshot()
+	if body != delivery.GenericAwayMessage {
+		t.Fatalf("body: want %q, got %q", delivery.GenericAwayMessage, body)
 	}
 }
 

internal/delivery/routing_table_test.go

@@ -208,10 +208,9 @@ func TestAwayChannelsGetMinimalBody(t *testing.T) {
 // TestCareAwayDropIsRecorded — DESIGN.md's drop is a decision ("a missed water
 // nudge is noise, a missed backup failure isn't"), so it should be visible
 // rather than vanish. Today drop is a bare `continue`: no nudge row, no outbox
-// attempt, no log — nothing an operator can see afterwards.
+// attempt, no log — nothing an operator can see afterwards. now it leaves a
+// 'dropped' outbox row.
 func TestCareAwayDropIsRecorded(t *testing.T) {
-	t.Skip("not implemented: dispatcher.go:149-151 skips a Drop channel with no record; there is no 'dropped' outcome in store/delivery.go:16-21")
-
 	ob := &fakeOutbox{}
 	d := NewDispatcher(Config{Voice: &fakeSink{}, Nudges: &fakeNudgeRecorder{}, Outbox: ob})
 

internal/delivery/telegramsink/telegramsink.go

@@ -2,11 +2,12 @@
 //
 // telegram is the away-channel for sev4 (ops hard) nudges — "disk-fire alarm
 // at 2am routes to telegram, repeat til ack." the message body is the
-// Sendable's Summary — the minimal-body rule from the spec ("disk low on
-// homesrv," not detail; no shoulder-surf exfil through the relay). voice gets
-// Body; away channels get Summary, enforced at the sink so a phraser bug can't
-// exfil. additionally, protect_content=true is passed on every send so the
-// message can't be forwarded out of the chat — locks the minimal body further.
+// delivery.AwayMessage — the minimal-body rule from the spec ("disk low on
+// homesrv," not detail; no shoulder-surf exfil through 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. additionally,
+// protect_content=true is passed on every send so the message can't be
+// forwarded out of the chat — locks the minimal body further.
 //
 // telegram's bot API is region-restricted for this homesrv — direct egress to
 // api.telegram.org is unreliable. the spec's "away channels leave the box —
@@ -140,18 +141,13 @@ type telegramResp struct {
 }
 
 // Send publishes one message to the configured telegram chat. the body is the
-// Sendable's Summary (minimal body); empty Summary falls back to Body (terse
-// full message beats no message). protect_content=true so a phraser bug (Body
-// leaking detail through Summary) can't be forwarded onward by the user or a
-// chat observer — locks the minimal-body rule at the channel's own last mile.
+// minimal away message (never the full body). protect_content=true so even
+// that can't be forwarded onward by the user or a chat observer — locks the
+// minimal-body rule at the channel's own last mile.
 func (s *Sink) Send(ctx context.Context, d delivery.Sendable) error {
-	body := d.Summary
-	if body == "" {
-		body = d.Body
-	}
-	if body == "" {
-		return fmt.Errorf("telegramsink: empty message for %s", d.Channel)
-	}
+	// never fall back to d.Body: telegram leaves the box, so an empty summary
+	// gets a generic line instead of the full detail.
+	body := delivery.AwayMessage(d)
 
 	payload := sendMessageReq{
 		ChatID:              s.cfg.ChatID,

internal/delivery/telegramsink/telegramsink_test.go

@@ -173,9 +173,9 @@ func TestSendBodyIsSummaryNotFullBody(t *testing.T) {
 	}
 }
 
-func TestSendFallsBackToBodyWhenSummaryEmpty(t *testing.T) {
-	// terse full message beats none; the phraser should produce a summary for
-	// away-bound severities, but don't silently drop.
+func TestSendNeverSendsTheBodyWhenSummaryEmpty(t *testing.T) {
+	// #368: this used to fall back to the full body. telegram leaves the box,
+	// so an empty summary gets a fixed generic line plus the rule name.
 	rs := newRecordingServer(t, 200, "")
 	srv := httptest.NewServer(rs.handler())
 	defer srv.Close()
@@ -188,12 +188,14 @@ func TestSendFallsBackToBodyWhenSummaryEmpty(t *testing.T) {
 	_, _, body, _, _ := rs.snapshot()
 	var req sendMessageReq
 	_ = json.Unmarshal([]byte(body), &req)
-	if req.Text != s.Body {
-		t.Fatalf("fallback text: want %q, got %q", s.Body, req.Text)
+	want := delivery.GenericAwayMessage + ": service_down"
+	if req.Text != want {
+		t.Fatalf("text: want %q, got %q", want, req.Text)
 	}
 }
 
-func TestSendRejectsEmptyMessage(t *testing.T) {
+func TestSendNeverSendsAnEmptyMessage(t *testing.T) {
+	// with nothing at all to say we still send the generic line.
 	rs := newRecordingServer(t, 200, "")
 	srv := httptest.NewServer(rs.handler())
 	defer srv.Close()
@@ -201,9 +203,15 @@ func TestSendRejectsEmptyMessage(t *testing.T) {
 	sink, _ := New(sinkCfg(srv.URL))
 	s := nudgeSendable(loop.Sev4, "")
 	s.Body = ""
-	err := sink.Send(context.Background(), s)
-	if err == nil {
-		t.Fatal("want error for empty message")
+	s.RuleName = ""
+	if err := sink.Send(context.Background(), s); err != nil {
+		t.Fatalf("Send: %v", err)
+	}
+	_, _, body, _, _ := rs.snapshot()
+	var req sendMessageReq
+	_ = json.Unmarshal([]byte(body), &req)
+	if req.Text != delivery.GenericAwayMessage {
+		t.Fatalf("text: want %q, got %q", delivery.GenericAwayMessage, req.Text)
 	}
 }
 

internal/dialogue/clarify.go

@@ -17,10 +17,11 @@ const (
 	SlotText  Slot = "text"  // Slots.Text
 )
 
-// MaxAttempts is 1 because Maven is not a nag (DESIGN.md § Non-goals). She asks
-// one clarifying question. If the answer still leaves the slot empty she drops
-// the request instead of asking again.
-const MaxAttempts = 1
+// DefaultMaxAttempts — how many questions she may ask about one request.
+// Three, because after three tries the likely problem is that she misheard the
+// whole request, not one slot — so another question about that slot won't help.
+// Configurable: voice.clarify_max_attempts.
+const DefaultMaxAttempts = 3
 
 // PendingQuestion is what Maven holds while she waits for an answer to an open
 // question. Unlike the yes/no confirms in cmd/mavend/voice.go, the answer here
@@ -32,7 +33,17 @@ type PendingQuestion struct {
 	Utterance string // the user's original raw words
 	Asked     time.Time
 	TTL       time.Duration
-	Attempts  int // questions already asked; capped by MaxAttempts
+	Attempts  int // questions already asked
+	// MaxAttempts caps Attempts. 0 ⇒ DefaultMaxAttempts.
+	MaxAttempts int
+}
+
+// maxAttempts is MaxAttempts with the default filled in.
+func (q *PendingQuestion) maxAttempts() int {
+	if q.MaxAttempts <= 0 {
+		return DefaultMaxAttempts
+	}
+	return q.MaxAttempts
 }
 
 func (q *PendingQuestion) IsExpired(now time.Time) bool {
@@ -41,12 +52,9 @@ func (q *PendingQuestion) IsExpired(now time.Time) bool {
 
 // CanAsk reports whether Maven may ask another question about this request.
 func (q *PendingQuestion) CanAsk() bool {
-	return q.Attempts < MaxAttempts
+	return q.Attempts < q.maxAttempts()
 }
 
-// TODO: the daemon will phrase the question text from Missing (one short ru
-// question per Slot, feminine self-reference) and speak it here.
-
 // ClarifyStore holds the parked questions. Same shape and locking as
 // SessionStore: keyed by dialogue id, expired entries dropped on read.
 type ClarifyStore struct {
@@ -67,8 +75,7 @@ func NewClarifyStore(defaultTTL time.Duration) *ClarifyStore {
 	}
 }
 
-// TODO: the daemon will Put a question here when Decision.Clarify fires, in
-// place of the flat "не разобрала" reply (cmd/mavend/voice.go).
+// Put parks a question. Called on a clarify decision (cmd/mavend/clarify.go).
 func (s *ClarifyStore) Put(id string, q *PendingQuestion) {
 	if q.TTL <= 0 {
 		q.TTL = s.defaultTTL
@@ -78,8 +85,7 @@ func (s *ClarifyStore) Put(id string, q *PendingQuestion) {
 	s.mu.Unlock()
 }
 
-// TODO: the daemon will Get on the next turn, parse that turn into Slots, call
-// Answer, and Delete — the open-question twin of resolveConfirm.
+// Get returns the live parked question, or nil when there is none.
 func (s *ClarifyStore) Get(id string, now time.Time) *PendingQuestion {
 	s.mu.RLock()
 	q, ok := s.questions[id]
@@ -94,6 +100,21 @@ func (s *ClarifyStore) Get(id string, now time.Time) *PendingQuestion {
 	return q
 }
 
+// TakeExpired reports whether a question was parked here but its TTL ran out,
+// and drops it. Get drops such a question silently, which leaves the user
+// thinking his request is still alive — the caller uses this to tell him it is
+// gone before treating his words as a fresh utterance.
+func (s *ClarifyStore) TakeExpired(id string, now time.Time) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	q, ok := s.questions[id]
+	if !ok || !q.IsExpired(now) {
+		return false
+	}
+	delete(s.questions, id)
+	return true
+}
+
 func (s *ClarifyStore) Delete(id string) {
 	s.mu.Lock()
 	delete(s.questions, id)
@@ -101,44 +122,45 @@ func (s *ClarifyStore) Delete(id string) {
 }
 
 // Answer merges the slots parsed from the user's answer into the parked ones.
-// Only the slots listed in Missing are filled, and an already filled slot is
-// never overwritten — the answer completes the original request, it does not
-// restate it. Parsing the answer text into `answer` is the caller's job; this
-// package must stay free of internal/router.
+// Only the slots listed in Missing are touched. Within those, a value the answer
+// carries WINS over what was parked: she asked about this slot, so «нет, в пять»
+// after «в три» must replace the time, not be thrown away.
+//
+// This is the clarify answer only. A correction in a fresh turn ("вообще-то
+// перенеси на пять") is a different code path (followUpMerge) — not here.
+//
+// Parsing the answer text into `answer` is the caller's job; this package must
+// stay free of internal/router.
 func (q *PendingQuestion) Answer(text string, answer Slots) Slots {
 	out := q.Slots
 	for _, slot := range q.Missing {
 		switch slot {
 		case SlotTime:
-			if !out.HasTime && answer.HasTime {
+			if answer.HasTime {
 				out.Time = answer.Time
 				out.HasTime = true
 			}
 		case SlotKey:
-			if !out.HasKey && answer.HasKey {
+			if answer.HasKey {
 				out.Key = answer.Key
 				out.HasKey = true
 			}
 		case SlotValue:
-			if out.Value == "" && answer.Value != "" {
+			if answer.Value != "" {
 				out.Value = answer.Value
 			}
 		case SlotFn:
-			if !out.HasFn && answer.HasFn {
+			if answer.HasFn {
 				out.Fn = answer.Fn
 				out.HasFn = true
-				if len(out.Args) == 0 {
-					out.Args = append([]string(nil), answer.Args...)
-				}
+				out.Args = append([]string(nil), answer.Args...)
 			}
 		case SlotText:
-			if out.Text == "" {
-				if answer.Text != "" {
-					out.Text = answer.Text
-				} else {
-					// No parse for a text slot — the raw answer IS the text.
-					out.Text = text
-				}
+			if answer.Text != "" {
+				out.Text = answer.Text
+			} else if out.Text == "" {
+				// No parse for a text slot — the raw answer IS the text.
+				out.Text = text
 			}
 		}
 	}

internal/dialogue/clarify_test.go

@@ -1,6 +1,7 @@
 package dialogue
 
 import (
+	"reflect"
 	"testing"
 	"time"
 )
@@ -59,6 +60,28 @@ func TestClarifyStoreGetPutDelete(t *testing.T) {
 	}
 }
 
+// TestClarifyStoreTakeExpired — TakeExpired reports (and drops) only a question
+// whose TTL ran out.
+func TestClarifyStoreTakeExpired(t *testing.T) {
+	s := NewClarifyStore(time.Minute)
+	if s.TakeExpired("voice", base) {
+		t.Fatal("nothing parked ⇒ nothing expired")
+	}
+	s.Put("voice", &PendingQuestion{Missing: []Slot{SlotTime}, Asked: base, TTL: time.Minute})
+	if s.TakeExpired("voice", base.Add(30*time.Second)) {
+		t.Fatal("a live question must not report as expired")
+	}
+	if s.Get("voice", base.Add(30*time.Second)) == nil {
+		t.Fatal("a live question must survive TakeExpired")
+	}
+	if !s.TakeExpired("voice", base.Add(2*time.Minute)) {
+		t.Fatal("a stale question must report as expired")
+	}
+	if s.TakeExpired("voice", base.Add(2*time.Minute)) {
+		t.Fatal("TakeExpired must drop the question, so the second call is false")
+	}
+}
+
 func TestNewClarifyStoreDefaultTTL(t *testing.T) {
 	s := NewClarifyStore(0)
 	q := &PendingQuestion{Asked: base}
@@ -89,12 +112,13 @@ func TestAnswerFillsOnlyMissingSlots(t *testing.T) {
 			want:    Slots{Text: "напомни позвонить", Time: answerTime, HasTime: true},
 		},
 		{
-			name:    "does not overwrite a filled time",
+			// He restated it: «нет, в пять». The new value wins.
+			name:    "a restated time overwrites the parked one",
 			parked:  Slots{Time: other, HasTime: true},
 			missing: []Slot{SlotTime},
-			text:    "в три",
+			text:    "нет, в три",
 			answer:  Slots{Time: answerTime, HasTime: true},
-			want:    Slots{Time: other, HasTime: true},
+			want:    Slots{Time: answerTime, HasTime: true},
 		},
 		{
 			name:    "ignores slots that were not missing",
@@ -121,12 +145,12 @@ func TestAnswerFillsOnlyMissingSlots(t *testing.T) {
 			want:    Slots{Fn: "restart", Args: []string{"nginx"}, HasFn: true},
 		},
 		{
-			name:    "keeps existing args when fn was already known",
+			name:    "a restated fn replaces the fn and its args",
 			parked:  Slots{Fn: "restart", Args: []string{"nginx"}, HasFn: true},
 			missing: []Slot{SlotFn},
 			text:    "останови postgres",
 			answer:  Slots{Fn: "stop", Args: []string{"postgres"}, HasFn: true},
-			want:    Slots{Fn: "restart", Args: []string{"nginx"}, HasFn: true},
+			want:    Slots{Fn: "stop", Args: []string{"postgres"}, HasFn: true},
 		},
 		{
 			name:    "raw answer becomes the text when nothing was parsed",
@@ -157,36 +181,34 @@ func TestAnswerFillsOnlyMissingSlots(t *testing.T) {
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
 			q := &PendingQuestion{Slots: tc.parked, Missing: tc.missing, Asked: base}
-			got := q.Answer(tc.text, tc.answer)
-			if got.Time != tc.want.Time || got.HasTime != tc.want.HasTime ||
-				got.Key != tc.want.Key || got.HasKey != tc.want.HasKey ||
-				got.Value != tc.want.Value || got.Text != tc.want.Text ||
-				got.Fn != tc.want.Fn || got.HasFn != tc.want.HasFn {
+			// Whole-struct compare: a new field in Slots is covered for free.
+			if got := q.Answer(tc.text, tc.answer); !reflect.DeepEqual(got, tc.want) {
 				t.Fatalf("Answer = %+v, want %+v", got, tc.want)
 			}
-			if len(got.Args) != len(tc.want.Args) {
-				t.Fatalf("Args = %v, want %v", got.Args, tc.want.Args)
-			}
-			for i := range got.Args {
-				if got.Args[i] != tc.want.Args[i] {
-					t.Fatalf("Args = %v, want %v", got.Args, tc.want.Args)
-				}
-			}
 		})
 	}
 }
 
-func TestCanAskCapsAtOneQuestion(t *testing.T) {
-	if MaxAttempts != 1 {
-		t.Fatalf("MaxAttempts = %d, want 1 (Maven asks once, she is not a nag)", MaxAttempts)
+func TestCanAskAllowsThreeQuestionsByDefault(t *testing.T) {
+	if DefaultMaxAttempts != 3 {
+		t.Fatalf("DefaultMaxAttempts = %d, want 3", DefaultMaxAttempts)
 	}
-	q := &PendingQuestion{Asked: base}
-	if !q.CanAsk() {
-		t.Fatal("a fresh question should be askable")
+	q := &PendingQuestion{Asked: base} // MaxAttempts unset ⇒ the default
+	for i := 0; i < 3; i++ {
+		if !q.CanAsk() {
+			t.Fatalf("question %d should be allowed", i+1)
+		}
+		q.Attempts++
 	}
-	q.Attempts = MaxAttempts
 	if q.CanAsk() {
-		t.Fatal("the question should not be asked twice")
+		t.Fatal("a fourth question must not be allowed")
+	}
+}
+
+func TestCanAskHonoursConfiguredMax(t *testing.T) {
+	q := &PendingQuestion{Asked: base, MaxAttempts: 1, Attempts: 1}
+	if q.CanAsk() {
+		t.Fatal("MaxAttempts 1 means one question only")
 	}
 }
 

internal/dialogue/session.go

@@ -1,8 +1,12 @@
 package dialogue
 
 import (
+	"context"
+	"encoding/json"
 	"sync"
 	"time"
+
+	"github.com/kami/maven/internal/store"
 )
 
 type Intent string
@@ -50,10 +54,22 @@ func (s *Session) IsExpired(now time.Time) bool {
 	return now.After(s.Timestamp.Add(s.TTL))
 }
 
+// SessionPersister — the bit of the store the session needs, as an interface
+// so tests can swap it out. Data is an opaque blob: the store never looks
+// inside, we encode the session as JSON here.
+type SessionPersister interface {
+	SaveDialogueSession(ctx context.Context, id string, data []byte, ts time.Time, ttl time.Duration) error
+	DeleteDialogueSession(ctx context.Context, id string) error
+	LoadDialogueSessions(ctx context.Context, now time.Time) ([]store.DialogueSessionRow, error)
+}
+
+// SessionStore keeps the live sessions in a map (the fast path) and mirrors
+// every write to the persister, so a daemon restart can load them back.
 type SessionStore struct {
 	mu         sync.RWMutex
 	sessions   map[string]*Session
 	defaultTTL time.Duration
+	persist    SessionPersister // may be nil: memory only (tests, no-store paths)
 }
 
 func NewSessionStore(defaultTTL time.Duration) *SessionStore {
@@ -66,6 +82,43 @@ func NewSessionStore(defaultTTL time.Duration) *SessionStore {
 	}
 }
 
+// NewPersistentSessionStore — same store, but writes also go to the DB.
+// Call Load once after this to bring back sessions from a previous run.
+func NewPersistentSessionStore(defaultTTL time.Duration, p SessionPersister) *SessionStore {
+	s := NewSessionStore(defaultTTL)
+	s.persist = p
+	return s
+}
+
+// Load — read the saved sessions back into memory. Anything past its TTL is
+// dropped (and deleted from the DB by the store), never revived.
+func (s *SessionStore) Load(ctx context.Context, now time.Time) error {
+	if s.persist == nil {
+		return nil
+	}
+	rows, err := s.persist.LoadDialogueSessions(ctx, now)
+	if err != nil {
+		return err
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, r := range rows {
+		var sess Session
+		if err := json.Unmarshal(r.Data, &sess); err != nil {
+			// A blob we can't read is not worth failing a startup over.
+			continue
+		}
+		if sess.TTL <= 0 {
+			sess.TTL = r.TTL
+		}
+		if sess.IsExpired(now) {
+			continue
+		}
+		s.sessions[r.ID] = &sess
+	}
+	return nil
+}
+
 func (s *SessionStore) Get(id string, now time.Time) *Session {
 	s.mu.RLock()
 	sess, ok := s.sessions[id]
@@ -87,12 +140,33 @@ func (s *SessionStore) Put(id string, sess *Session) {
 	s.mu.Lock()
 	s.sessions[id] = sess
 	s.mu.Unlock()
+	s.save(id, sess)
 }
 
 func (s *SessionStore) Delete(id string) {
 	s.mu.Lock()
 	delete(s.sessions, id)
 	s.mu.Unlock()
+	if s.persist != nil {
+		_ = s.persist.DeleteDialogueSession(context.Background(), id)
+	}
+}
+
+// save — mirror one session to the DB. Best effort: memory already has it, so
+// a write error costs us the restart safety net, not the current turn.
+func (s *SessionStore) save(id string, sess *Session) {
+	if s.persist == nil {
+		return
+	}
+	data, err := json.Marshal(sess)
+	if err != nil {
+		return
+	}
+	ts := sess.Timestamp
+	if ts.IsZero() {
+		ts = time.Now()
+	}
+	_ = s.persist.SaveDialogueSession(context.Background(), id, data, ts, sess.TTL)
 }
 
 func InheritSlots(prev, cur Slots) Slots {

internal/dialogue/session_persist_test.go

@@ -0,0 +1,118 @@
+package dialogue
+
+import (
+	"context"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/store"
+)
+
+// openStore — a store on disk, so a second handle can reopen the same file.
+func openStore(t *testing.T, path string) *store.Store {
+	t.Helper()
+	s, err := store.Open(context.Background(), path)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	t.Cleanup(func() { _ = s.Close() })
+	return s
+}
+
+// A session written before a restart comes back and still merges a follow-up.
+func TestSessionSurvivesRestart(t *testing.T) {
+	ctx := context.Background()
+	path := filepath.Join(t.TempDir(), "maven_test.db")
+	now := time.Now().UTC().Truncate(time.Millisecond)
+
+	first := openStore(t, path)
+	before := NewPersistentSessionStore(2*time.Minute, first)
+	before.Put("voice", &Session{
+		Intent:    IntentReminder,
+		Slots:     Slots{Text: "полить цветы", Time: now.Add(time.Hour), HasTime: true},
+		Timestamp: now,
+		TTL:       2 * time.Minute,
+	})
+	if err := first.Close(); err != nil {
+		t.Fatalf("close: %v", err)
+	}
+
+	// fresh handle, fresh in-memory map — as after a daemon restart
+	after := openStore(t, path)
+	reloaded := NewPersistentSessionStore(2*time.Minute, after)
+	if err := reloaded.Load(ctx, now.Add(10*time.Second)); err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	sess := reloaded.Get("voice", now.Add(10*time.Second))
+	if sess == nil {
+		t.Fatal("session did not survive the restart")
+	}
+	if sess.Intent != IntentReminder {
+		t.Fatalf("intent = %q, want reminder", sess.Intent)
+	}
+	// the follow-up carries no text of its own; it must inherit the old one
+	merged := InheritSlots(sess.Slots, Slots{Time: now.Add(2 * time.Hour), HasTime: true})
+	if merged.Text != "полить цветы" {
+		t.Fatalf("merged text = %q, want the earlier turn's text", merged.Text)
+	}
+	if !merged.Time.Equal(now.Add(2 * time.Hour)) {
+		t.Fatalf("merged time = %v, want the follow-up's time", merged.Time)
+	}
+}
+
+// A session past its TTL is dead: a restart must not bring it back.
+func TestExpiredSessionNotResurrected(t *testing.T) {
+	ctx := context.Background()
+	path := filepath.Join(t.TempDir(), "maven_test.db")
+	now := time.Now().UTC().Truncate(time.Millisecond)
+
+	first := openStore(t, path)
+	before := NewPersistentSessionStore(2*time.Minute, first)
+	before.Put("voice", &Session{
+		Intent:    IntentReminder,
+		Slots:     Slots{Text: "полить цветы"},
+		Timestamp: now,
+		TTL:       time.Minute,
+	})
+	if err := first.Close(); err != nil {
+		t.Fatalf("close: %v", err)
+	}
+
+	after := openStore(t, path)
+	reloaded := NewPersistentSessionStore(2*time.Minute, after)
+	later := now.Add(5 * time.Minute) // well past the 1-min TTL
+	if err := reloaded.Load(ctx, later); err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if sess := reloaded.Get("voice", later); sess != nil {
+		t.Fatalf("expired session came back: %+v", sess)
+	}
+	// and it is gone from the DB too, not just from memory
+	rows, err := after.LoadDialogueSessions(ctx, later)
+	if err != nil {
+		t.Fatalf("LoadDialogueSessions: %v", err)
+	}
+	if len(rows) != 0 {
+		t.Fatalf("expired row still in the DB: %+v", rows)
+	}
+}
+
+// Delete removes the row as well, so an ended conversation stays ended.
+func TestDeleteRemovesPersistedSession(t *testing.T) {
+	ctx := context.Background()
+	path := filepath.Join(t.TempDir(), "maven_test.db")
+	now := time.Now().UTC().Truncate(time.Millisecond)
+
+	s := openStore(t, path)
+	ss := NewPersistentSessionStore(2*time.Minute, s)
+	ss.Put("voice", &Session{Intent: IntentChat, Timestamp: now, TTL: time.Minute})
+	ss.Delete("voice")
+	rows, err := s.LoadDialogueSessions(ctx, now)
+	if err != nil {
+		t.Fatalf("LoadDialogueSessions: %v", err)
+	}
+	if len(rows) != 0 {
+		t.Fatalf("row survived Delete: %+v", rows)
+	}
+}

internal/kiwix/client.go

@@ -0,0 +1,116 @@
+// Package kiwix reads a local Kiwix server (offline Wikipedia and friends).
+//
+// Why: the resident model is a 0.8B and invents facts. Letting her read a local
+// article snippet beats letting her recall. Nothing here talks to the internet;
+// the Kiwix server is on the same box.
+//
+// This is search only. Full articles are ~100KB of HTML, far too big for a 4096
+// token context, so the unit of context is the search snippet (~500 chars).
+package kiwix
+
+import (
+	"context"
+	"encoding/xml"
+	"fmt"
+	"html"
+	"io"
+	"net/http"
+	"net/url"
+	"regexp"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Result is one search hit.
+type Result struct {
+	Title     string // article title, e.g. "Rayleigh scattering"
+	Path      string // e.g. /content/wikipedia_en_all_maxi_2026-02/Rayleigh_scattering
+	Snippet   string // plain text, tags stripped, entities decoded
+	WordCount int    // 0 if the server did not say
+}
+
+// Client is a Kiwix HTTP client. Boring on purpose: no retries, no cache.
+type Client struct {
+	base string
+	http *http.Client
+}
+
+// New makes a client for a Kiwix base URL like http://127.0.0.1:8034.
+func New(baseURL string) *Client {
+	return &Client{
+		base: strings.TrimRight(baseURL, "/"),
+		http: &http.Client{Timeout: 10 * time.Second},
+	}
+}
+
+// Search runs a keyword search in one ZIM (book) and returns up to limit hits.
+//
+// Ranking is keyword based, not semantic: "Rayleigh scattering" finds the right
+// article, "why is the sky blue" finds a TV episode. Pass keywords, not questions.
+func (c *Client) Search(ctx context.Context, pattern, book string, limit int) ([]Result, error) {
+	if limit <= 0 {
+		limit = 5
+	}
+	q := url.Values{}
+	q.Set("pattern", pattern)
+	q.Set("books.name", book)
+	q.Set("format", "xml")
+	q.Set("pageLength", strconv.Itoa(limit))
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, c.base+"/search?"+q.Encode(), nil)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("kiwix search: http %d", resp.StatusCode)
+	}
+	return ParseSearchRSS(resp.Body)
+}
+
+// rss mirrors just the bits of the RSS 2.0 reply we use.
+type rss struct {
+	Items []struct {
+		Title string `xml:"title"`
+		Link  string `xml:"link"`
+		// innerxml keeps the <b> match markers so we can strip them ourselves.
+		Description struct {
+			Inner string `xml:",innerxml"`
+		} `xml:"description"`
+		WordCount   string `xml:"wordCount"`
+	} `xml:"channel>item"`
+}
+
+var tagRE = regexp.MustCompile(`<[^>]*>`)
+
+// ParseSearchRSS turns a Kiwix search reply into results. Exported so the parser
+// is testable from a captured response, with no server running.
+func ParseSearchRSS(r io.Reader) ([]Result, error) {
+	var doc rss
+	if err := xml.NewDecoder(r).Decode(&doc); err != nil {
+		return nil, fmt.Errorf("kiwix search: bad xml: %w", err)
+	}
+	out := make([]Result, 0, len(doc.Items))
+	for _, it := range doc.Items {
+		n, _ := strconv.Atoi(strings.ReplaceAll(it.WordCount, ",", ""))
+		out = append(out, Result{
+			Title:     strings.TrimSpace(it.Title),
+			Path:      strings.TrimSpace(it.Link),
+			Snippet:   plainText(it.Description.Inner),
+			WordCount: n,
+		})
+	}
+	return out, nil
+}
+
+// plainText drops markup and decodes entities, leaving text a model can read.
+func plainText(s string) string {
+	s = tagRE.ReplaceAllString(s, "")
+	s = html.UnescapeString(s)
+	return strings.TrimSpace(strings.Join(strings.Fields(s), " "))
+}

internal/kiwix/kiwix_test.go

@@ -0,0 +1,90 @@
+package kiwix
+
+import (
+	"context"
+	"os"
+	"strings"
+	"testing"
+	"time"
+)
+
+// A real reply from the live server, trimmed to two items.
+const sampleRSS = `<?xml version="1.0" encoding="UTF-8"?>
+<rss version="2.0" xmlns:opensearch="http://a9.com/-/spec/opensearch/1.1/">
+  <channel>
+    <title>Search: Rayleigh scattering</title>
+    <opensearch:totalResults>800</opensearch:totalResults>
+    <item>
+      <title>Rayleigh scattering</title>
+      <link>/content/wikipedia_en_all_maxi_2026-02/Rayleigh_scattering</link>
+      <description><b>Rayleigh</b> scattering causes the blue color of the sky &amp; yellow colors near the Sun.[1]</description>
+      <book><title>Wikipedia</title></book>
+      <wordCount>2,818</wordCount>
+    </item>
+    <item>
+      <title>Hyper–Rayleigh scattering</title>
+      <link>/content/wikipedia_en_all_maxi_2026-02/Hyper%E2%80%93Rayleigh_scattering</link>
+      <description>...<b>Rayleigh</b> scattering" is a nonlinear optical counterpart.</description>
+      <book><title>Wikipedia</title></book>
+      <wordCount>914</wordCount>
+    </item>
+  </channel>
+</rss>`
+
+func TestParseSearchRSS(t *testing.T) {
+	got, err := ParseSearchRSS(strings.NewReader(sampleRSS))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(got) != 2 {
+		t.Fatalf("want 2 results, got %d", len(got))
+	}
+	if got[0].Title != "Rayleigh scattering" {
+		t.Errorf("title = %q", got[0].Title)
+	}
+	if got[0].Path != "/content/wikipedia_en_all_maxi_2026-02/Rayleigh_scattering" {
+		t.Errorf("path = %q", got[0].Path)
+	}
+	if got[0].WordCount != 2818 {
+		t.Errorf("wordCount = %d", got[0].WordCount)
+	}
+	want := "Rayleigh scattering causes the blue color of the sky & yellow colors near the Sun.[1]"
+	if got[0].Snippet != want {
+		t.Errorf("snippet = %q, want %q", got[0].Snippet, want)
+	}
+	if strings.Contains(got[1].Snippet, "<b>") {
+		t.Errorf("second snippet still has tags: %q", got[1].Snippet)
+	}
+}
+
+func TestParseSearchRSSBadXML(t *testing.T) {
+	if _, err := ParseSearchRSS(strings.NewReader("not xml at all")); err == nil {
+		t.Fatal("want an error on junk input")
+	}
+}
+
+// Opt-in: needs a live Kiwix server. CI has none.
+// MAVEN_KIWIX_URL=http://127.0.0.1:8034 no_proxy=127.0.0.1,localhost go test -run Retrieval -v ./internal/kiwix/
+func TestRetrievalEval(t *testing.T) {
+	base := os.Getenv("MAVEN_KIWIX_URL")
+	if base == "" {
+		t.Skip("set MAVEN_KIWIX_URL to run the retrieval eval")
+	}
+	noProxyLoopback(t)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	rep, err := RunRetrievalEval(ctx, New(base), 5)
+	if err != nil {
+		t.Fatalf("eval: %v", err)
+	}
+	// No pass bar on purpose: the number is the finding.
+	t.Log("\n" + rep.String() + rep.Detail())
+}
+
+// noProxyLoopback stops the box's SOCKS bridge from eating loopback requests.
+func noProxyLoopback(t *testing.T) {
+	t.Setenv("no_proxy", "127.0.0.1,localhost")
+	t.Setenv("NO_PROXY", "127.0.0.1,localhost")
+}

internal/kiwix/knowledge_v1.json

@@ -0,0 +1,63 @@
+{
+  "name": "kiwix-knowledge-v1",
+  "book": "wikipedia_en_all_maxi_2026-02",
+  "note": "The 9 knowledge cases from internal/phraser/eval/talk_v1.json. Queries are hand-written English keywords on purpose: Kiwix ranks by keyword, not meaning, so a natural question fails. Writing them by hand separates 'retrieval is broken' from 'the model writes bad queries'.",
+  "cases": [
+    {
+      "id": "know-sky-blue",
+      "question": "почему небо синее?",
+      "query": "Rayleigh scattering sky blue",
+      "want_titles": ["Rayleigh scattering", "Diffuse sky radiation"]
+    },
+    {
+      "id": "know-boil-egg",
+      "question": "сколько варить яйцо вкрутую?",
+      "query": "boiled egg cooking",
+      "want_titles": ["Boiled egg", "Egg as food"]
+    },
+    {
+      "id": "know-ssd-vs-hdd",
+      "question": "чем ssd отличается от hdd?",
+      "query": "solid-state drive",
+      "want_titles": ["Solid-state drive", "Hard disk drive"]
+    },
+    {
+      "id": "know-cat-purr",
+      "question": "почему кошки мурчат?",
+      "query": "cat purr",
+      "want_titles": ["Purr", "Cat communication"]
+    },
+    {
+      "id": "know-hiccups",
+      "question": "как быстро избавиться от икоты?",
+      "query": "hiccup",
+      "want_titles": ["Hiccup"]
+    },
+    {
+      "id": "know-polite-form",
+      "question": "не могли бы вы объяснить, что такое vpn?",
+      "query": "virtual private network",
+      "want_titles": ["Virtual private network"]
+    },
+    {
+      "id": "know-dont-know",
+      "question": "как зовут моего соседа снизу?",
+      "query": "name of my downstairs neighbour",
+      "want_titles": [],
+      "expect_miss": true,
+      "note": "Unanswerable by design. Retrieval SHOULD find nothing useful. Counted as a hit only when nothing relevant comes back."
+    },
+    {
+      "id": "know-water-per-day",
+      "question": "сколько воды в день надо пить?",
+      "query": "human daily water requirement drinking",
+      "want_titles": ["Drinking water", "Water", "Dehydration", "Hydration"]
+    },
+    {
+      "id": "know-thunder-delay",
+      "question": "почему гром слышно позже молнии?",
+      "query": "thunder speed of sound lightning",
+      "want_titles": ["Thunder", "Lightning"]
+    }
+  ]
+}

internal/kiwix/retrieval_eval.go

@@ -0,0 +1,135 @@
+package kiwix
+
+// This scores retrieval alone: no LLM. For each general-knowledge question we
+// hand-write English keywords and ask whether the article that would answer it
+// comes back in the top N hits. If this score is low, reading Wikipedia cannot
+// help the model no matter how good the prompt is.
+//
+// The unanswerable case (know-dont-know) is not scored. Whether the junk it
+// returns is "nothing useful" is a human judgement, so the report just prints
+// the titles and leaves the score to the 8 answerable cases.
+
+import (
+	"context"
+	_ "embed"
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+//go:embed knowledge_v1.json
+var knowledgeFixtureJSON []byte
+
+// EvalCase — one question with hand-written keywords.
+type EvalCase struct {
+	ID         string   `json:"id"`
+	Question   string   `json:"question"`
+	Query      string   `json:"query"`
+	WantTitles []string `json:"want_titles"`
+	ExpectMiss bool     `json:"expect_miss"`
+}
+
+type fixture struct {
+	Name  string     `json:"name"`
+	Book  string     `json:"book"`
+	Cases []EvalCase `json:"cases"`
+}
+
+// Outcome — what one case retrieved.
+type Outcome struct {
+	Case   EvalCase
+	Titles []string // titles of the top N hits, in rank order
+	Rank   int      // 1-based rank of the first wanted title, 0 if none
+	Err    error
+}
+
+// Hit is true when a wanted title came back.
+func (o Outcome) Hit() bool { return o.Rank > 0 }
+
+// Report — the score plus per-case detail.
+type Report struct {
+	Name     string
+	Book     string
+	TopN     int
+	Scored   int // answerable cases
+	Hits     int
+	Errors   int
+	Outcomes []Outcome
+}
+
+// Accuracy over the answerable cases.
+func (r Report) Accuracy() float64 {
+	if r.Scored == 0 {
+		return 0
+	}
+	return float64(r.Hits) / float64(r.Scored)
+}
+
+// RunRetrievalEval searches for every fixture case.
+func RunRetrievalEval(ctx context.Context, c *Client, topN int) (Report, error) {
+	var f fixture
+	if err := json.Unmarshal(knowledgeFixtureJSON, &f); err != nil {
+		return Report{}, err
+	}
+	rep := Report{Name: f.Name, Book: f.Book, TopN: topN}
+	for _, cs := range f.Cases {
+		res, err := c.Search(ctx, cs.Query, f.Book, topN)
+		o := Outcome{Case: cs, Err: err}
+		if err != nil {
+			rep.Errors++
+		}
+		for i, hit := range res {
+			o.Titles = append(o.Titles, hit.Title)
+			if o.Rank == 0 && matches(cs.WantTitles, hit.Title) {
+				o.Rank = i + 1
+			}
+		}
+		if !cs.ExpectMiss {
+			rep.Scored++
+			if o.Hit() {
+				rep.Hits++
+			}
+		}
+		rep.Outcomes = append(rep.Outcomes, o)
+	}
+	return rep, nil
+}
+
+func matches(want []string, title string) bool {
+	for _, w := range want {
+		if strings.EqualFold(strings.TrimSpace(title), w) {
+			return true
+		}
+	}
+	return false
+}
+
+// String — the headline number.
+func (r Report) String() string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "%s: %d/%d answerable questions retrieve a wanted article in top %d (%.1f%%), %d errors\n",
+		r.Name, r.Hits, r.Scored, r.TopN, 100*r.Accuracy(), r.Errors)
+	fmt.Fprintf(&b, "  book: %s\n", r.Book)
+	return b.String()
+}
+
+// Detail — per case: what was asked, what was searched, what came back.
+func (r Report) Detail() string {
+	var b strings.Builder
+	for _, o := range r.Outcomes {
+		mark := "MISS"
+		switch {
+		case o.Case.ExpectMiss:
+			mark = "n/a "
+		case o.Hit():
+			mark = fmt.Sprintf("hit@%d", o.Rank)
+		}
+		fmt.Fprintf(&b, "  %-6s %-20s q=%q\n", mark, o.Case.ID, o.Case.Query)
+		if o.Err != nil {
+			fmt.Fprintf(&b, "         error: %v\n", o.Err)
+			continue
+		}
+		fmt.Fprintf(&b, "         got: %s\n", strings.Join(o.Titles, " | "))
+	}
+	return b.String()
+}

internal/llm/modelid.go

@@ -1,4 +1,4 @@
-package eval
+package llm
 
 import (
 	"context"
@@ -6,8 +6,20 @@ import (
 	"fmt"
 	"net/http"
 	"strings"
+	"time"
 )
 
+// UnknownModel is the label to print when the server would not say what it has
+// loaded. Deliberately ugly: an honest "unknown" is fine, a plausible-looking
+// but wrong model name is the bug this whole file exists to prevent.
+const UnknownModel = "unknown-model"
+
+// llama-server is local, so never send this through a proxy: this box's
+// http_proxy answers 503 for loopback, which would look like "server won't say
+// which model it has" when the server is right there and fine.
+// A Transport with no Proxy set bypasses http_proxy entirely.
+var modelHTTP = &http.Client{Timeout: 10 * time.Second, Transport: &http.Transport{}}
+
 // ModelID asks llama-server which model it has loaded, so a scoring run can
 // label itself. Without this a bake-off between two models produces two tables
 // that look identical, and the operator has to remember which server was up.
@@ -19,7 +31,7 @@ func ModelID(ctx context.Context, base string) (string, error) {
 	if err != nil {
 		return "", err
 	}
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := modelHTTP.Do(req)
 	if err != nil {
 		return "", err
 	}
@@ -38,14 +50,21 @@ func ModelID(ctx context.Context, base string) (string, error) {
 	if len(out.Data) == 0 {
 		return "", fmt.Errorf("models: empty list")
 	}
-	return shortModelID(out.Data[0].ID), nil
+	short := shortModelID(out.Data[0].ID)
+	if short == "" {
+		// Server answered but the id field was missing or blank. Say so
+		// instead of handing back an empty label that reads as a real name.
+		return "", fmt.Errorf("models: no id in response")
+	}
+	return short, nil
 }
 
 // shortModelID trims the path and the .gguf suffix — llama-server reports the
 // file name it was started with, which is too long for a table header.
 func shortModelID(id string) string {
+	id = strings.TrimSpace(id)
 	if i := strings.LastIndexAny(id, "/\\"); i >= 0 {
 		id = id[i+1:]
 	}
-	return strings.TrimSuffix(id, ".gguf")
+	return strings.TrimSpace(strings.TrimSuffix(id, ".gguf"))
 }

internal/llm/modelid_test.go

@@ -0,0 +1,64 @@
+package llm
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+// The point of these tests: a wrong-but-plausible model label is the bug, so
+// every path that cannot learn the real name must return an error instead of a
+// guess. No llama-server needed — a stub server stands in.
+func TestModelID(t *testing.T) {
+	cases := []struct {
+		name string
+		body string
+		code int
+		want string // "" ⇒ expect an error
+	}{
+		{"full path", `{"data":[{"id":"/mnt/hdd1/llms/qwen3.5/Qwen3.5-0.8B.Q4_K_M.gguf"}]}`, 200, "Qwen3.5-0.8B.Q4_K_M"},
+		{"bare name", `{"data":[{"id":"LFM2.5-1.2B"}]}`, 200, "LFM2.5-1.2B"},
+		{"empty list", `{"data":[]}`, 200, ""},
+		{"id missing", `{"data":[{}]}`, 200, ""},
+		{"id blank", `{"data":[{"id":"  "}]}`, 200, ""},
+		{"server error", `nope`, 500, ""},
+		{"not json", `<html>`, 200, ""},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if r.URL.Path != "/v1/models" {
+					t.Errorf("asked for %s, want /v1/models", r.URL.Path)
+				}
+				w.WriteHeader(c.code)
+				_, _ = w.Write([]byte(c.body))
+			}))
+			defer srv.Close()
+
+			got, err := ModelID(context.Background(), srv.URL+"/")
+			if c.want == "" {
+				if err == nil {
+					t.Fatalf("want an error, got label %q", got)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("ModelID: %v", err)
+			}
+			if got != c.want {
+				t.Errorf("got %q, want %q", got, c.want)
+			}
+		})
+	}
+}
+
+func TestModelIDUnreachable(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {}))
+	url := srv.URL
+	srv.Close() // nothing listening now
+
+	if got, err := ModelID(context.Background(), url); err == nil {
+		t.Fatalf("want an error from a dead server, got label %q", got)
+	}
+}

internal/memory/gate.go

@@ -0,0 +1,38 @@
+package memory
+
+// Confidence gate for a recall. Two checks, both must pass before Maven says a
+// note back:
+//
+//   - minScore — an absolute cosine floor.
+//   - minMargin — the top hit must beat the runner-up by more than this.
+//
+// The margin is the one that carries the weight. The e5 embedder packs every
+// score into a narrow high band (0.79-0.89 on the recall fixture), so an
+// absolute floor cannot tell a real hit from a confident-looking miss: every
+// value under the band admits everything, every value above it answers nothing.
+// A margin asks a different question — "is this note clearly the best one, or
+// is the whole shelf equally close?" — and a made-up question has no clear best.
+//
+// With one hit and no runner-up there is nothing to compare, so only the floor
+// applies.
+
+// ConfidentScores reports whether the top score clears both gates. scores must
+// be sorted highest first. minMargin <= 0 turns the margin check off.
+func ConfidentScores(scores []float64, minScore, minMargin float64) bool {
+	if len(scores) == 0 || scores[0] < minScore {
+		return false
+	}
+	if minMargin > 0 && len(scores) > 1 && scores[0]-scores[1] <= minMargin {
+		return false
+	}
+	return true
+}
+
+// Confident is ConfidentScores for search results.
+func Confident(results []Result, minScore, minMargin float64) bool {
+	scores := make([]float64, len(results))
+	for i, r := range results {
+		scores[i] = r.Score
+	}
+	return ConfidentScores(scores, minScore, minMargin)
+}

internal/memory/gate_test.go

@@ -0,0 +1,45 @@
+package memory
+
+import "testing"
+
+func TestConfidentScores(t *testing.T) {
+	cases := []struct {
+		name      string
+		scores    []float64
+		minScore  float64
+		minMargin float64
+		want      bool
+	}{
+		{"no hits", nil, 0.55, 0.008, false},
+		{"below the floor", []float64{0.40, 0.10}, 0.55, 0.008, false},
+		{"clear winner", []float64{0.86, 0.70}, 0.55, 0.008, true},
+		{"runner-up too close", []float64{0.860, 0.858}, 0.55, 0.008, false},
+		// The rule is "beats the runner-up by MORE than delta". Not testing an
+		// exactly-equal margin: no pair of these decimals subtracts to exactly
+		// 0.008 in binary float, so such a test would pin rounding, not the rule.
+		{"margin just under delta", []float64{0.8079, 0.8}, 0.55, 0.008, false},
+		{"margin just over delta", []float64{0.8081, 0.8}, 0.55, 0.008, true},
+		// One hit: nothing to compare against, so only the floor applies.
+		{"single hit clears", []float64{0.86}, 0.55, 0.008, true},
+		{"single hit below floor", []float64{0.10}, 0.55, 0.008, false},
+		// Margin off — the old absolute-only behaviour.
+		{"margin off admits a tie", []float64{0.86, 0.86}, 0.55, 0, true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := ConfidentScores(c.scores, c.minScore, c.minMargin); got != c.want {
+				t.Errorf("got %v, want %v", got, c.want)
+			}
+		})
+	}
+}
+
+func TestConfidentReadsResultScores(t *testing.T) {
+	res := []Result{{ID: "a", Score: 0.86}, {ID: "b", Score: 0.858}}
+	if Confident(res, 0.55, 0.008) {
+		t.Error("thin margin passed the gate")
+	}
+	if !Confident(res, 0.55, 0) {
+		t.Error("margin off should fall back to the floor alone")
+	}
+}

internal/memory/recalleval/recalleval.go

@@ -177,6 +177,8 @@ type Outcome struct {
 	Tied    bool
 	TopID   string
 	TopScor float64
+	// Margin — top1 − top2. 0 when fewer than two hits came back.
+	Margin  float64
 	Reasons []string
 }
 
@@ -184,8 +186,10 @@ type Outcome struct {
 // ranks first but is silenced by query_min_score is a threshold problem, and a
 // note that never ranks first is an embedder problem. Those are different fixes.
 type Report struct {
-	Name       string
-	MinScore   float64
+	Name     string
+	MinScore float64
+	// MinMargin — how far the top hit must beat the runner-up. 0 ⇒ off.
+	MinMargin  float64
 	Total      int
 	Answerable int
 	Rank1      int
@@ -212,9 +216,14 @@ type Report struct {
 	// where the right note ranked first, and for the no-answer cases. The gap
 	// between these two distributions is what a defensible query_min_score
 	// would have to sit inside; if they overlap, no threshold separates them.
-	CorrectTop    []float64
-	NoAnswerTop   []float64
-	P50, P95, Max time.Duration
+	CorrectTop  []float64
+	NoAnswerTop []float64
+	// CorrectMargin / NoAnswerMargin — the same two groups, but top1 − top2
+	// instead of top1. This is the pair the margin gate has to separate, and
+	// unlike the absolute scores it is what the sweep reads.
+	CorrectMargin  []float64
+	NoAnswerMargin []float64
+	P50, P95, Max  time.Duration
 }
 
 // TagStat — passed/total for one slice of the fixture.
@@ -245,13 +254,14 @@ func ratio(n, d int) float64 {
 // the run on an embed or search error: an erroring case scores as a miss and is
 // counted in Errors, because "the embedder was down" and "the embedder was
 // wrong" are different numbers.
-func Score(ctx context.Context, name string, emb router.Embedder, newStore NewStore, minScore float64, f Fixture) (Report, error) {
+func Score(ctx context.Context, name string, emb router.Embedder, newStore NewStore, minScore, minMargin float64, f Fixture) (Report, error) {
 	rep := Report{
-		Name:     name,
-		MinScore: minScore,
-		Total:    len(f.Cases),
-		ByTag:    map[string]TagStat{},
-		ByLang:   map[string]TagStat{},
+		Name:      name,
+		MinScore:  minScore,
+		MinMargin: minMargin,
+		Total:     len(f.Cases),
+		ByTag:     map[string]TagStat{},
+		ByLang:    map[string]TagStat{},
 	}
 	lat := make([]time.Duration, 0, len(f.Cases))
 
@@ -261,7 +271,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 		} else {
 			rep.NoAnswer++
 		}
-		o, err := scoreCase(ctx, emb, newStore, minScore, c, f.Filler)
+		o, err := scoreCase(ctx, emb, newStore, minScore, minMargin, c, f.Filler)
 		if err != nil {
 			return Report{}, err
 		}
@@ -285,6 +295,11 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 			} else if !o.Rank1 {
 				rep.WrongTop++
 			}
+			if o.Rank1 {
+				// Margins are collected on rank, not on the gate, so the
+				// distribution does not move as the sweep changes the gate.
+				rep.CorrectMargin = append(rep.CorrectMargin, o.Margin)
+			}
 			if o.Rank1 && o.Recalled != "" {
 				rep.CorrectTop = append(rep.CorrectTop, o.TopScor)
 			}
@@ -293,6 +308,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 				rep.FalseRecall++
 			}
 			rep.NoAnswerTop = append(rep.NoAnswerTop, o.TopScor)
+			rep.NoAnswerMargin = append(rep.NoAnswerMargin, o.Margin)
 		}
 
 		if o.Pass {
@@ -307,6 +323,8 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 
 	sort.Float64s(rep.CorrectTop)
 	sort.Float64s(rep.NoAnswerTop)
+	sort.Float64s(rep.CorrectMargin)
+	sort.Float64s(rep.NoAnswerMargin)
 	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
 	rep.P50, rep.P95 = percentile(lat, 0.50), percentile(lat, 0.95)
 	if len(lat) > 0 {
@@ -318,7 +336,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 // scoreCase inserts the case's notes into a fresh store, then runs the read
 // path the daemon runs. The returned error is fatal (the harness is broken);
 // an embedder or store failure on the query lands in Outcome.Err instead.
-func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minScore float64, c Case, filler []StoredNote) (Outcome, error) {
+func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minScore, minMargin float64, c Case, filler []StoredNote) (Outcome, error) {
 	st, release, err := newStore()
 	if err != nil {
 		return Outcome{}, fmt.Errorf("%s: new store: %w", c.ID, err)
@@ -357,7 +375,10 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 
 	if len(hits) > 0 {
 		o.TopID, o.TopScor = hits[0].ID, hits[0].Score
-		o.Recalled = bestRecall(hits, minScore)
+		if len(hits) > 1 {
+			o.Margin = hits[0].Score - hits[1].Score
+		}
+		o.Recalled = bestRecall(hits, minScore, minMargin)
 	}
 	for i, h := range hits {
 		if h.ID != c.Want {
@@ -378,14 +399,14 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 	switch {
 	case !c.Answerable():
 		if o.Recalled != "" {
-			o.Reasons = append(o.Reasons, fmt.Sprintf("false recall: %q at %.3f, want silence", o.TopID, o.TopScor))
+			o.Reasons = append(o.Reasons, fmt.Sprintf("false recall: %q at %.3f (margin %.3f), want silence", o.TopID, o.TopScor, o.Margin))
 		}
 	case o.Tied:
 		o.Reasons = append(o.Reasons, fmt.Sprintf("tie at %.3f — the right note is on top only by sort order", o.TopScor))
 	case !o.Rank1:
 		o.Reasons = append(o.Reasons, fmt.Sprintf("top hit %q (%.3f), want %q%s", o.TopID, o.TopScor, c.Want, rankNote(o.Rank3)))
 	case o.Recalled == "":
-		o.Reasons = append(o.Reasons, fmt.Sprintf("right note ranked first but scored %.3f < gate %.2f — daemon says \"не знаю\"", o.TopScor, minScore))
+		o.Reasons = append(o.Reasons, fmt.Sprintf("right note ranked first at %.3f (margin %.3f) but the gate silenced it — daemon says \"не знаю\"", o.TopScor, o.Margin))
 	}
 	o.Pass = len(o.Reasons) == 0
 	return o, nil
@@ -401,8 +422,10 @@ func rankNote(inTop3 bool) string {
 // bestRecall mirrors cmd/mavend/recall.go — the gate the daemon actually
 // applies to a memory hit. Duplicated rather than imported because package main
 // is not importable; recalleval_test.go asserts the two agree in behaviour.
-func bestRecall(results []memory.Result, min float64) string {
-	if len(results) == 0 || results[0].Score < min {
+// The daemon returns the whole hit (a note and a fact are said differently);
+// the harness only scores what came back, so it keeps returning the text.
+func bestRecall(results []memory.Result, minScore, minMargin float64) string {
+	if !memory.Confident(results, minScore, minMargin) {
 		return ""
 	}
 	return results[0].Meta["text"]
@@ -437,7 +460,7 @@ func percentile(sorted []time.Duration, p float64) time.Duration {
 // the slices that name where the path is weak.
 func (r Report) String() string {
 	var b strings.Builder
-	fmt.Fprintf(&b, "%s: %d/%d cases pass (gate %.2f)\n", r.Name, r.Passed, r.Total, r.MinScore)
+	fmt.Fprintf(&b, "%s: %d/%d cases pass (gate %.2f, margin %.3f)\n", r.Name, r.Passed, r.Total, r.MinScore, r.MinMargin)
 	fmt.Fprintf(&b, "  recall@1 %.1f%% (%d/%d)  recall@3 %.1f%% (%d/%d)  answered after gate %.1f%% (%d/%d)\n",
 		100*r.Recall1(), r.Rank1, r.Answerable,
 		100*r.Recall3(), r.Rank3, r.Answerable,
@@ -448,6 +471,8 @@ func (r Report) String() string {
 		100*r.FalseRecallRate(), r.FalseRecall, r.NoAnswer)
 	fmt.Fprintf(&b, "  top-1 score, right note first: %s\n", spread(r.CorrectTop))
 	fmt.Fprintf(&b, "  top-1 score, must be silent:   %s\n", spread(r.NoAnswerTop))
+	fmt.Fprintf(&b, "  margin top1-top2, right note first: %s\n", spread(r.CorrectMargin))
+	fmt.Fprintf(&b, "  margin top1-top2, must be silent:  %s\n", spread(r.NoAnswerMargin))
 	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))

internal/memory/recalleval/recalleval_test.go

@@ -142,21 +142,40 @@ func words(s string) []string {
 // cmd/mavend/recall.go (package main is not importable). This pins the copy to
 // the original's three rules: no hits, below the gate, or no text ⇒ silence.
 func TestBestRecallMatchesDaemon(t *testing.T) {
-	if got := bestRecall(nil, 0.55); got != "" {
+	if got := bestRecall(nil, 0.55, 0); got != "" {
 		t.Errorf("no hits: got %q, want silence", got)
 	}
 	low := []memory.Result{{ID: "a", Score: 0.4, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(low, 0.55); got != "" {
+	if got := bestRecall(low, 0.55, 0); got != "" {
 		t.Errorf("below gate: got %q, want silence", got)
 	}
 	noText := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{}}}
-	if got := bestRecall(noText, 0.55); got != "" {
+	if got := bestRecall(noText, 0.55, 0); got != "" {
 		t.Errorf("no text: got %q, want silence", got)
 	}
 	ok := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(ok, 0.55); got != "чай" {
+	if got := bestRecall(ok, 0.55, 0); got != "чай" {
 		t.Errorf("above gate: got %q, want %q", got, "чай")
 	}
+	// Margin: a close runner-up means the embedder cannot tell the two apart,
+	// so Maven stays silent even though both clear the absolute floor.
+	close := []memory.Result{
+		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
+		{ID: "b", Score: 0.85, Meta: map[string]string{"text": "кофе"}},
+	}
+	if got := bestRecall(close, 0.55, 0.03); got != "" {
+		t.Errorf("thin margin: got %q, want silence", got)
+	}
+	if got := bestRecall(close, 0.55, 0); got != "чай" {
+		t.Errorf("margin off: got %q, want %q", got, "чай")
+	}
+	clear := []memory.Result{
+		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
+		{ID: "b", Score: 0.70, Meta: map[string]string{"text": "кофе"}},
+	}
+	if got := bestRecall(clear, 0.55, 0.03); got != "чай" {
+		t.Errorf("wide margin: got %q, want %q", got, "чай")
+	}
 }
 
 // TestHashRecallBaseline — the CI ratchet. HashEmbedder, so it needs no model
@@ -171,14 +190,15 @@ func TestHashRecallBaseline(t *testing.T) {
 		t.Fatalf("Load: %v", err)
 	}
 	rep, err := Score(context.Background(), "recall+hash", router.NewHashEmbedder(hashDim), InMemory,
-		config.DefaultQueryMinScore, f)
+		config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score: %v", err)
 	}
 	t.Log("\n" + rep.String() + rep.Failures())
 	t.Log("\ngate sweep:\n" + sweep(t, router.NewHashEmbedder(hashDim), f))
 
-	// 0.32 sits under the observed 0.360 recall@1.
+	// 0.32 sits under the observed 0.370 recall@1 (was 0.360 over 25 answerable
+	// cases; the two mixed note+fact cases added with #373 make it 27).
 	const floorRecall1 = 0.32
 	if rep.Recall1() < floorRecall1 {
 		t.Errorf("recall@1 %.3f below ratchet %.2f — note recall regressed", rep.Recall1(), floorRecall1)
@@ -201,11 +221,11 @@ func TestPersistentStoreScoresTheSame(t *testing.T) {
 		t.Fatalf("Load: %v", err)
 	}
 	emb := router.NewHashEmbedder(hashDim)
-	inMem, err := Score(context.Background(), "recall+hash+memory", emb, InMemory, config.DefaultQueryMinScore, f)
+	inMem, err := Score(context.Background(), "recall+hash+memory", emb, InMemory, config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score in-memory: %v", err)
 	}
-	persistent, err := Score(context.Background(), "recall+hash+sqlite", emb, sqliteStores(t), config.DefaultQueryMinScore, f)
+	persistent, err := Score(context.Background(), "recall+hash+sqlite", emb, sqliteStores(t), config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score sqlite: %v", err)
 	}
@@ -263,14 +283,16 @@ func TestONNXRecall(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	rep, err := Score(context.Background(), "recall+onnx", emb, InMemory, config.DefaultQueryMinScore, f)
+	rep, err := Score(context.Background(), "recall+onnx", emb, InMemory, config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score: %v", err)
 	}
 	t.Log("\n" + rep.String() + rep.Failures())
-	// Cached for the sweep only: the headline run above must pay the real
+	// Cached for the sweeps only: the headline run above must pay the real
 	// embedder cost so its latency numbers mean something.
-	t.Log("\ngate sweep:\n" + sweep(t, Cache(emb), f))
+	cached := Cache(emb)
+	t.Log("\ngate sweep (margin off):\n" + sweep(t, cached, f))
+	t.Log("\nmargin sweep (gate 0.55):\n" + marginSweep(t, cached, f))
 }
 
 // sweep scores the fixture at a range of gates and renders one line each. Two
@@ -281,7 +303,7 @@ func sweep(t *testing.T, emb router.Embedder, f Fixture) string {
 	t.Helper()
 	var b strings.Builder
 	for _, gate := range []float64{0.0, 0.30, 0.40, 0.50, 0.55, 0.60, 0.70, 0.80, 0.90} {
-		rep, err := Score(context.Background(), "sweep", emb, InMemory, gate, f)
+		rep, err := Score(context.Background(), "sweep", emb, InMemory, gate, 0, f)
 		if err != nil {
 			t.Fatalf("sweep at %.2f: %v", gate, err)
 		}
@@ -290,3 +312,21 @@ func sweep(t *testing.T, emb router.Embedder, f Fixture) string {
 	}
 	return b.String()
 }
+
+// marginSweep is the same idea for the margin gate (top1 − top2 > delta), with
+// the absolute gate held at its default. The absolute score cannot separate a
+// real hit from a made-up question under e5 — every score lands in one narrow
+// band — so this sweep is the one that picks a number.
+func marginSweep(t *testing.T, emb router.Embedder, f Fixture) string {
+	t.Helper()
+	var b strings.Builder
+	for _, d := range []float64{0, 0.002, 0.005, 0.008, 0.01, 0.012, 0.015, 0.02, 0.025, 0.03, 0.04, 0.05, 0.06} {
+		rep, err := Score(context.Background(), "margin sweep", emb, InMemory, config.DefaultQueryMinScore, d, f)
+		if err != nil {
+			t.Fatalf("margin sweep at %.3f: %v", d, err)
+		}
+		fmt.Fprintf(&b, "  delta %.3f: answered %d/%d (%.0f%%)  false recall %d/%d\n",
+			d, rep.Rank1-rep.Gated, rep.Answerable, 100*rep.Answered(), rep.FalseRecall, rep.NoAnswer)
+	}
+	return b.String()
+}

internal/memory/recalleval/ru_recall_v1.json

@@ -387,6 +387,32 @@
         {"id": "n2", "text": "wifi channel is 6", "kind": "note"},
         {"id": "n3", "text": "the guest network is off", "kind": "note"}
       ]
+    },
+    {
+      "id": "ru-mixed-031",
+      "lang": "ru",
+      "tags": ["mixed", "paraphrase", "hard"],
+      "note": "notes and facts in one store and the note is the answer — the daemon indexes both (Vikunja #373)",
+      "query": "куда я спрятал второй ключ от квартиры",
+      "want": "n1",
+      "notes": [
+        {"id": "n1", "text": "запасной ключ от квартиры лежит в синей коробке на полке", "kind": "note"},
+        {"id": "x1", "text": "поменял замок в двери двадцатого июня", "kind": "fact"},
+        {"id": "x2", "text": "отдал ключ соседке в мае", "kind": "fact"}
+      ]
+    },
+    {
+      "id": "ru-mixed-032",
+      "lang": "ru",
+      "tags": ["mixed", "distractor"],
+      "note": "the mirror of ru-mixed-031: the fact answers and the notes are the distractors",
+      "query": "когда я в последний раз заливал бензин",
+      "want": "x1",
+      "notes": [
+        {"id": "x1", "text": "залил полный бак в четверг вечером", "kind": "fact"},
+        {"id": "n1", "text": "на заправке у моста дешевле бензин", "kind": "note"},
+        {"id": "n2", "text": "надо поменять зимние шины", "kind": "note"}
+      ]
     }
   ]
 }

internal/persona/persona.go

@@ -0,0 +1,138 @@
+// Package persona builds the one shared context block that goes in front of
+// every LLM system prompt: who the owner is, how to address him, and what
+// time it is right now.
+//
+// Why one block and not a line pasted into each prompt: there are five
+// prompts (nudges, action replies, chat, note queries, general knowledge) and
+// the "address him as ты" rule had only reached two of them. Five copies drift.
+// One block cannot.
+//
+// The rules here are defaults in code, not config. Maven is feminine and the
+// owner is a man addressed informally — that is a hard constraint of the
+// product, so it must hold with an empty config file. Config only ADDS
+// optional facts (his name, his city).
+package persona
+
+import (
+	"fmt"
+	"strings"
+	"time"
+)
+
+// Facts — the optional, deployment-specific half of the block. All fields may
+// be empty; the block is still correct and useful without them.
+type Facts struct {
+	OwnerName string // his name, e.g. "Ками"
+	City      string // where he is, e.g. "Москва"
+	Static    string // the free-text `persona` config string, appended verbatim
+
+	// The two config-gated capabilities. They are listed only when this
+	// deployment actually has them, because a capability she names and cannot
+	// do is worse than one she never mentions.
+	Weather  bool // an open-meteo provider is configured
+	Telegram bool // a telegram bot token + chat id are configured
+	Tools    bool // at least one shell act is on the allowlist
+}
+
+var ruWeekdays = [...]string{"воскресенье", "понедельник", "вторник", "среда", "четверг", "пятница", "суббота"}
+
+var ruMonths = [...]string{
+	"января", "февраля", "марта", "апреля", "мая", "июня",
+	"июля", "августа", "сентября", "октября", "ноября", "декабря",
+}
+
+// Block renders the context block for one turn. Russian even in front of the
+// English prompts: the rules it states are Russian grammar (ты/тебя, feminine
+// verbs), and a Russian rule reads best stated in Russian.
+//
+// Keep it short. It ships on every turn to a 0.8B on laptop CPU, so every
+// line here is latency.
+func (f Facts) Block(now time.Time) string {
+	var b strings.Builder
+
+	b.WriteString("Ты — Maven, домашняя ассистентка. О себе говоришь в женском роде: \"я записала\", \"я проверила\".\n")
+
+	// The address form gets its own line. It is the thing that kept getting
+	// lost when it was buried in prose.
+	b.WriteString("ОБРАЩЕНИЕ: владелец — мужчина, всегда на \"ты\" (ты, тебя, тебе, твой) и в единственном числе (\"выпей\", \"посмотри\"). Никогда \"вы\"/\"вас\"/\"ваш\". Никогда \"он\"/\"его\" о нём — ты говоришь ему, а не о нём. Глаголы о нём — в мужском роде (\"ты забыл\").\n")
+
+	if who := f.who(); who != "" {
+		b.WriteString(who + "\n")
+	}
+
+	b.WriteString(fmt.Sprintf("Сейчас: %s, %d %s %d, %02d:%02d (местное время).\n",
+		ruWeekdays[int(now.Weekday())], now.Day(), ruMonths[int(now.Month())-1], now.Year(),
+		now.Hour(), now.Minute()))
+
+	b.WriteString("Умеешь: " + strings.Join(f.can(), "; ") +
+		". Других ДЕЙСТВИЙ не умеешь — если просят такое, скажи прямо.\n")
+
+	if s := strings.TrimSpace(f.Static); s != "" {
+		b.WriteString(s + "\n")
+	}
+	return b.String()
+}
+
+// can lists what she can really do. Every entry here is a code path that
+// exists in the daemon today:
+//   - reminders: IntentReminder → CoreAPI.CreateReminder, fired by the tick.
+//   - notes and facts: IntentNote/IntentFact write, IntentQuery reads them back.
+//   - calendar: IntentQuery answers "что у меня сегодня" from CalendarEvents.
+//   - weather / telegram / shell acts: only when configured (see Facts).
+//
+// Nothing speculative goes in this list. A capability she offers and cannot
+// perform is worse than one she never mentions.
+func (f Facts) can() []string {
+	c := []string{
+		// Talking comes first, and the closing line says "действий" rather than
+		// "ничего", because this same block sits in front of the chat and
+		// general-knowledge prompts. A flat "you can do nothing else" would
+		// tell her to refuse the exact thing those two prompts are for.
+		"разговаривать и отвечать на вопросы",
+		"ставить напоминания",
+		"записывать заметки и факты и отвечать по ним",
+		"смотреть календарь",
+	}
+	if f.Weather {
+		c = append(c, "говорить погоду")
+	}
+	if f.Telegram {
+		c = append(c, "писать в телеграм")
+	}
+	if f.Tools {
+		c = append(c, "запускать разрешённые команды на сервере")
+	}
+	return c
+}
+
+// who renders the optional name/city line, or "" when neither is configured.
+//
+// Written as labels ("Имя владельца: ..."), not as a sentence with pronouns:
+// the block's own "ты" is Maven, so "тебя зовут" would read as her name and
+// "его" would model the third-person form she must never use about him.
+func (f Facts) who() string {
+	name := strings.TrimSpace(f.OwnerName)
+	city := strings.TrimSpace(f.City)
+	switch {
+	case name != "" && city != "":
+		return "Имя владельца: " + name + ". Город: " + city + "."
+	case name != "":
+		return "Имя владельца: " + name + "."
+	case city != "":
+		return "Город: " + city + "."
+	}
+	return ""
+}
+
+// Prepend puts the block in front of a system prompt. Nil-safe: a nil renderer
+// (tests, the stub paths) returns the prompt untouched.
+func Prepend(block func() string, prompt string) string {
+	if block == nil {
+		return prompt
+	}
+	s := strings.TrimSpace(block())
+	if s == "" {
+		return prompt
+	}
+	return s + "\n\n" + prompt
+}

internal/persona/persona_test.go

@@ -0,0 +1,72 @@
+package persona
+
+import (
+	"strings"
+	"testing"
+	"time"
+)
+
+var ref = time.Date(2026, 7, 31, 14, 5, 0, 0, time.UTC)
+
+// The block must be correct with an empty config: the address form and the
+// gender rules are hard constraints, not preferences.
+func TestBlockWorksWithZeroConfig(t *testing.T) {
+	b := Facts{}.Block(ref)
+	for _, want := range []string{"женском роде", "ОБРАЩЕНИЕ", "\"ты\"", "31 июля 2026", "пятница", "14:05"} {
+		if !strings.Contains(b, want) {
+			t.Errorf("block missing %q:\n%s", want, b)
+		}
+	}
+}
+
+func TestBlockAddsOptionalFacts(t *testing.T) {
+	b := Facts{OwnerName: "Ками", City: "Москва", Static: "Будь краткой."}.Block(ref)
+	for _, want := range []string{"Ками", "Москва", "Будь краткой."} {
+		if !strings.Contains(b, want) {
+			t.Errorf("block missing %q:\n%s", want, b)
+		}
+	}
+}
+
+// The time changes between turns, so two renders must differ.
+func TestBlockRendersTimePerTurn(t *testing.T) {
+	a := Facts{}.Block(ref)
+	c := Facts{}.Block(ref.Add(time.Hour))
+	if a == c {
+		t.Errorf("block did not change with the clock:\n%s", a)
+	}
+}
+
+// She may only offer what this deployment actually has.
+func TestCapabilitiesAreConfigGated(t *testing.T) {
+	bare := Facts{}.Block(ref)
+	for _, want := range []string{"напоминания", "заметки", "календарь"} {
+		if !strings.Contains(bare, want) {
+			t.Errorf("block missing always-on capability %q:\n%s", want, bare)
+		}
+	}
+	for _, unwanted := range []string{"погоду", "телеграм", "команды"} {
+		if strings.Contains(bare, unwanted) {
+			t.Errorf("block offers unconfigured %q:\n%s", unwanted, bare)
+		}
+	}
+
+	full := Facts{Weather: true, Telegram: true, Tools: true}.Block(ref)
+	for _, want := range []string{"погоду", "телеграм", "команды"} {
+		if !strings.Contains(full, want) {
+			t.Errorf("block missing configured capability %q:\n%s", want, full)
+		}
+	}
+}
+
+func TestPrependNilIsSafe(t *testing.T) {
+	if got := Prepend(nil, "PROMPT"); got != "PROMPT" {
+		t.Errorf("Prepend(nil) = %q", got)
+	}
+	if got := Prepend(func() string { return "  " }, "PROMPT"); got != "PROMPT" {
+		t.Errorf("Prepend(blank) = %q", got)
+	}
+	if got := Prepend(func() string { return "CTX" }, "PROMPT"); got != "CTX\n\nPROMPT" {
+		t.Errorf("Prepend = %q", got)
+	}
+}

internal/phraser/context_block_test.go

@@ -0,0 +1,33 @@
+package phraser
+
+import (
+	"strings"
+	"testing"
+)
+
+// Every phrasing prompt must carry the shared context block. This is the
+// regression guard for the bug that started this: the "ты" rule reached only
+// two of the five prompts because each prompt had its own copy of the rules.
+func TestEveryPromptCarriesTheContextBlock(t *testing.T) {
+	block := func() string { return "CTXBLOCK" }
+	p := &LLMPhraser{cfg: Config{ContextBlock: block}}
+
+	prompts := map[string]string{
+		"nudge": p.systemPrompt(),
+		"query": p.querySystemPrompt(),
+		"chat":  chatSystemPrompt(block),
+	}
+	for name, got := range prompts {
+		if !strings.HasPrefix(got, "CTXBLOCK\n\n") {
+			t.Errorf("%s prompt does not start with the context block:\n%s", name, got)
+		}
+	}
+}
+
+// Without a block the prompts are unchanged — the stub and test paths pass nil.
+func TestPromptsWithoutBlockAreUnchanged(t *testing.T) {
+	p := &LLMPhraser{}
+	if p.systemPrompt() != nudgeSystem {
+		t.Errorf("nudge prompt changed with no block set")
+	}
+}

internal/phraser/eval/address_multi_test.go

@@ -0,0 +1,33 @@
+package eval
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestAddressReportsEveryBreak — the real reply from a nudge eval run broke in
+// two ways at once and the check named only the plural. Both must print: a
+// half-reported failure reads as a milder problem than it is.
+func TestAddressReportsEveryBreak(t *testing.T) {
+	body := "Смотрите на его потребление воды."
+	res := checkAddress(body)
+	if res.Pass {
+		t.Fatalf("checkAddress passed %q", body)
+	}
+	for _, want := range []string{"смотрите", "его"} {
+		if !strings.Contains(res.Detail, want) {
+			t.Errorf("detail %q does not name %q", res.Detail, want)
+		}
+	}
+}
+
+// One word repeated is one problem, so the detail must not say it twice.
+func TestAddressDeduplicates(t *testing.T) {
+	res := checkAddress("Вам стоит поесть, вам это нужно.")
+	if res.Pass {
+		t.Fatal("expected failure")
+	}
+	if n := strings.Count(res.Detail, "formal"); n != 1 {
+		t.Errorf("detail repeats the same break %d times: %q", n, res.Detail)
+	}
+}

internal/phraser/eval/address_time_test.go

@@ -0,0 +1,42 @@
+package eval
+
+import "testing"
+
+func TestAddressTimeWordDoesNotBlind(t *testing.T) {
+	// A nudge that opens with a time word must still be caught. Without the
+	// time words in the stoplist, "сегодня" was read as the third party.
+	for _, s := range []string{
+		"сегодня он не ел 11 дней",
+		"вчера он не пил воду",
+		"опять он забыл про таблетки",
+	} {
+		if r := checkAddress(s); r.Pass {
+			t.Errorf("checkAddress(%q) passed, want a third-person failure", s)
+		}
+	}
+	// Still must not fire when a third party really is named.
+	for _, s := range []string{
+		"сегодня сервис упал, он не отвечает",
+		"ты не пил воду четыре часа",
+	} {
+		if r := checkAddress(s); !r.Pass {
+			t.Errorf("checkAddress(%q) failed: %s", s, r.Detail)
+		}
+	}
+}
+
+// TestAddressVerbIsNotAnAntecedent — a nudge is mostly verbs, and a verb is
+// never who "он" refers to. This exact string passed the check before.
+func TestAddressVerbIsNotAnAntecedent(t *testing.T) {
+	s := "попробуй встать и отдохнуть — у него есть перерыв"
+	if r := checkAddress(s); r.Pass {
+		t.Errorf("checkAddress(%q) passed, want a third-person failure", s)
+	}
+	// Still missed, and this is the documented hole: "выпей воды, он не пил" has
+	// a real noun ("воды") before the pronoun, so the scan believes somebody
+	// else was named. Telling that apart needs a parser, not a suffix rule.
+	// A named third party still wins over the verbs around it.
+	if r := checkAddress("сервис упал, он не отвечает"); !r.Pass {
+		t.Errorf("checkAddress on a real third party failed: %s", r.Detail)
+	}
+}

internal/phraser/eval/checks.go

@@ -17,10 +17,18 @@ const (
 	CheckFeminine = "feminine" // her self-reference is feminine (hard constraint)
 	CheckCringe   = "cringe"   // DESIGN.md § Non-goals, "not a relationship"
 	CheckOnTopic  = "ontopic"  // says the thing the rule is about
+
+	// CheckHisGender — the other half of the persona rule: SHE is feminine, HE
+	// is male. "ты давно не отдыхала" addresses the operator as a woman.
+	CheckHisGender = "hisgender"
+
+	// CheckAddress — she talks TO him, informally, one to one. Not "вы", not
+	// "он". See the comment block above checkAddress.
+	CheckAddress = "address"
 )
 
 // CheckNames — report order.
-var CheckNames = []string{CheckMood, CheckLang, CheckLength, CheckFeminine, CheckCringe, CheckOnTopic}
+var CheckNames = []string{CheckMood, CheckLang, CheckLength, CheckFeminine, CheckHisGender, CheckAddress, CheckCringe, CheckOnTopic}
 
 // Result — one check on one message.
 type Result struct {
@@ -52,6 +60,8 @@ func RunChecks(c Case, body, mood string) []Result {
 		checkLang(body),
 		checkLength(body),
 		checkFeminine(body),
+		checkHisGender(body),
+		checkAddress(body),
 		checkCringe(body),
 		checkOnTopic(c, body),
 	}
@@ -176,6 +186,315 @@ func checkFeminine(body string) Result {
 	return Result{CheckFeminine, true, ""}
 }
 
+// --- he is male ----------------------------------------------------------
+//
+// The mirror of checkFeminine, and the failure it was written for: the model
+// wrote "ты давно не отдыхала", which addresses the operator as a woman. That
+// scored clean, because checkFeminine only ever looks at how SHE speaks about
+// herself.
+//
+// How it works: Russian past tense is gendered by suffix, -л (m) / -ла (f). So
+// this looks for feminine past-tense words in a sentence that also talks TO him
+// ("ты", "тебя", "тебе", "твой", …). A feminine verb that belongs to her ("я
+// заметила", "напомнила тебе") is skipped — that one is correct.
+//
+// Honest about the limits: this is a suffix rule, not a parser.
+//   - False positives: a feminine noun can be the subject in the same sentence
+//     ("зарядка была утром, ты её пропустил"). The guard below skips a verb whose
+//     previous word looks like a feminine noun, which helps but will not always
+//     be right.
+//   - False negatives: gender also shows up outside the past tense (short
+//     adjectives, "сама"), and none of that is checked here.
+//
+// That is acceptable for an eval check. It is a signal to read the message, not
+// a grammar verdict, and every hit prints the word it tripped on so a human can
+// disagree.
+
+// hisMarkers — words that mean the sentence is addressed to him.
+var hisMarkers = map[string]bool{
+	"ты": true, "тебя": true, "тебе": true, "тобой": true, "тобою": true,
+	"твой": true, "твоя": true, "твоё": true, "твое": true, "твои": true, "твою": true,
+}
+
+// notFeminineVerb — ordinary words ending in "-ла" that are not verbs. Small on
+// purpose: it only has to cover words a nudge might actually use.
+//
+// Words that are both a noun and a verb are deliberately NOT here. "села",
+// "мыла" and "стекла" are nouns on paper, but in a nudge they are almost always
+// verbs ("ты села", "ты мыла"), and listing them would make the check miss the
+// exact thing it is for. Missing a real hit is worse than one false alarm.
+var notFeminineVerb = map[string]bool{
+	"школа": true, "скала": true, "игла": true, "метла": true, "смола": true,
+	"дела": true, "тела": true, "масла": true, "весла": true,
+	"зола": true, "пчела": true, "числа": true,
+}
+
+// femininePast reports whether a word looks like a feminine past-tense verb:
+// "отдыхала", "поела", "выспалась".
+func femininePast(w string) bool {
+	if len([]rune(w)) < 3 || notFeminineVerb[w] {
+		return false
+	}
+	return strings.HasSuffix(w, "ла") || strings.HasSuffix(w, "лась")
+}
+
+// looksFeminineNoun — a crude guard against "зарядка была": a word right before
+// the verb that ends in "а"/"я" and is not itself a verb is probably the subject.
+func looksFeminineNoun(w string) bool {
+	if femininePast(w) || len([]rune(w)) < 3 {
+		return false
+	}
+	return strings.HasSuffix(w, "а") || strings.HasSuffix(w, "я")
+}
+
+// sentenceRE splits on sentence-ending punctuation, so a feminine verb in one
+// sentence is not blamed on a "ты" in the next.
+var sentenceRE = regexp.MustCompile(`[.!?;…]+`)
+
+func checkHisGender(body string) Result {
+	for _, sentence := range sentenceRE.Split(strings.ToLower(body), -1) {
+		words := wordRE.FindAllString(sentence, -1)
+		addressed := false
+		for _, w := range words {
+			if hisMarkers[w] {
+				addressed = true
+			}
+		}
+		if !addressed {
+			continue
+		}
+		for i, w := range words {
+			if !femininePast(w) || hersNotHis(words, i) {
+				continue
+			}
+			if i > 0 && looksFeminineNoun(prevWord(words, i)) {
+				continue
+			}
+			return Result{CheckHisGender, false,
+				fmt.Sprintf("feminine %q addressed to him — he is male", w)}
+		}
+	}
+	return Result{CheckHisGender, true, ""}
+}
+
+// hersNotHis — the verb is Maven's own if "я" comes shortly before it, or if the
+// thing she did was done to him ("напомнила тебе", "проверила за тебя").
+func hersNotHis(words []string, i int) bool {
+	for j := i - 1; j >= 0 && j >= i-3; j-- {
+		if words[j] == "я" {
+			return true
+		}
+	}
+	if i+1 < len(words) {
+		switch words[i+1] {
+		case "тебе", "тебя", "за", "тобой":
+			return true
+		}
+	}
+	return false
+}
+
+// prevWord — the word before i, skipping "не" and punctuation, so "не отдыхала"
+// still sees the subject.
+func prevWord(words []string, i int) string {
+	for j := i - 1; j >= 0; j-- {
+		w := words[j]
+		if w == "не" || w == "ни" || !unicode.Is(unicode.Cyrillic, []rune(w)[0]) {
+			continue
+		}
+		return w
+	}
+	return ""
+}
+
+// --- how she addresses him ------------------------------------------------
+//
+// Persona hard constraint: Maven speaks TO him, informally, one to one. The
+// phrasing eval produced two breaks of it, and both scored clean:
+//
+//   - "Приходите… Жду вас" — the formal plural. Correct is ты/тебя/тебе and a
+//     singular imperative ("приходи", "жду тебя").
+//   - "Он не ел 11 дней" — she talks ABOUT him, in the third person, as if
+//     reporting to somebody else. Correct is "ты не ел 11 дней".
+//
+// Like checkHisGender this is a keyword + suffix heuristic, NOT a parser. Every
+// hit prints the word it tripped on, so a false alarm is obvious at a glance and
+// can be dismissed.
+//
+// Part 1, formal address. Two signals:
+//   - the "вы" pronoun family, matched as whole words, so there is nothing to
+//     exclude — "вы" and "вас" are never anything else.
+//   - a plural verb ending: -ите/-ете/-йте/-ьте ("приходите", "выпейте",
+//     "не забудьте", "хотите"). Nouns in the prepositional case share those
+//     endings ("в интернете", "в свете"), so a word right after a preposition is
+//     skipped. That is the whole exclusion list, on purpose: a bigger one would
+//     start swallowing real imperatives.
+//
+// Part 2, third person. "он" is perfectly fine when the message really is about
+// somebody or something else ("сервис упал, он не отвечает"). The way to tell
+// them apart: a legitimate third person has an ANTECEDENT — the thing it refers
+// to was named earlier in the message. So "он" is only flagged when nothing
+// before it in the message could be that thing.
+//
+// Where this gives up, plainly:
+//   - it only looks BACKWARD. "Он не отвечает, сервис упал" names the subject
+//     after the pronoun and is flagged wrongly.
+//   - any noun earlier in the message counts as an antecedent, even when it is
+//     not one ("после обеда он не ел", "выпей воды, он не пил" — both missed).
+//     Verbs and time words no longer count, which covers the usual nudge, but a
+//     plain noun before the pronoun still blinds it. The
+//     common time words are stoplisted so the usual nudge opening does not
+//     blind it, but a message with any other noun in front still slips through.
+//     This is the check's real hole; widening it further would start flagging
+//     legitimate third-party messages, so it stops here.
+//   - a message that opens with "ты" and only later slips into "он" is missed,
+//     because "ты" itself is skipped but the words around it are not.
+//   - formal address outside these endings (short adjectives, "вашими" style
+//     forms not listed) is missed.
+
+// addressWordRE also takes Latin words, because "him"/"he" is the same break in
+// English.
+var addressWordRE = regexp.MustCompile(`[\p{Cyrillic}]+|[a-zA-Z]+|[,.;:!?…—-]`)
+
+// formalPronouns — the "вы" family. Whole-word match, so no false hits.
+var formalPronouns = map[string]bool{
+	"вы": true, "вас": true, "вам": true, "вами": true,
+	"ваш": true, "ваша": true, "ваше": true, "ваши": true,
+	"вашего": true, "вашей": true, "вашему": true, "вашим": true,
+	"вашими": true, "вашу": true,
+}
+
+// prepositions — used twice: to skip prepositional-case nouns that look like
+// plural verbs, and as words that cannot be what "он" refers to.
+var prepositions = map[string]bool{
+	"в": true, "во": true, "на": true, "о": true, "об": true, "обо": true,
+	"при": true, "по": true, "за": true, "из": true, "с": true, "со": true,
+	"к": true, "ко": true, "до": true, "от": true, "у": true, "над": true,
+	"под": true, "про": true, "без": true, "для": true, "через": true,
+}
+
+// pluralVerb reports whether a word looks like a plural/formal verb form:
+// "приходите", "выпейте", "забудьте", "хотите".
+func pluralVerb(w string) bool {
+	if len([]rune(w)) < 5 {
+		return false
+	}
+	return strings.HasSuffix(w, "ите") || strings.HasSuffix(w, "ете") ||
+		strings.HasSuffix(w, "йте") || strings.HasSuffix(w, "ьте")
+}
+
+// thirdPersonHim — pronouns that would be talking about him instead of to him.
+var thirdPersonHim = map[string]bool{
+	"он": true, "его": true, "ему": true, "него": true, "нему": true, "ним": true,
+	"he": true, "him": true, "his": true,
+}
+
+// notAnAntecedent — words that cannot be the thing "он" refers to: pronouns,
+// particles, conjunctions, adverbs of time. If only these come before "он", the
+// message never named a third party and "он" is him.
+var notAnAntecedent = map[string]bool{
+	"не": true, "ни": true, "и": true, "а": true, "но": true, "да": true,
+	"же": true, "бы": true, "ли": true, "вот": true, "уже": true,
+	"ещё": true, "еще": true, "тоже": true, "там": true, "тут": true,
+	"здесь": true, "это": true, "что": true, "как": true, "когда": true,
+	"чтобы": true, "потому": true, "сейчас": true, "потом": true,
+	// Time words. A nudge almost always opens with one ("сегодня он не ел"),
+	// and without them the very next word is read as the person being talked
+	// about, so the check misses the exact break it was written for.
+	"сегодня": true, "вчера": true, "завтра": true, "послезавтра": true,
+	"утром": true, "днём": true, "днем": true, "вечером": true, "ночью": true,
+	"опять": true, "снова": true, "весь": true, "всю": true, "целый": true,
+	"я": true, "мне": true, "меня": true, "мной": true, "мы": true, "нас": true,
+	"ты": true, "тебя": true, "тебе": true, "тобой": true,
+	"твой": true, "твоя": true, "твоё": true, "твое": true, "твои": true, "твою": true,
+}
+
+// looksVerb — a verb is never the thing "он" refers to, so it must not count as
+// an antecedent. Past tense keeps "сервис упал, он не отвечает" working off
+// "сервис"; the infinitive and imperative endings are here because a nudge is
+// mostly made of them ("попробуй встать и отдохнуть — у него есть перерыв"
+// slipped through with "попробуй" taken for the person being talked about).
+func looksVerb(w string) bool {
+	r := []rune(w)
+	if len(r) < 3 {
+		return false
+	}
+	for _, suf := range []string{
+		"л", "ла", "ло", "ли", // past tense
+		"ть", "ться", "ти", "чь", // infinitive
+		"й", "йся", "йте", // imperative
+	} {
+		if strings.HasSuffix(w, suf) {
+			return true
+		}
+	}
+	return false
+}
+
+func checkAddress(body string) Result {
+	words := addressWordRE.FindAllString(strings.ToLower(body), -1)
+
+	// Every break, not just the first. A bad reply usually breaks in more than
+	// one way at once — "Смотрите на его потребление воды" is a plural imperative
+	// AND third person about him — and reporting only the first hid the second,
+	// which made the failure look milder than it was.
+	var breaks []string
+	seen := map[string]bool{}
+	add := func(msg string) {
+		if seen[msg] {
+			return // the same word twice in one message is one problem, not two
+		}
+		seen[msg] = true
+		breaks = append(breaks, msg)
+	}
+
+	for i, w := range words {
+		if formalPronouns[w] {
+			add(fmt.Sprintf("formal %q — she says ты/тебя/тебе", w))
+		}
+		if pluralVerb(w) && !(i > 0 && prepositions[words[i-1]]) {
+			add(fmt.Sprintf("plural imperative %q — she uses the singular", w))
+		}
+	}
+
+	for i, w := range words {
+		if !thirdPersonHim[w] {
+			continue
+		}
+		named := false
+		for j := 0; j < i; j++ {
+			p := words[j]
+			if !unicode.Is(unicode.Cyrillic, []rune(p)[0]) && !isLatinWord(p) {
+				continue // punctuation
+			}
+			// pluralVerb as well as looksVerb: looksVerb knows the imperative in
+			// -й/-йте but not the -те plural ("смотрите"), so "Смотрите на его
+			// потребление воды" counted "смотрите" as the person being talked
+			// about and the "его" never printed. Third time a verb form has
+			// blinded this check — if a fourth turns up, the antecedent test
+			// wants a real morphology table, not another suffix.
+			if notAnAntecedent[p] || prepositions[p] || thirdPersonHim[p] || looksVerb(p) || pluralVerb(p) {
+				continue
+			}
+			named = true
+			break
+		}
+		if !named {
+			add(fmt.Sprintf("third person %q with nobody else named — she talks to him, not about him", w))
+		}
+	}
+
+	if len(breaks) > 0 {
+		return Result{CheckAddress, false, strings.Join(breaks, " + ")}
+	}
+	return Result{CheckAddress, true, ""}
+}
+
+func isLatinWord(w string) bool {
+	r := []rune(w)[0]
+	return (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z')
+}
+
 // --- the cringe checks ---------------------------------------------------
 //
 // "Think Jarvis without the cringe part". DESIGN.md § Non-goals: "Not a
@@ -276,12 +595,47 @@ func checkCringe(body string) Result {
 // checkOnTopic — the message must name the thing the rule is about. A nudge
 // that never mentions water leaves the operator with a chime and no action.
 func checkOnTopic(c Case, body string) Result {
+	return checkOnTopicAny(c.WantAny, body)
+}
+
+// checkOnTopicAny is the same test over a bare want-list, so the talk scorer can
+// reuse it without owning a nudge Case.
+func checkOnTopicAny(wantAny []string, body string) Result {
 	low := strings.ToLower(body)
-	for _, want := range c.WantAny {
+	for _, want := range wantAny {
 		if strings.Contains(low, strings.ToLower(want)) {
 			return Result{CheckOnTopic, true, ""}
 		}
 	}
 	return Result{CheckOnTopic, false,
-		fmt.Sprintf("mentions none of %v", c.WantAny)}
+		fmt.Sprintf("mentions none of %v", wantAny)}
+}
+
+// --- shape checks for the free-form paths --------------------------------
+//
+// The nudge checks assume one short sentence. Chat and query replies are longer
+// by design, so the only shape worth testing there is that the model produced a
+// reply at all and did not trail off. Both are failure modes the fallbacks in
+// llmphraser.go hide: a truncated or empty generation still returns nil error.
+
+const (
+	CheckNonEmpty = "nonempty" // she said something
+	CheckEllipsis = "ellipsis" // she finished the sentence
+)
+
+func checkNonEmpty(body string) Result {
+	if strings.TrimSpace(body) == "" {
+		return Result{CheckNonEmpty, false, "empty reply"}
+	}
+	return Result{CheckNonEmpty, true, ""}
+}
+
+// checkEllipsis — a reply ending in "…" or "..." is a generation that ran out of
+// tokens, not a stylistic pause. Mid-sentence ellipses are left alone.
+func checkEllipsis(body string) Result {
+	trimmed := strings.TrimRight(strings.TrimSpace(body), `"'»)`)
+	if strings.HasSuffix(trimmed, "…") || strings.HasSuffix(trimmed, "...") {
+		return Result{CheckEllipsis, false, "reply trails off in an ellipsis — likely truncated"}
+	}
+	return Result{CheckEllipsis, true, ""}
 }

internal/phraser/eval/eval.go

@@ -281,7 +281,7 @@ func (r Report) String() string {
 	fmt.Fprintf(&b, "%s: %d/%d cases pass every check (%.1f%%), %d errors\n",
 		r.Name, r.Passed, r.Total, 100*r.Accuracy(), r.Errors)
 	for _, name := range CheckNames {
-		fmt.Fprintf(&b, "  %-9s %d/%d\n", name, r.ByCheck[name], r.Total)
+		fmt.Fprintf(&b, "  %-10s %d/%d\n", name, r.ByCheck[name], r.Total)
 	}
 	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
 	fmt.Fprintf(&b, "  by rule: %s\n", renderStats(r.ByRule))

internal/phraser/eval/eval_test.go

@@ -72,10 +72,12 @@ func TestStubBaseline(t *testing.T) {
 		// ceiling ("you've been at your desk for 4 hours without a break — step
 		// away for a bit." is 76 chars but 16 words). Left failing rather than
 		// raising the ceiling to hide it.
-		CheckLength:   12,
-		CheckFeminine: 15,
-		CheckCringe:   15,
-		CheckOnTopic:  12,
+		CheckLength:    12,
+		CheckFeminine:  15,
+		CheckHisGender: 15,
+		CheckAddress:   15,
+		CheckCringe:    15,
+		CheckOnTopic:   12,
 	}
 	for name, floor := range floors {
 		if rep.ByCheck[name] < floor {
@@ -104,6 +106,14 @@ func TestChecksCatchWhatTheyClaim(t *testing.T) {
 		{"masculine predicative", "я должен сказать: попей воды.", CheckFeminine},
 		// The other direction: HE is male, so second-person masculine is right.
 		{"second person masculine ok", "ты не пил воду четыре часа.", ""},
+		// The real observed failure: she addressed him as a woman.
+		{"feminine second person", "ты давно не отдыхала — попей воды.", CheckHisGender},
+		{"feminine second person no dash", "ты пила воду четыре часа назад.", CheckHisGender},
+		// Her own feminine verb next to "ты" is correct and must not be flagged.
+		{"her feminine verb near ты", "я заметила, что ты не пил воду.", ""},
+		{"her feminine verb about him", "напомнила тебе про воду.", ""},
+		// A feminine noun subject in the same sentence is not him.
+		{"feminine noun subject ok", "зарядка была утром, ты её пропустил, попей воды.", ""},
 		{"feminine self ok", "я заметила: воды не было четыре часа.", ""},
 		{"pet name", "милый, попей воды.", CheckCringe},
 		{"emoji", "попей воды 💧", CheckCringe},
@@ -114,6 +124,10 @@ func TestChecksCatchWhatTheyClaim(t *testing.T) {
 		{"asks how he feels", "как ты себя чувствуешь? попей воды.", CheckCringe},
 		{"praise", "молодец! теперь попей воды.", CheckCringe},
 		{"off topic", "пора бы уже что-то сделать.", CheckOnTopic},
+		// The two recorded persona breaks from the phrasing eval run. Pinned as
+		// unit tests because an eval run is sampled and may not reproduce them.
+		{"formal plural", "Приходите… Жду вас", CheckAddress},
+		{"third person about him", "Он не ел 11 дней", CheckAddress},
 	}
 
 	for _, tc := range cases {
@@ -135,6 +149,36 @@ func TestChecksCatchWhatTheyClaim(t *testing.T) {
 	}
 }
 
+// TestAddressCheck — the address check on its own, so the messages that must NOT
+// trip it can be written without also having to satisfy the on-topic check.
+func TestAddressCheck(t *testing.T) {
+	bad := []string{
+		"Приходите… Жду вас",        // the recorded formal-plural break
+		"Он не ел 11 дней",          // the recorded third-person break
+		"Выпейте воды, пожалуйста.", // plural imperative on its own
+		"Ваш обед был давно.",       // formal possessive
+	}
+	for _, body := range bad {
+		if r := checkAddress(body); r.Pass {
+			t.Errorf("persona break not caught: %q", body)
+		} else {
+			t.Logf("%q -> %s", body, r.Detail)
+		}
+	}
+
+	good := []string{
+		"ты не пил воду четыре часа — попей.",   // correct informal address
+		"сервис netdata упал, он не отвечает.",  // legitimately about a third party
+		"я заметила, что зарядка была утром.",   // no address at all
+		"в интернете опять тихо, всё работает.", // "интернете" is a noun, not an imperative
+	}
+	for _, body := range good {
+		if r := checkAddress(body); !r.Pass {
+			t.Errorf("clean message flagged: %q -> %s", body, r.Detail)
+		}
+	}
+}
+
 func TestMoodCheckUsesTheEnum(t *testing.T) {
 	if r := checkMood("cheerful"); r.Pass {
 		t.Error("mood outside the enum passed")

internal/phraser/eval/llmphraser_test.go

@@ -7,6 +7,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/phraser"
 )
 
@@ -32,6 +34,7 @@ func TestLLMPhrasingBaseline(t *testing.T) {
 	// case as a phrasing error and read as "the model cannot phrase".
 	noProxyLoopback(t)
 
+	ctx := context.Background()
 	f, err := Load()
 	if err != nil {
 		t.Fatalf("Load: %v", err)
@@ -41,10 +44,25 @@ func TestLLMPhrasingBaseline(t *testing.T) {
 	// Generous: an unconstrained 0.8B can spend a minute thinking before it
 	// writes a word, and a timeout would be scored as a model failure.
 	cfg.Timeout = 5 * time.Minute
+	// The same shared context block the daemon prepends (internal/persona),
+	// with an empty config — that is the deployment we actually ship.
+	cfg.ContextBlock = func() string { return persona.Facts{}.Block(time.Now()) }
 	p := phraser.NewLLMPhraserAt(base, cfg)
 	defer p.Close()
 
-	rep, err := Score(context.Background(), "llm (0.8B, built-in persona)", p, f)
+	// Label the run with whatever gguf the server actually has loaded. It used
+	// to say "0.8B" no matter what, so two runs of two different models came
+	// out named the same and were easy to mix up when comparing.
+	model, err := llm.ModelID(ctx, base)
+	if err != nil {
+		// An unlabelled score is still a score, but say so loudly — a made-up
+		// name in a bake-off table is worse than no name.
+		t.Logf("could not read model id from %s: %v — report will say %q", base, err, llm.UnknownModel)
+		model = llm.UnknownModel
+	}
+	t.Logf("scoring model %s at %s", model, base)
+
+	rep, err := Score(ctx, "llm ("+model+", built-in persona)", p, f)
 	if err != nil {
 		t.Fatalf("Score: %v", err)
 	}

internal/phraser/eval/talk.go

@@ -0,0 +1,265 @@
+package eval
+
+// This file scores the CONVERSATIONAL paths, the ones the nudge fixture never
+// touches: chat, query-with-notes, and general knowledge. All three now carry
+// the shared persona block (internal/persona), and all three produce long
+// free-form Russian — which is exactly where a persona break (formality, third
+// person, masculine self-reference) is most likely and where, until this file,
+// nothing could see one.
+//
+// Why a second fixture instead of more nudge cases: the checks differ. A nudge
+// must be one short sentence with no question in it; a chat reply is allowed
+// 1-3 sentences and a follow-up question is a FEATURE there. Mixing them would
+// need per-case check masks, and the nudge scorer stays untouched this way.
+//
+// Why per-path reporting: a chat regression and a knowledge regression have
+// different causes (chat prompt vs router.KnowledgePrompt), and one blended
+// percentage cannot tell them apart.
+
+import (
+	"context"
+	_ "embed"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/dialogue"
+)
+
+//go:embed talk_v1.json
+var talkFixtureJSON []byte
+
+// The three phrasing paths under test. Values match the fixture's "path" field.
+const (
+	PathChat      = "chat"      // PhraseChat
+	PathQuery     = "query"     // PhraseQuery with notes
+	PathKnowledge = "knowledge" // PhraseQuery with no notes
+)
+
+// TalkPaths — report order.
+var TalkPaths = []string{PathChat, PathQuery, PathKnowledge}
+
+// TalkCheckNames — the checks that apply to a free-form reply, in report order.
+// Deliberately a subset of CheckNames: length, mood and "no questions" are nudge
+// properties and would fail a correct chat reply. These paths return no mood at
+// all, so there is nothing to check there.
+var TalkCheckNames = []string{
+	CheckNonEmpty, CheckEllipsis, CheckLang, CheckFeminine, CheckAddress, CheckOnTopic,
+}
+
+// TalkCase — one turn as the daemon would present it.
+//
+// History is flat text because that is all PhraseChat uses (it concatenates
+// turn texts into one user message); intents and slots would be dead fields.
+// Notes are what the store would have matched for a query.
+//
+// WantAny is the on-topic contract: at least one lowercased fragment must appear
+// in the reply. Fragments are stems ("пароль" → "парол") so declension does not
+// defeat them.
+type TalkCase struct {
+	ID        string   `json:"id"`
+	Path      string   `json:"path"`
+	Utterance string   `json:"utterance"`
+	History   []string `json:"history,omitempty"`
+	Notes     []string `json:"notes,omitempty"`
+	WantAny   []string `json:"want_any"`
+	Tags      []string `json:"tags,omitempty"`
+	Note      string   `json:"note,omitempty"`
+}
+
+// TalkFixture — the versioned envelope, same gating as Fixture.
+type TalkFixture struct {
+	SchemaVersion int        `json:"schema_version"`
+	Name          string     `json:"name"`
+	Notes         []string   `json:"notes"`
+	Cases         []TalkCase `json:"cases"`
+}
+
+// LoadTalk returns the embedded conversational fixture.
+func LoadTalk() (TalkFixture, error) {
+	var f TalkFixture
+	if err := json.Unmarshal(talkFixtureJSON, &f); err != nil {
+		return TalkFixture{}, fmt.Errorf("parse talk fixture: %w", err)
+	}
+	if f.SchemaVersion != SchemaVersion {
+		return TalkFixture{}, fmt.Errorf("talk fixture schema_version %d, want %d", f.SchemaVersion, SchemaVersion)
+	}
+	if len(f.Cases) == 0 {
+		return TalkFixture{}, fmt.Errorf("talk fixture has no cases")
+	}
+	return f, nil
+}
+
+// Talker — the two methods a conversational path must have to be scorable.
+// *phraser.LLMPhraser satisfies it; same trick as Nudger.
+type Talker interface {
+	PhraseChat(ctx context.Context, utterance string, history []dialogue.Turn) (string, error)
+	PhraseQuery(ctx context.Context, utterance string, notes []string) (string, error)
+}
+
+// TalkOutcome — one scored case.
+type TalkOutcome struct {
+	Case    TalkCase
+	Reply   string
+	Err     error
+	Latency time.Duration
+	Pass    bool
+	Failed  []string
+	Reasons []string
+}
+
+// TalkReport — the aggregate. ByPath is the point of this scorer.
+type TalkReport struct {
+	Name     string
+	Total    int
+	Passed   int
+	Errors   int
+	ByCheck  map[string]int
+	ByPath   map[string]TagStat
+	Outcomes []TalkOutcome
+	P50      time.Duration
+	P95      time.Duration
+	Max      time.Duration
+}
+
+// Accuracy — fraction of cases that passed every check.
+func (r TalkReport) Accuracy() float64 {
+	if r.Total == 0 {
+		return 0
+	}
+	return float64(r.Passed) / float64(r.Total)
+}
+
+// ScoreTalk runs every case through t and aggregates. A phrasing error scores as
+// a miss and is counted separately: "the model was down" and "the model wrote
+// something bad" must not be the same number.
+func ScoreTalk(ctx context.Context, name string, t Talker, f TalkFixture) (TalkReport, error) {
+	rep := TalkReport{
+		Name:    name,
+		Total:   len(f.Cases),
+		ByCheck: map[string]int{},
+		ByPath:  map[string]TagStat{},
+	}
+	for _, n := range TalkCheckNames {
+		rep.ByCheck[n] = 0
+	}
+	lat := make([]time.Duration, 0, len(f.Cases))
+
+	for _, c := range f.Cases {
+		start := time.Now()
+		reply, err := c.run(ctx, t)
+		o := TalkOutcome{Case: c, Reply: reply, Err: err, Latency: time.Since(start)}
+		lat = append(lat, o.Latency)
+
+		if err != nil {
+			rep.Errors++
+			o.Failed = append(o.Failed, "call")
+			o.Reasons = append(o.Reasons, fmt.Sprintf("phrase error: %v", err))
+		} else {
+			for _, res := range RunTalkChecks(c, reply) {
+				if res.Pass {
+					rep.ByCheck[res.Name]++
+					continue
+				}
+				o.Failed = append(o.Failed, res.Name)
+				o.Reasons = append(o.Reasons, res.Name+": "+res.Detail)
+			}
+		}
+
+		o.Pass = len(o.Failed) == 0
+		if o.Pass {
+			rep.Passed++
+		}
+		bump(rep.ByPath, c.Path, o.Pass)
+		rep.Outcomes = append(rep.Outcomes, o)
+	}
+
+	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
+	rep.P50, rep.P95 = percentile(lat, 0.50), percentile(lat, 0.95)
+	if len(lat) > 0 {
+		rep.Max = lat[len(lat)-1]
+	}
+	return rep, nil
+}
+
+// run dispatches the case to its path. knowledge and query are the same method;
+// the empty notes slice is what selects the no-notes branch inside PhraseQuery.
+func (c TalkCase) run(ctx context.Context, t Talker) (string, error) {
+	switch c.Path {
+	case PathChat:
+		return t.PhraseChat(ctx, c.Utterance, c.turns())
+	case PathQuery:
+		return t.PhraseQuery(ctx, c.Utterance, c.Notes)
+	case PathKnowledge:
+		return t.PhraseQuery(ctx, c.Utterance, nil)
+	}
+	return "", fmt.Errorf("unknown path %q", c.Path)
+}
+
+func (c TalkCase) turns() []dialogue.Turn {
+	turns := make([]dialogue.Turn, 0, len(c.History))
+	for _, h := range c.History {
+		turns = append(turns, dialogue.Turn{Text: h})
+	}
+	return turns
+}
+
+// RunTalkChecks scores one reply. Order matches TalkCheckNames.
+func RunTalkChecks(c TalkCase, reply string) []Result {
+	return []Result{
+		checkNonEmpty(reply),
+		checkEllipsis(reply),
+		checkLang(reply),
+		checkFeminine(reply),
+		checkAddress(reply),
+		checkOnTopicAny(c.WantAny, reply),
+	}
+}
+
+// String renders the comparison table — composite, then per-check so a
+// regression names the property, then per-path so it names the prompt.
+func (r TalkReport) String() string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "%s: %d/%d cases pass every check (%.1f%%), %d errors\n",
+		r.Name, r.Passed, r.Total, 100*r.Accuracy(), r.Errors)
+	for _, name := range TalkCheckNames {
+		fmt.Fprintf(&b, "  %-10s %d/%d\n", name, r.ByCheck[name], r.Total)
+	}
+	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
+	fmt.Fprintf(&b, "  by path: %s\n", renderStats(r.ByPath))
+	return b.String()
+}
+
+// Failures — per-case detail, sorted by ID so two runs diff cleanly.
+func (r TalkReport) Failures() string {
+	var b strings.Builder
+	for _, o := range r.sorted() {
+		if o.Pass {
+			continue
+		}
+		fmt.Fprintf(&b, "  %s %q\n    %s\n", o.Case.ID, o.Reply, strings.Join(o.Reasons, "; "))
+	}
+	return b.String()
+}
+
+// Replies — every generated reply verbatim. This is what a human reads to judge
+// tone; the score only says which checks fired.
+func (r TalkReport) Replies() string {
+	var b strings.Builder
+	for _, o := range r.sorted() {
+		mark := "ok  "
+		if !o.Pass {
+			mark = "FAIL"
+		}
+		fmt.Fprintf(&b, "  %s %-9s %-22s %q\n", mark, o.Case.Path, o.Case.ID, o.Reply)
+	}
+	return b.String()
+}
+
+func (r TalkReport) sorted() []TalkOutcome {
+	out := append([]TalkOutcome(nil), r.Outcomes...)
+	sort.Slice(out, func(i, j int) bool { return out[i].Case.ID < out[j].Case.ID })
+	return out
+}

internal/phraser/eval/talk_test.go

@@ -0,0 +1,163 @@
+package eval
+
+import (
+	"context"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/dialogue"
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/persona"
+	"github.com/kami/maven/internal/phraser"
+)
+
+// perPathMinimum — the resolution floor. A per-path score built on a handful of
+// cases moves by 12% when a single reply changes, which cannot distinguish a
+// prompt regression from noise.
+const perPathMinimum = 8
+
+// TestTalkFixture — the fixture itself has to be sound before any score off it
+// means anything.
+func TestTalkFixture(t *testing.T) {
+	f, err := LoadTalk()
+	if err != nil {
+		t.Fatalf("LoadTalk: %v", err)
+	}
+
+	seen := map[string]bool{}
+	byPath := map[string]int{}
+	for _, c := range f.Cases {
+		if seen[c.ID] {
+			t.Errorf("duplicate case id %q", c.ID)
+		}
+		seen[c.ID] = true
+
+		switch c.Path {
+		case PathChat, PathQuery, PathKnowledge:
+		default:
+			t.Errorf("%s: unknown path %q", c.ID, c.Path)
+		}
+		byPath[c.Path]++
+
+		if strings.TrimSpace(c.Utterance) == "" {
+			t.Errorf("%s: empty utterance", c.ID)
+		}
+		if len(c.WantAny) == 0 {
+			t.Errorf("%s: no want_any — the reply cannot be checked for topic", c.ID)
+		}
+		// A query case with no notes would silently score the knowledge path.
+		if c.Path == PathQuery && len(c.Notes) == 0 {
+			t.Errorf("%s: query case has no notes", c.ID)
+		}
+		if c.Path == PathKnowledge && len(c.Notes) > 0 {
+			t.Errorf("%s: knowledge case must have no notes", c.ID)
+		}
+	}
+
+	for _, p := range TalkPaths {
+		if byPath[p] < perPathMinimum {
+			t.Errorf("path %s has %d cases, want at least %d", p, byPath[p], perPathMinimum)
+		}
+	}
+}
+
+// fakeTalker — a scripted Talker, so the scorer is testable without a model.
+type fakeTalker struct{ reply string }
+
+func (f fakeTalker) PhraseChat(context.Context, string, []dialogue.Turn) (string, error) {
+	return f.reply, nil
+}
+func (f fakeTalker) PhraseQuery(context.Context, string, []string) (string, error) {
+	return f.reply, nil
+}
+
+// TestScoreTalkCounts — a reply that fails on purpose must be counted on every
+// path, so a real run cannot report a hidden zero.
+func TestScoreTalkCounts(t *testing.T) {
+	f, err := LoadTalk()
+	if err != nil {
+		t.Fatalf("LoadTalk: %v", err)
+	}
+	// Formal address, off-topic, trailing ellipsis: three checks fail at once.
+	rep, err := ScoreTalk(context.Background(), "fake", fakeTalker{"Приходите, я вас жду…"}, f)
+	if err != nil {
+		t.Fatalf("ScoreTalk: %v", err)
+	}
+	if rep.Total != len(f.Cases) || rep.Passed != 0 {
+		t.Errorf("got %d/%d passing, want 0/%d", rep.Passed, rep.Total, len(f.Cases))
+	}
+	if rep.ByCheck[CheckAddress] != 0 {
+		t.Errorf("formal reply passed the address check %d times", rep.ByCheck[CheckAddress])
+	}
+	if rep.ByCheck[CheckEllipsis] != 0 {
+		t.Errorf("truncated reply passed the ellipsis check %d times", rep.ByCheck[CheckEllipsis])
+	}
+	for _, p := range TalkPaths {
+		if rep.ByPath[p].Total == 0 {
+			t.Errorf("path %s missing from the report", p)
+		}
+	}
+	if !strings.Contains(rep.String(), "by path") {
+		t.Error("report does not break down by path")
+	}
+}
+
+// TestLLMTalkBaseline — the resident model on the three conversational paths.
+// Opt-in exactly like TestLLMPhrasingBaseline: CI has no model and a run costs
+// minutes on the CPU target.
+//
+//	MAVEN_LLM_URL=http://127.0.0.1:18099 \
+//	  go test -run TestLLMTalkBaseline ./internal/phraser/eval/
+//
+// Reports, does not assert a quality bar — the numbers are the input to tuning
+// the persona prompt. The one thing worth failing on is a harness fault.
+func TestLLMTalkBaseline(t *testing.T) {
+	base := os.Getenv("MAVEN_LLM_URL")
+	if base == "" {
+		t.Skip("MAVEN_LLM_URL unset — point it at a running llama-server (see doc comment)")
+	}
+	noProxyLoopback(t)
+
+	ctx := context.Background()
+	f, err := LoadTalk()
+	if err != nil {
+		t.Fatalf("LoadTalk: %v", err)
+	}
+
+	cfg := phraser.DefaultConfig("")
+	cfg.Timeout = 5 * time.Minute
+	cfg.ContextBlock = func() string { return persona.Facts{}.Block(time.Now()) }
+	p := phraser.NewLLMPhraserAt(base, cfg)
+	defer p.Close()
+
+	// Unreachable server is fatal here, not a logged warning, and that differs
+	// from the nudge test on purpose. PhraseNudge returns its errors, so a dead
+	// server there shows up honestly in the Errors column. PhraseChat and
+	// PhraseQuery do NOT: they swallow every failure and return a canned string
+	// ("поговорили.", "не знаю.", "вот что я нашла: …"). So on these three paths
+	// a dead server produces a full report with 0 errors and a terrible score —
+	// a number that looks like bad phrasing and is really no phrasing at all.
+	// Refusing to score without a confirmed model is the only guard available
+	// until the phraser reports its failures (Vikunja #397).
+	model, err := llm.ModelID(ctx, base)
+	if err != nil {
+		t.Fatalf("no model at %s: %v — refusing to score, these paths hide their errors "+
+			"and would report a plausible-looking result off a dead server", base, err)
+	}
+	t.Logf("scoring model %s at %s", model, base)
+
+	rep, err := ScoreTalk(ctx, "llm ("+model+", built-in persona)", p, f)
+	if err != nil {
+		t.Fatalf("ScoreTalk: %v", err)
+	}
+	t.Log("\n" + rep.String() + "\nreplies:\n" + rep.Replies() + "\nfailures:\n" + rep.Failures())
+
+	// And again afterwards: the run takes minutes, and a server that died or got
+	// OOM-killed halfway through would leave the first cases scored and the rest
+	// silently canned. Checking only at the start would not catch that.
+	if _, err := llm.ModelID(ctx, base); err != nil {
+		t.Fatalf("model at %s went away during the run: %v — the score above is not trustworthy", base, err)
+	}
+}

internal/phraser/eval/talk_v1.json

@@ -0,0 +1,227 @@
+{
+  "schema_version": 1,
+  "name": "ru-talk-v1",
+  "notes": [
+    "Scores the three conversational phrasing paths: chat (PhraseChat), query (PhraseQuery with notes) and knowledge (PhraseQuery with no notes). The nudge fixture does not cover any of them.",
+    "Nine cases per path, not five. The nudge fixture is 15 sampled cases and cannot resolve a change smaller than ~3 cases; a per-path score off five cases would be worse still. More cases per path is the point of this fixture.",
+    "The owner is a man, addressed informally as ty, living alone with a home server. Every utterance is written the way he actually talks to her.",
+    "chat-formality-bait and chat-about-me exist to provoke the two persona breaks the nudge eval caught: the formal vy/vas plural, and talking about him in the third person.",
+    "want_any fragments are stems so Russian declension does not defeat the on-topic check. They are lowercased before comparison.",
+    "want_any is a plain substring test, so a fragment that is too short passes by accident: \"ты\" matches inside \"работы\", \"нет\" inside \"интернет\". Keep every fragment to three or more letters of a real stem.",
+    "Notes are written as the store would have them: short, first person, no punctuation discipline."
+  ],
+  "cases": [
+    {
+      "id": "chat-how-are-you",
+      "path": "chat",
+      "utterance": "привет, как дела?",
+      "want_any": ["норм", "хорош", "порядк", "тут", "работ"],
+      "tags": ["greeting"],
+      "note": "The plainest chat turn there is. If the persona breaks anywhere it breaks here first."
+    },
+    {
+      "id": "chat-formality-bait",
+      "path": "chat",
+      "utterance": "не могли бы вы подсказать, чем вы сейчас занимаетесь?",
+      "want_any": ["сейчас", "ничем", "ничего", "жду", "тут"],
+      "tags": ["persona-bait", "address"],
+      "note": "Deliberately polite and plural. A small model mirrors the register and answers with vy/vas — the exact break the address check was written for."
+    },
+    {
+      "id": "chat-about-me",
+      "path": "chat",
+      "utterance": "расскажи обо мне",
+      "want_any": ["теб"],
+      "tags": ["persona-bait", "third-person"],
+      "note": "Baits the third person: she should say 'ты живёшь один', not 'он живёт один', as if reporting to somebody else."
+    },
+    {
+      "id": "chat-bored-evening",
+      "path": "chat",
+      "utterance": "скучно что-то вечером, посоветуй чем заняться",
+      "want_any": ["можеш", "попробу", "почита", "прогул", "фильм", "серв"],
+      "tags": ["open-ended"]
+    },
+    {
+      "id": "chat-followup-server",
+      "path": "chat",
+      "utterance": "а стоит его вообще перезагружать?",
+      "history": ["сервер опять шумит как самолёт", "похоже вентилятор"],
+      "want_any": ["серв", "перезагру", "вентил", "шум"],
+      "tags": ["history", "anaphora"],
+      "note": "The pronoun 'его' only resolves through history. Also the one case where 'он' about the server is legitimate."
+    },
+    {
+      "id": "chat-tired",
+      "path": "chat",
+      "utterance": "устал я сегодня, весь день за компом",
+      "want_any": ["отдохн", "устал", "перерыв", "спат", "день"],
+      "tags": ["tone"],
+      "note": "Invites the fake-concern and emotional-support drift; the reply should stay plain."
+    },
+    {
+      "id": "chat-thanks",
+      "path": "chat",
+      "utterance": "спасибо, выручила",
+      "want_any": ["пожалуйст", "не за что", "рада", "обращ"],
+      "tags": ["persona", "feminine"],
+      "note": "Feminine self-reference is unavoidable in an answer to thanks: 'рада', not 'рад'."
+    },
+    {
+      "id": "chat-what-can-you-do",
+      "path": "chat",
+      "utterance": "что ты вообще умеешь?",
+      "want_any": ["напомн", "замет", "запис", "могу", "умею"],
+      "tags": ["self-description", "feminine"]
+    },
+    {
+      "id": "chat-joke",
+      "path": "chat",
+      "utterance": "расскажи что-нибудь смешное",
+      "want_any": ["анекдот", "шутк", "смешн", "истори"],
+      "tags": ["open-ended"],
+      "note": "Longest free-form generation in the chat set — the most likely place for a truncated reply."
+    },
+    {
+      "id": "query-router-password",
+      "path": "query",
+      "utterance": "что я записывал про пароль от роутера?",
+      "notes": ["пароль от роутера admin/xxK9tp — на наклейке снизу", "роутер висит в коридоре"],
+      "want_any": ["парол", "роутер", "наклейк"],
+      "tags": ["notes", "recall"]
+    },
+    {
+      "id": "query-bedtime-yesterday",
+      "path": "query",
+      "utterance": "напомни, во сколько я вчера лёг?",
+      "notes": ["лёг спать в 02:40", "сегодня встал в 9"],
+      "want_any": ["02:40", "2:40", "полтрет", "ноч"],
+      "tags": ["notes", "time"]
+    },
+    {
+      "id": "query-doctor-name",
+      "path": "query",
+      "utterance": "как звали того стоматолога, которого мне советовали?",
+      "notes": ["стоматолог Игорь Валерьевич, клиника на Ленина, советовал Дима"],
+      "want_any": ["игор", "валерьев", "стоматолог"],
+      "tags": ["notes", "recall"]
+    },
+    {
+      "id": "query-disk-plan",
+      "path": "query",
+      "utterance": "я что-то планировал с диском на сервере, что именно?",
+      "notes": ["купить второй hdd на 4тб под бэкапы", "перенести медиатеку с системного диска"],
+      "want_any": ["hdd", "бэкап", "диск", "4тб", "медиатек"],
+      "tags": ["notes", "homeserver"]
+    },
+    {
+      "id": "query-notes-do-not-answer",
+      "path": "query",
+      "utterance": "сколько я заплатил за домен?",
+      "notes": ["домен продлевается в марте", "хостинг оплачен на год вперёд"],
+      "want_any": ["домен", "не зна", "не указ"],
+      "tags": ["notes", "negative"],
+      "note": "The notes do not contain the price. The prompt tells her to say so; a made-up number is the failure being watched for."
+    },
+    {
+      "id": "query-single-note",
+      "path": "query",
+      "utterance": "где лежит запасной ключ?",
+      "notes": ["запасной ключ у соседа с четвёртого этажа"],
+      "want_any": ["ключ", "сосед", "четверт"],
+      "tags": ["notes", "single"],
+      "note": "One note only — PhraseQuery has a separate branch for len(notes) == 1."
+    },
+    {
+      "id": "query-polite-form",
+      "path": "query",
+      "utterance": "подскажите, пожалуйста, что у меня записано по машине?",
+      "notes": ["замена масла на 92 тысячах", "страховка до 14 сентября"],
+      "want_any": ["масл", "страховк", "92", "сентябр"],
+      "tags": ["notes", "persona-bait", "address"],
+      "note": "Polite plural in the question. The answer must still be ty."
+    },
+    {
+      "id": "query-shopping",
+      "path": "query",
+      "utterance": "что мне надо было купить?",
+      "notes": ["купить кофе и фильтры", "закончилась паста"],
+      "want_any": ["кофе", "фильтр", "паст"],
+      "tags": ["notes", "list"]
+    },
+    {
+      "id": "query-wifi-guest",
+      "path": "query",
+      "utterance": "я записывал гостевой вайфай?",
+      "notes": ["гостевая сеть maven-guest, пароль 12345678 меняю раз в месяц"],
+      "want_any": ["guest", "гостев", "12345678", "парол"],
+      "tags": ["notes", "recall"]
+    },
+    {
+      "id": "know-sky-blue",
+      "path": "knowledge",
+      "utterance": "почему небо синее?",
+      "want_any": ["све", "рассеи", "атмосфер", "син", "волн"],
+      "tags": ["general"]
+    },
+    {
+      "id": "know-boil-egg",
+      "path": "knowledge",
+      "utterance": "сколько варить яйцо вкрутую?",
+      "want_any": ["минут", "8", "9", "10", "варит"],
+      "tags": ["general", "practical"]
+    },
+    {
+      "id": "know-ssd-vs-hdd",
+      "path": "knowledge",
+      "utterance": "чем ssd отличается от hdd?",
+      "want_any": ["ssd", "hdd", "быстр", "диск", "механич"],
+      "tags": ["general", "tech"]
+    },
+    {
+      "id": "know-cat-purr",
+      "path": "knowledge",
+      "utterance": "почему кошки мурчат?",
+      "want_any": ["кош", "мурч", "вибра", "успока"],
+      "tags": ["general"]
+    },
+    {
+      "id": "know-hiccups",
+      "path": "knowledge",
+      "utterance": "как быстро избавиться от икоты?",
+      "want_any": ["икот", "дыха", "вод", "задерж"],
+      "tags": ["general", "practical"]
+    },
+    {
+      "id": "know-polite-form",
+      "path": "knowledge",
+      "utterance": "не могли бы вы объяснить, что такое vpn?",
+      "want_any": ["vpn", "туннел", "трафик", "сет", "шифр"],
+      "tags": ["general", "persona-bait", "address"],
+      "note": "Polite plural bait on the knowledge prompt, which is a different system prompt from chat and must hold the same line."
+    },
+    {
+      "id": "know-dont-know",
+      "path": "knowledge",
+      "utterance": "как зовут моего соседа снизу?",
+      "want_any": ["не зна", "не мог"],
+      "tags": ["general", "negative"],
+      "note": "Unanswerable without notes. Admitting it beats inventing a name; watching for the invention."
+    },
+    {
+      "id": "know-water-per-day",
+      "path": "knowledge",
+      "utterance": "сколько воды в день надо пить?",
+      "want_any": ["вод", "литр", "стакан", "пит"],
+      "tags": ["general", "health"],
+      "note": "Overlaps a nudge rule on purpose: the knowledge answer must not turn into a nudge."
+    },
+    {
+      "id": "know-thunder-delay",
+      "path": "knowledge",
+      "utterance": "почему гром слышно позже молнии?",
+      "want_any": ["звук", "све", "быстр", "гром", "молни"],
+      "tags": ["general"]
+    }
+  ]
+}

internal/phraser/grammar_test.go

@@ -0,0 +1,124 @@
+package phraser
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/kami/maven/internal/loop"
+)
+
+// grammarSpy stands in for llama-server: it records the grammar field of every
+// request and always answers with a contract-shaped reply.
+type grammarSpy struct {
+	srv      *httptest.Server
+	grammars []string
+}
+
+func newGrammarSpy(t *testing.T) *grammarSpy {
+	t.Helper()
+	s := &grammarSpy{}
+	s.srv = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var req chatReq
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Errorf("spy: decode request: %v", err)
+		}
+		s.grammars = append(s.grammars, req.Grammar)
+		w.Header().Set("Content-Type", "application/json")
+		w.Write([]byte(`{"choices":[{"message":{"content":"{\"response\": \"ага\", \"mood\": \"neutral\"}"}}]}`))
+	}))
+	t.Cleanup(s.srv.Close)
+	return s
+}
+
+// callAllPhrasingPaths hits every path that expects the JSON contract.
+func callAllPhrasingPaths(t *testing.T, p *LLMPhraser) {
+	t.Helper()
+	ctx := context.Background()
+	if _, err := p.PhraseNudge(ctx, loop.Candidate{Rule: loop.WaterRule(), Severity: loop.Sev1}); err != nil {
+		t.Fatalf("PhraseNudge: %v", err)
+	}
+	if _, err := p.PhraseChat(ctx, "привет", nil); err != nil {
+		t.Fatalf("PhraseChat: %v", err)
+	}
+	// Both branches: no notes (general knowledge) and with notes (grounded).
+	if _, err := p.PhraseQuery(ctx, "сколько воды я выпил", nil); err != nil {
+		t.Fatalf("PhraseQuery (no notes): %v", err)
+	}
+	if _, err := p.PhraseQuery(ctx, "сколько воды я выпил", []string{"два литра"}); err != nil {
+		t.Fatalf("PhraseQuery (notes): %v", err)
+	}
+}
+
+func TestGrammarIsAttachedToEveryPhrasingRequest(t *testing.T) {
+	if strings.TrimSpace(responseGrammar) == "" {
+		t.Fatal("responseGrammar is empty")
+	}
+	spy := newGrammarSpy(t)
+	p := NewLLMPhraserAt(spy.srv.URL, Config{})
+
+	callAllPhrasingPaths(t, p)
+
+	if len(spy.grammars) != 4 {
+		t.Fatalf("expected 4 requests, got %d", len(spy.grammars))
+	}
+	for i, g := range spy.grammars {
+		if g != responseGrammar {
+			t.Errorf("request %d carries grammar %q, want responseGrammar", i, g)
+		}
+	}
+}
+
+func TestNoGrammarConfigDisablesIt(t *testing.T) {
+	spy := newGrammarSpy(t)
+	p := NewLLMPhraserAt(spy.srv.URL, Config{NoGrammar: true})
+
+	callAllPhrasingPaths(t, p)
+
+	for i, g := range spy.grammars {
+		if g != "" {
+			t.Errorf("request %d still carries a grammar with NoGrammar set: %q", i, g)
+		}
+	}
+}
+
+// The grammar's string rule must accept any codepoint, not just ASCII. Replies
+// are Russian: an ASCII-only class would constrain the model into empty replies.
+func TestGrammarStringRuleIsNotASCIIOnly(t *testing.T) {
+	if !strings.Contains(responseGrammar, `([^"\\] | "\\" ["\\/bfnrt])`) {
+		t.Error("string rule is not the any-codepoint-except-quote-and-backslash class; Cyrillic replies would be impossible")
+	}
+}
+
+// What the grammar describes must survive the parser that reads it back — a
+// Russian body with an escaped quote inside, hand-built to test the contract.
+func TestGrammarShapedJSONParses(t *testing.T) {
+	raw := `{"response": "он сказал \"привет\" и ушёл.\nвот так.", "mood": "confused"}`
+	text, mood := parseResponseMood(raw)
+	if want := "он сказал \"привет\" и ушёл.\nвот так."; text != want {
+		t.Errorf("response = %q, want %q", text, want)
+	}
+	if mood != "confused" {
+		t.Errorf("mood = %q, want confused", mood)
+	}
+}
+
+// Every mood the grammar permits is one the contract knows, and all five are there.
+func TestGrammarMoodEnumMatchesTheContract(t *testing.T) {
+	for _, m := range []string{"neutral", "happy", "thinking", "tired", "confused"} {
+		if !strings.Contains(responseGrammar, `"\"`+m+`\""`) {
+			t.Errorf("mood %q missing from the grammar", m)
+		}
+	}
+	// No sixth mood: the enum line lists exactly five alternatives.
+	for _, line := range strings.Split(responseGrammar, "\n") {
+		if strings.HasPrefix(line, "mood") {
+			if n := strings.Count(line, "|") + 1; n != 5 {
+				t.Errorf("mood rule lists %d alternatives, want 5: %s", n, line)
+			}
+		}
+	}
+}

internal/phraser/llmphraser.go

@@ -18,6 +18,7 @@ import (
 	"github.com/kami/maven/internal/delivery"
 	"github.com/kami/maven/internal/dialogue"
 	"github.com/kami/maven/internal/loop"
+	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/router"
 )
 
@@ -39,7 +40,18 @@ type Config struct {
 	NGpuLayers int
 	NCtx       int
 	Timeout    time.Duration
-	Persona    string // optional prompt prefix tuning maven's character
+
+	// ContextBlock renders the shared context block (who he is, how to
+	// address him, the time) fresh for each turn. See internal/persona.
+	// nil ⇒ no block, the prompts stand alone.
+	ContextBlock func() string
+
+	// NoGrammar turns the GBNF constraint off (zero value ⇒ grammar ON).
+	// The escape hatch exists because the target resident model — the
+	// locally CPT'd Qwen3-1.7B — does not exist yet: if its chat template
+	// ever fights the grammar, the fix should be a config flip on the
+	// deploy box, not a code change and a rebuild.
+	NoGrammar bool
 }
 
 func DefaultConfig(modelPath string) Config {
@@ -184,7 +196,10 @@ func (p *LLMPhraser) PhraseNudge(ctx context.Context, c loop.Candidate) (deliver
 		body, _ = parsePhrase(resp)
 	}
 	if body == "" {
-		body = fmt.Sprintf("%s — %s", c.Rule.Name, sevLabel(c.Severity))
+		// The model said nothing usable. Say it in Russian anyway — this text
+		// goes straight to a Russian piper voice, so the old "water — care"
+		// fallback was unspeakable.
+		body = fallbackNudge(c)
 	}
 	if mood == "" {
 		mood = "neutral"
@@ -199,7 +214,7 @@ func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []
 	if len(notes) == 0 {
 		// General knowledge — no notes to ground the answer. The system
 		// prompt is the single tested source in router.KnowledgePrompt.
-		sys := router.KnowledgePrompt()
+		sys := persona.Prepend(p.cfg.ContextBlock, router.KnowledgePrompt())
 		prompt := fmt.Sprintf("Пользователь спрашивает: \"%s\".", utterance)
 		resp, err := p.chatWithSystem(ctx, sys, prompt, 256)
 		if err != nil || resp == "" {
@@ -235,7 +250,7 @@ func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []
 // message array from dialogue history + the current user utterance. Falls back
 // to a simple greeting on any LLM error — better to say something than nothing.
 func (p *LLMPhraser) PhraseChat(ctx context.Context, utterance string, history []dialogue.Turn) (string, error) {
-	sys := chatSystemPrompt(p.cfg.Persona)
+	sys := chatSystemPrompt(p.cfg.ContextBlock)
 	msgs := []chatMsg{
 		{Role: "system", Content: sys},
 	}
@@ -264,17 +279,14 @@ func (p *LLMPhraser) PhraseChat(ctx context.Context, utterance string, history [
 }
 
 // chatSystemPrompt returns the system prompt for conversational chat.
-// Prepends the configured persona when set.
-func chatSystemPrompt(persona string) string {
+// Prepends the shared context block when the phraser has one.
+func chatSystemPrompt(block func() string) string {
 	base := `You are maven, a self-hosted personal assistant. You're talking with your owner. 
 Keep replies brief (1-3 sentences) and natural. You're helpful, curious, and a little warm.
 Respond in the user's language (Russian or English, matching their last message). 
 Never roleplay emotions you don't have, but stay friendly. 
 Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}. "response" is your reply text; "mood" reflects your tone (neutral/happy/thinking/tired/confused).`
-	if persona != "" {
-		base = persona + "\n\n" + base
-	}
-	return base
+	return persona.Prepend(block, base)
 }
 
 // chatWithMessages sends a full message array (system + history + current) to
@@ -285,6 +297,7 @@ func (p *LLMPhraser) chatWithMessages(ctx context.Context, msgs []chatMsg, maxTo
 		Messages:    msgs,
 		Temperature: 0.7,
 		MaxTokens:   maxTokens,
+		Grammar:     p.grammar(),
 	}
 	body, err := json.Marshal(req)
 	if err != nil {
@@ -364,6 +377,37 @@ type chatReq struct {
 	Messages    []chatMsg `json:"messages"`
 	Temperature float64   `json:"temperature"`
 	MaxTokens   int       `json:"max_tokens"`
+	// Grammar is llama-server's `grammar` field (GBNF). Same wiring as
+	// internal/llm.Req.Grammar. Empty ⇒ unconstrained sampling.
+	Grammar string `json:"grammar,omitempty"`
+}
+
+// responseGrammar — GBNF constraining the model to the documented phrasing
+// contract and nothing else: {"response": "<text>", "mood": "<enum>"}.
+//
+// Without it a 0.8B answers roughly one chat turn in three with open reasoning
+// as plain text ("Thinking Process:" …), which no tag-stripper can remove and
+// which eats the token budget before the JSON closes. Modelled on
+// routeGrammar in internal/router/llmrouter.go so the two read alike.
+//
+// text accepts ANY codepoint except the two JSON must escape — the replies are
+// Russian, so an ASCII-only rule would make every reply empty. The escape rule
+// is what lets the model close a string it opened with a quote inside. Length
+// is bounded so a repetition loop truncates the field, not the JSON object.
+const responseGrammar = `
+root   ::= "{" ws "\"response\"" ws ":" ws string ws "," ws "\"mood\"" ws ":" ws mood ws "}"
+mood   ::= "\"neutral\"" | "\"happy\"" | "\"thinking\"" | "\"tired\"" | "\"confused\""
+string ::= "\"" ([^"\\] | "\\" ["\\/bfnrt]){0,400} "\""
+ws     ::= [ \t\n]*
+`
+
+// grammar returns the GBNF to attach to a phrasing request, or "" when the
+// operator turned it off.
+func (p *LLMPhraser) grammar() string {
+	if p.cfg.NoGrammar {
+		return ""
+	}
+	return responseGrammar
 }
 
 type chatResp struct {
@@ -388,6 +432,7 @@ func (p *LLMPhraser) chatWithSystem(ctx context.Context, system, user string, ma
 		},
 		Temperature: 0.7,
 		MaxTokens:   maxTokens,
+		Grammar:     p.grammar(),
 	}
 	body, err := json.Marshal(req)
 	if err != nil {
@@ -428,40 +473,155 @@ func (p *LLMPhraser) chatWithSystem(ctx context.Context, system, user string, ma
 	return stripThink(content), nil
 }
 
+// nudgeSystem — the phrasing contract for nudges.
+//
+// Written as filled-in examples, not as a schema with "..." in it. A 0.8B
+// copies whatever sits in the response slot, so a literal placeholder there
+// teaches it to answer with the placeholder. Measured: 7/15 nudges came back
+// as "..." before this. See PHRASING-EVAL-31-07-2026.md.
+//
+// Russian only, feminine self-reference, second person masculine (the owner is
+// a man). She talks TO him, informally, singular — never "вы", never "он".
+// One short sentence — the nudge is spoken aloud.
+const nudgeSystem = `Ты — Maven, домашняя ассистентка. О себе говоришь в женском роде ("я проверила", "я записала"). Владелец — мужчина, обращайся к нему в мужском роде ("ты пил", "ты забыл").
+Говоришь с ним на "ты", в единственном числе ("выпей", "встань"). Никогда не "вы"/"вас"/"ваш" и никогда "он"/"его" — ты говоришь ему, а не о нём.
+
+Пиши ОДНО короткое напоминание по-русски: не больше 120 символов и не больше 16 слов. Только по делу.
+
+Запрещено: обращения ("дорогой", "милый"), эмодзи, извинения ("прости", "извини"), вопросы о самочувствии, похвала, больше одного восклицательного знака, английские слова кроме имён сервисов.
+
+Отвечай ТОЛЬКО одним объектом JSON с полями "response" и "mood".
+"response" — сам текст напоминания.
+"mood" — ровно одно из: neutral, happy, thinking, tired, confused.
+
+Так выглядит правильный ответ по форме. Темы здесь посторонние — их в запросе не будет:
+{"response": "Стиральная машина закончила. Развесь бельё.", "mood": "neutral"}
+{"response": "Ноутбук на трёх процентах. Я поставила его на зарядку.", "mood": "confused"}
+
+Это примеры ФОРМЫ, а не темы. Пиши только про ту ситуацию, которую тебе дали в запросе. Не копируй примеры и никогда не пиши "..." в поле response.`
+
 func (p *LLMPhraser) systemPrompt() string {
-	base := `You are maven, a self-hosted personal assistant. Generate brief, natural nudge messages in the user's language (Russian or English). Respond ONLY with valid JSON: {"response": "full voice message", "mood": "neutral"}. "response" is what the user hears; "mood" reflects maven's tone (neutral/happy/thinking/tired/confused).`
-	if p.cfg.Persona != "" {
-		base = p.cfg.Persona + "\n\n" + base
-	}
-	return base
+	return persona.Prepend(p.cfg.ContextBlock, nudgeSystem)
 }
 
 // querySystemPrompt returns the system prompt for PhraseQuery (notes + general
 // knowledge). Prepends the configured persona when set.
 func (p *LLMPhraser) querySystemPrompt() string {
 	base := "You are maven, a self-hosted personal assistant answering from your notes. Answer briefly and naturally in Russian starting with \"вот что я нашла: \". Respond ONLY with valid JSON: {\"response\": \"...\", \"mood\": \"neutral\"}."
-	if p.cfg.Persona != "" {
-		base = p.cfg.Persona + "\n\n" + base
+	return persona.Prepend(p.cfg.ContextBlock, base)
+}
+
+// ruleTopics — Russian gloss for each built-in rule name. The rule names are
+// English identifiers; a 0.8B asked to nudge about "netdata_critical" writes
+// about nothing. The daemon knows what its own rules mean, so it says so.
+var ruleTopics = map[string]string{
+	"water":            "он давно не пил воду",
+	"meal":             "он давно не ел",
+	"break":            "он давно без перерыва, пора встать и размяться",
+	"service_down":     "сервис не отвечает, лежит",
+	"netdata_critical": "критический алярм в netdata, проблема с диском или местом",
+}
+
+// ruleKeywords — the word the message must contain. The 0.8B drifts to
+// whatever topic it saw last unless the required word is named outright.
+var ruleKeywords = map[string]string{
+	"water":            "воду",
+	"meal":             "поешь",
+	"break":            "перерыв",
+	"service_down":     "сервис",
+	"netdata_critical": "диск",
+}
+
+// ruleTopic turns a rule name into a Russian description of the situation.
+// "routine:зарядка" and "morning:утро" carry their own Russian suffix.
+func ruleTopic(rule string) string {
+	if t, ok := ruleTopics[rule]; ok {
+		return t
 	}
-	return base
+	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
+		switch rule[:i] {
+		case "morning":
+			return "утро, пора начать день: " + rule[i+1:]
+		default:
+			return "пора сделать по распорядку: " + rule[i+1:]
+		}
+	}
+	return rule
+}
+
+// ruleKeyword — the word the nudge must contain, or "" when the rule name's
+// own Russian suffix already is that word.
+func ruleKeyword(rule string) string {
+	if k, ok := ruleKeywords[rule]; ok {
+		return k
+	}
+	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
+		return rule[i+1:]
+	}
+	return ""
+}
+
+// ruDur — duration in Russian. humanDur is English and its output was landing
+// verbatim in the message.
+func ruDur(d time.Duration) string {
+	if d < 0 {
+		d = 0
+	}
+	h, m := int(d.Hours()), int(d.Minutes())%60
+	switch {
+	case h >= 2:
+		return fmt.Sprintf("%d ч", h)
+	case h == 1 && m >= 30:
+		return "полтора часа"
+	case h == 1:
+		return "час"
+	default:
+		return fmt.Sprintf("%d мин", m)
+	}
+}
+
+// fallbackNudge — plain Russian for when the model returns nothing parseable.
+var fallbackNudges = map[string]string{
+	"water":            "Ты давно не пил воду.",
+	"meal":             "Ты давно не ел, поешь.",
+	"break":            "Пора сделать перерыв.",
+	"service_down":     "Сервис не отвечает.",
+	"netdata_critical": "Критический алярм: проверь диск.",
+}
+
+func fallbackNudge(c loop.Candidate) string {
+	if s, ok := fallbackNudges[c.Rule.Name]; ok {
+		return s
+	}
+	if kw := ruleKeyword(c.Rule.Name); kw != "" {
+		return "Напоминаю: " + kw + "."
+	}
+	return "Напоминаю о деле."
 }
 
 func buildNudgePrompt(c loop.Candidate) string {
 	var ctxParts []string
-	ctxParts = append(ctxParts, fmt.Sprintf("Rule: %s", c.Rule.Name))
-	ctxParts = append(ctxParts, fmt.Sprintf("Severity: %s", sevLabel(c.Severity)))
-
+	ctxParts = append(ctxParts, "Ситуация: "+ruleTopic(c.Rule.Name))
+	if f, ok := c.State.Facts[c.Rule.Name]; ok && f.Key != "" && f.Key != c.Rule.Name {
+		ctxParts = append(ctxParts, "Что именно: "+f.Key)
+	}
 	if d, ok := c.State.Since(c.Rule.Name); ok {
-		ctxParts = append(ctxParts, fmt.Sprintf("Duration since last event: %s", humanDur(d)))
+		ctxParts = append(ctxParts, "Прошло: "+ruDur(d))
+	}
+	switch sevLabel(c.Severity) {
+	case "alarm":
+		ctxParts = append(ctxParts, "Срочно, скажи прямо.")
+	case "ops":
+		ctxParts = append(ctxParts, "Это про сервер, не про здоровье.")
+	}
+	tail := "Напиши напоминание про эту ситуацию. Одно предложение, по-русски, в JSON."
+	if kw := ruleKeyword(c.Rule.Name); kw != "" {
+		// Last line on purpose: a 0.8B weights the end of the prompt hardest,
+		// and without the required word it drifts back to the examples.
+		tail += " Ответ ДОЛЖЕН содержать слово «" + kw + "»."
 	}
 
-	return fmt.Sprintf(
-		`Generate a nudge message. Context:
-%s
-
-Respond as JSON: {"response": "...", "mood": "..."}`,
-		strings.Join(ctxParts, "\n"),
-	)
+	return strings.Join(ctxParts, "\n") + "\n\n" + tail
 }
 
 type responseMood struct {

internal/router/embedder.go

@@ -2,6 +2,7 @@ package router
 
 import (
 	"context"
+	"fmt"
 	"math"
 	"unicode"
 )
@@ -19,6 +20,24 @@ type Embedder interface {
 	Close() error
 }
 
+// IdentifiedEmbedder — an embedder that can name itself. The name goes into
+// the DB next to the vectors it wrote, so a later model swap is caught instead
+// of silently returning nonsense scores (Vikunja #378).
+type IdentifiedEmbedder interface {
+	Embedder
+	ID() string
+}
+
+// EmbedderID is the stable string stored alongside the vectors. It comes from
+// the embedder itself — nobody hand-types a model name twice — and changes
+// whenever the model or its dimension changes.
+func EmbedderID(e Embedder) string {
+	if i, ok := e.(IdentifiedEmbedder); ok {
+		return i.ID()
+	}
+	return fmt.Sprintf("unknown@%d", e.Dim())
+}
+
 // AsymmetricEmbedder — an embedder that wants to know whether a text is a
 // search query or a stored passage. Recall is asymmetric: a short question
 // goes in, a longer note comes out. The e5 family is trained for exactly that
@@ -70,6 +89,10 @@ func NewHashEmbedder(dim int) *HashEmbedder {
 
 func (h *HashEmbedder) Dim() int { return h.dim }
 
+// ID names this embedder for the DB marker. The dimension is part of it
+// because a HashEmbedder of another width is a different vector space.
+func (h *HashEmbedder) ID() string { return fmt.Sprintf("hash@%d", h.dim) }
+
 func (h *HashEmbedder) Close() error { return nil }
 
 func (h *HashEmbedder) Embed(_ context.Context, text string) ([]float32, error) {

internal/router/embedderid_test.go

@@ -0,0 +1,24 @@
+package router
+
+import "testing"
+
+func TestEmbedderIDFromModelPath(t *testing.T) {
+	got := modelIDFromPath("/opt/maven/models/embedder/multilingual-e5-small.onnx")
+	if got != "multilingual-e5-small@384" {
+		t.Fatalf("modelIDFromPath = %q", got)
+	}
+	// A different model file must produce a different id, even at 384 dim.
+	old := modelIDFromPath("/opt/maven/models/embedder/paraphrase-multilingual-MiniLM-L12-v2.onnx")
+	if old == got {
+		t.Fatal("two different models share one id")
+	}
+}
+
+func TestEmbedderIDIncludesDim(t *testing.T) {
+	if id := EmbedderID(NewHashEmbedder(1024)); id != "hash@1024" {
+		t.Fatalf("EmbedderID = %q", id)
+	}
+	if EmbedderID(NewHashEmbedder(1024)) == EmbedderID(NewHashEmbedder(384)) {
+		t.Fatal("dimension not part of the id")
+	}
+}

internal/router/eval/llmrouter_test.go

@@ -1,11 +1,8 @@
 package eval
 
 import (
-	"bytes"
 	"context"
-	"encoding/json"
 	"fmt"
-	"net/http"
 	"os"
 	"strings"
 	"testing"
@@ -29,13 +26,20 @@ import (
 // a bake-off across checkpoints (#278, #250) produces tables you can tell
 // apart. Point the variable at one server at a time.
 //
-// Three configurations, because "the LLM router" is ambiguous and the three
-// numbers answer different questions:
+// Two configurations, because "the LLM router" is ambiguous and the two numbers
+// answer different questions:
 //
-//	llm-only        — the model alone. Measures the prompt + grammar contract.
-//	cascade+llm     — what #320 would actually ship: stage-0 grammar, then the
-//	                  model, then the classifier as the failure floor.
-//	llm-no-thinking — diagnostic only, not a shippable path (see below).
+//	llm-only    — the model alone. Measures the prompt + grammar contract.
+//	cascade+llm — what #320 would actually ship: stage-0 grammar, then the
+//	              model, then the classifier as the failure floor.
+//
+// There used to be a third, "thinking off", which looked 6 points better. It is
+// gone: it was measured with a hand-rolled HTTP client that quietly dropped
+// repeat_penalty, so the gap was the missing penalty and not the thinking mode.
+// Re-measured with everything else held equal, thinking off scores exactly the
+// same, case for case — and a direct probe shows this llama-server build ignores
+// enable_thinking / reasoning_budget for this model anyway, so there was nothing
+// to turn off. Full write-up in ROUTING-EVAL-31-07-2026.md (Vikunja #376).
 func TestLLMRouterBaseline(t *testing.T) {
 	base := os.Getenv("MAVEN_LLM_URL")
 	if base == "" {
@@ -57,12 +61,12 @@ func TestLLMRouterBaseline(t *testing.T) {
 	}
 
 	ctx := context.Background()
-	model, err := ModelID(ctx, base)
+	model, err := llm.ModelID(ctx, base)
 	if err != nil {
 		// Not fatal: an unlabelled score is still a score. But say so loudly,
 		// because an unlabelled row in a bake-off table is worthless.
-		t.Logf("could not read model id from %s: %v — reports will say %q", base, err, "unknown-model")
-		model = "unknown-model"
+		t.Logf("could not read model id from %s: %v — reports will say %q", base, err, llm.UnknownModel)
+		model = llm.UnknownModel
 	}
 	t.Logf("scoring model %s at %s", model, base)
 	lr := router.NewLLMRouter(client)
@@ -96,104 +100,18 @@ func TestLLMRouterBaseline(t *testing.T) {
 	}
 	t.Log("\n" + repCascade.String() + repCascade.Failures())
 
-	// llm-no-thinking: same prompt and grammar with the chat template's
-	// thinking mode off. Qwen3.5's template defaults thinking=1, so under a
-	// grammar the constrained JSON lands in reasoning_content with content
-	// empty — llm.Client's ReasoningContent fallback is what makes the router
-	// work at all today, by accident rather than design.
-	//
-	// MEASURED 2026-07-31: this variant scores identically to as-deployed
-	// (18/76, 48.7% intent-only, 2 errors, same p50). Thinking mode is a
-	// non-issue under a grammar — llama.cpp constrains the same token stream
-	// either way. Kept so the question stays answered instead of being
-	// re-asked, and so internal/llm does NOT grow a chat_template_kwargs field
-	// for a problem that does not exist.
-	repNoThink, err := Score(ctx, "llm-only ("+model+", thinking off) [diagnostic]",
-		RouterFunc(func(ctx context.Context, u string, now time.Time) (router.Decision, error) {
-			d, ok, err := router.NewLLMRouter(&noThinkCompleter{base: base, http: &http.Client{Timeout: 60 * time.Second}}).Route(ctx, u, now)
-			if err != nil {
-				return d, err
-			}
-			if !ok {
-				return d, fmt.Errorf("llm router declined without an error")
-			}
-			return d, nil
-		}), f)
-	if err != nil {
-		t.Fatalf("Score no-thinking: %v", err)
-	}
-	t.Log("\n" + repNoThink.String() + repNoThink.Failures())
-
 	// Reports rather than asserts — the numbers are inputs to the #320
 	// decision, and an assertion here would be this test inventing the bar.
 	// The one thing worth failing on is a harness fault: if every single case
 	// errors, the run measured infrastructure, not routing, and the report
 	// must not be mistaken for a score.
-	for _, rep := range []Report{repLLM, repCascade, repNoThink} {
+	for _, rep := range []Report{repLLM, repCascade} {
 		if rep.Errors == rep.Total {
 			t.Errorf("%s: all %d cases errored — harness fault, not a measurement", rep.Name, rep.Total)
 		}
 	}
 }
 
-// noThinkCompleter — llm.Client with chat_template_kwargs.enable_thinking
-// false. A test-local copy rather than a change to internal/llm: whether the
-// daemon should send it is the open question, and answering it here by adding
-// the field would prejudge #320.
-type noThinkCompleter struct {
-	base string
-	http *http.Client
-}
-
-func (c *noThinkCompleter) Complete(ctx context.Context, r llm.Req) (string, error) {
-	payload := map[string]any{
-		"messages": []map[string]string{
-			{"role": "system", "content": r.System},
-			{"role": "user", "content": r.User},
-		},
-		"max_tokens":           r.MaxTokens,
-		"temperature":          0,
-		"grammar":              r.Grammar,
-		"chat_template_kwargs": map[string]any{"enable_thinking": false},
-	}
-	b, err := json.Marshal(payload)
-	if err != nil {
-		return "", err
-	}
-	req, err := http.NewRequestWithContext(ctx, "POST", c.base+"/v1/chat/completions", bytes.NewReader(b))
-	if err != nil {
-		return "", err
-	}
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := c.http.Do(req)
-	if err != nil {
-		return "", err
-	}
-	defer resp.Body.Close()
-	if resp.StatusCode != 200 {
-		return "", fmt.Errorf("status %d", resp.StatusCode)
-	}
-	var out struct {
-		Choices []struct {
-			Message struct {
-				Content          string `json:"content"`
-				ReasoningContent string `json:"reasoning_content"`
-			} `json:"message"`
-		} `json:"choices"`
-	}
-	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
-		return "", err
-	}
-	if len(out.Choices) == 0 {
-		return "", fmt.Errorf("no choices")
-	}
-	m := out.Choices[0].Message
-	if m.Content != "" {
-		return m.Content, nil
-	}
-	return m.ReasoningContent, nil
-}
-
 func ping(ctx context.Context, c *llm.Client) error {
 	ctx, cancel := context.WithTimeout(ctx, 90*time.Second)
 	defer cancel()

internal/router/eval/ru_routing_v1.json

@@ -22,6 +22,7 @@
     { "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-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" },

internal/router/knowledge.go

@@ -3,5 +3,8 @@ package router
 // KnowledgePrompt returns the system prompt for general knowledge questions
 // that the phraser uses when no notes match the query.
 func KnowledgePrompt() string {
-	return `Ты — Мавена, персональный ассистент. Ответь кратко из своих знаний. Если не знаешь — скажи "не знаю". Не выдумывай. Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}.`
+	// No self-introduction here: the shared persona block already says who she
+	// is, and this line used to disagree with it — a different name ("Мавена")
+	// and a masculine noun ("ассистент") in front of a feminine persona.
+	return `Ответь кратко из своих знаний. Если не знаешь — скажи "не знаю". Не выдумывай. Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}.`
 }

internal/router/knowledge_test.go

@@ -11,10 +11,18 @@ func TestKnowledgePrompt(t *testing.T) {
 		t.Fatal("KnowledgePrompt returned empty string")
 	}
 	// Must contain key instructions
-	checks := []string{"Мавена", "не знаю", "не выдумывай"}
+	checks := []string{"не знаю", "не выдумывай"}
 	for _, c := range checks {
 		if !strings.Contains(strings.ToLower(prompt), strings.ToLower(c)) {
 			t.Errorf("KnowledgePrompt should mention %q", c)
 		}
 	}
+	// Who she is comes from the shared persona block now. This prompt used to
+	// say it too, with a different name and a masculine noun, which is the
+	// drift the block exists to stop.
+	for _, w := range []string{"Мавена", "ассистент"} {
+		if strings.Contains(prompt, w) {
+			t.Errorf("KnowledgePrompt should not introduce her (%q) — the persona block does", w)
+		}
+	}
 }

internal/router/llmrouter.go

@@ -44,10 +44,21 @@ ws     ::= [ \t\n]*
 // Changed again 31-07-2026: added the "unknown" escape hatch so the model can
 // admit it cannot route (Vikunja #359).
 //
+// Changed again 31-07-2026: added the clock/calendar rule (Vikunja #374). The
+// prompt never said which side "который час" or "какое число завтра" belong on,
+// so the model guessed — `system→query ×4` in every eval run. The rule sits
+// above the question test on purpose: these utterances all carry a question
+// word, so a later rule would never be reached. The boundary is what the
+// daemon can actually answer: only replySystem in cmd/mavend/voice.go owns the
+// clock and the calendar formatter, while the agenda ("что у меня завтра") is
+// answered inside the query branch, so that side stays query.
+//
 // The training workspace keeps its own copy of this prompt for relabelling, and
 // `llm/check_prompt_parity.py` there compares the two. That copy is in another
 // repo and was not touched, so parity will fail until it gets the same edits —
-// both the rule reorder and the "unknown" wording (Vikunja #362).
+// both the rule reorder and the "unknown" wording (Vikunja #362) — and now the
+// clock/calendar rule too. The training workspace is not checked out on this
+// box at all, so it could not be updated here; #362 still covers the catch-up.
 const routeSystem = `Классифицируй ровно одно сообщение пользователя. Верни ОДИН JSON-массив действий.
 
 Ровно одно намерение: fact, reminder, note, query, act, chat, system.
@@ -56,19 +67,21 @@ const routeSystem = `Классифицируй ровно одно сообще
 Классифицируй по цели пользователя. Порядок решения:
 1. Хочет напоминание в будущем → reminder
 2. Явно просит сохранить информацию → note
-3. Задаёт вопрос: есть вопросительное слово (сколько, что, какой, когда, где, кто, почему, как) или знак «?» → query
-4. Хочет получить информацию, в том числе о своих же данных → query
-5. Утверждает: сообщает или обновляет текущее состояние/событие → fact
-6. Просит выполнить работу → act
-7. Про ассистента, настройки или память → system
-8. Реплика — обрывок или указание на неназванное («это», «то», «потом»), и без него непонятно, что именно нужно сделать → unknown
-9. Иначе → chat
+3. Спрашивает только «который час» / «какое число» / «какой день недели» — сами часы или календарная дата, без своих данных → system
+4. Задаёт вопрос: есть вопросительное слово (сколько, что, какой, когда, где, кто, почему, как) или знак «?» → query
+5. Хочет получить информацию, в том числе о своих же данных → query
+6. Утверждает: сообщает или обновляет текущее состояние/событие → fact
+7. Просит выполнить работу → act
+8. Про ассистента, настройки или память → system
+9. Реплика — обрывок или указание на неназванное («это», «то», «потом»), и без него непонятно, что именно нужно сделать → unknown
+10. Иначе → chat
 
 Различия:
 - note — сохранить информацию, без напоминания. text = суть.
 - reminder — уведомить позже. text = что напомнить.
 - fact — неявное обновление: пользователь сообщает, что что-то в мире изменилось (текущее/изменённое состояние, случившееся событие). key/value.
 - unknown — редкий случай. Ставь его, только если в самой реплике нет ни предмета, ни действия. Короткая, простая или незнакомая тема — это не причина для unknown: приветствие и болтовня — это chat, вопрос на любую тему — это query, просьба сделать что-то названное — это act.
+- system против query — часы и календарная дата сами по себе (сколько времени, какое число, какой день недели — можно и про завтра, и про другой город) — это system. А что записано в календаре или в памяти («что у меня завтра», «какие есть напоминания») — это query. Если в реплике есть просьба (напомни, запиши, сделай), то названное время — просто деталь просьбы, и это не system.
 - query против fact — решает форма реплики, а не тема. Вопрос о состоянии — это query, даже если названо то же самое, что бывает в fact. Только утверждение — это fact.
 
 Примеры:
@@ -82,6 +95,8 @@ const routeSystem = `Классифицируй ровно одно сообще
 "что такое docker?" → {"intent":"query","text":"что такое docker"}
 "напиши письмо" → {"intent":"act","verb":"написать письмо"}
 "очисти память" → {"intent":"system"}
+"который час?" → {"intent":"system"}
+"какое число завтра?" → {"intent":"system"}
 "привет" → {"intent":"chat","text":"привет"}
 "сделай это" → {"intent":"unknown"}
 "ну это" → {"intent":"unknown"}

internal/router/onnxembedder.go

@@ -33,6 +33,7 @@ const (
 type onnxEmbedder struct {
 	tokenizer *unigramTokenizer
 	session   *ort.DynamicSession[int64, float32]
+	id        string
 }
 
 func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, error) {
@@ -58,11 +59,31 @@ func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, e
 	return &onnxEmbedder{
 		tokenizer: tok,
 		session:   session,
+		id:        modelIDFromPath(modelPath),
 	}, nil
 }
 
 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.
+func (e *onnxEmbedder) ID() string { return e.id }
+
+// modelIDFromPath turns /opt/.../multilingual-e5-small.onnx into
+// "multilingual-e5-small@384".
+func modelIDFromPath(modelPath string) string {
+	name := modelPath
+	if i := strings.LastIndexAny(name, "/\\"); i >= 0 {
+		name = name[i+1:]
+	}
+	name = strings.TrimSuffix(name, ".onnx")
+	if name == "" {
+		name = "onnx"
+	}
+	return fmt.Sprintf("%s@%d", name, embedDim)
+}
+
 // Embed treats the text as a query. The classifier compares one short
 // utterance to another short seed phrase, so both sides get the same prefix
 // and the comparison stays fair. The recall path must call EmbedQuery and

internal/router/slots.go

@@ -449,16 +449,30 @@ func (AnaphoraResolver) Resolve(text string) (ref string, ok bool) {
 	return "", false
 }
 
-// ParseCalendarDate detects RU calendar date words in text and returns the
-// resolved time (midnight UTC+0 for "сегодня"/"today", next day for "завтра"/"tomorrow").
-// Returns zero time + false if no match.
+// ParseCalendarDate detects RU/EN calendar day words in text and returns
+// midnight of that day in now's own time zone. Handles "сегодня", "завтра",
+// "послезавтра", "вчера" (and the English words). Returns zero time + false
+// if no match.
+//
+// "послезавтра" is checked before "завтра" because it contains it.
 func ParseCalendarDate(text string, now time.Time) (time.Time, bool) {
 	lower := strings.ToLower(text)
-	if strings.Contains(lower, "сегодня") || strings.Contains(lower, "today") {
-		return now.Truncate(24 * time.Hour), true
-	}
-	if strings.Contains(lower, "завтра") || strings.Contains(lower, "tomorrow") {
-		return now.Truncate(24 * time.Hour).Add(24 * time.Hour), true
+	switch {
+	case strings.Contains(lower, "сегодня") || strings.Contains(lower, "today"):
+		return midnight(now, 0), true
+	case strings.Contains(lower, "послезавтра") || strings.Contains(lower, "day after tomorrow"):
+		return midnight(now, 2), true
+	case strings.Contains(lower, "завтра") || strings.Contains(lower, "tomorrow"):
+		return midnight(now, 1), true
+	case strings.Contains(lower, "вчера") || strings.Contains(lower, "yesterday"):
+		return midnight(now, -1), true
 	}
 	return time.Time{}, false
 }
+
+// midnight returns the start of the day that is `days` away from now, in
+// now's time zone (now.Truncate(24h) would cut on a UTC boundary instead).
+func midnight(now time.Time, days int) time.Time {
+	y, m, d := now.AddDate(0, 0, days).Date()
+	return time.Date(y, m, d, 0, 0, 0, 0, now.Location())
+}

internal/router/slots_ru_test.go

@@ -45,6 +45,9 @@ func TestParseCalendarDate(t *testing.T) {
 		{"расписание на завтра", time.Date(2026, 7, 7, 0, 0, 0, 0, time.UTC), true},
 		{"what's today", time.Date(2026, 7, 6, 0, 0, 0, 0, time.UTC), true},
 		{"tomorrow plans", time.Date(2026, 7, 7, 0, 0, 0, 0, time.UTC), true},
+		{"какое число послезавтра", time.Date(2026, 7, 8, 0, 0, 0, 0, time.UTC), true},
+		{"что было вчера", time.Date(2026, 7, 5, 0, 0, 0, 0, time.UTC), true},
+		{"yesterday plans", time.Date(2026, 7, 5, 0, 0, 0, 0, time.UTC), true},
 		{"какая погода", time.Time{}, false},
 		{"сколько времени", time.Time{}, false},
 		{"", time.Time{}, false},

internal/store/backfill.go

@@ -0,0 +1,161 @@
+package store
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"time"
+)
+
+// EmbedFunc embeds one piece of stored text. The caller passes
+// router.EmbedPassage — the STORE side of the query/passage asymmetry, which is
+// the side every vector in the DB was written with. (Passing the query side
+// would put the stored vectors in the wrong half of the space and quietly halve
+// recall.) A func instead of an interface keeps this package free of any
+// dependency on internal/router.
+type EmbedFunc func(ctx context.Context, text string) ([]float32, error)
+
+// BackfillResult is what the re-embed run did, for logging.
+type BackfillResult struct {
+	Skipped  bool // marker already matched — nothing to do
+	Notes    int  // rows rewritten in the notes table
+	Facts    int  // fact rows rewritten in memory_vectors
+	MemNotes int  // note rows rewritten in memory_vectors
+	NoText   int  // memory_vectors rows with no text in their meta, left alone
+	Took     time.Duration
+}
+
+// ReembedAll rewrites every stored vector with the currently configured
+// embedder and then records that embedder as the one that owns the DB.
+//
+// Both places a vector lives are rewritten in the same pass: the `notes` table
+// `embedding` column and the `memory_vectors` rows (notes AND facts). Doing
+// only one would leave the two indexes disagreeing, which is worse than leaving
+// both stale.
+//
+// Safe to re-run: if the marker already names the current embedder there is
+// nothing to fix, so it returns immediately with Skipped set.
+//
+// Crash safety: everything — every vector and the marker — happens inside one
+// transaction. If anything fails or the process dies partway, the transaction
+// rolls back: no vectors changed and no marker written, so the next run does
+// the whole job again. The marker is never set unless the full rewrite
+// committed.
+func (s *Store) ReembedAll(ctx context.Context, currentID string, embed EmbedFunc) (BackfillResult, error) {
+	start := time.Now()
+	var res BackfillResult
+
+	stored, err := s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		return res, err
+	}
+	if stored == currentID {
+		res.Skipped = true
+		res.Took = time.Since(start)
+		return res, nil
+	}
+
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return res, fmt.Errorf("reembed: begin: %w", err)
+	}
+	defer tx.Rollback() // no-op once committed
+
+	// ----- notes table -----
+	type noteRow struct {
+		id   int64
+		text string
+	}
+	var notes []noteRow
+	rows, err := tx.QueryContext(ctx, `SELECT id, text FROM notes WHERE text != ''`)
+	if err != nil {
+		return res, fmt.Errorf("reembed: read notes: %w", err)
+	}
+	for rows.Next() {
+		var n noteRow
+		if err := rows.Scan(&n.id, &n.text); err != nil {
+			rows.Close()
+			return res, fmt.Errorf("reembed: note row: %w", err)
+		}
+		notes = append(notes, n)
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return res, fmt.Errorf("reembed: notes: %w", err)
+	}
+
+	for _, n := range notes {
+		vec, err := embed(ctx, n.text)
+		if err != nil {
+			return res, fmt.Errorf("reembed: embed note %d: %w", n.id, err)
+		}
+		if _, err := tx.ExecContext(ctx,
+			`UPDATE notes SET embedding = ? WHERE id = ?`, floatsToBlob(vec), n.id); err != nil {
+			return res, fmt.Errorf("reembed: write note %d: %w", n.id, err)
+		}
+		res.Notes++
+	}
+
+	// ----- memory_vectors (the unified index: notes AND facts) -----
+	// The text to re-embed is the one carried in the row's meta blob, which is
+	// exactly the text that was embedded when the row was written.
+	type vecRow struct {
+		id, text, kind string
+	}
+	var vecs []vecRow
+	rows, err = tx.QueryContext(ctx, `SELECT id, meta FROM memory_vectors`)
+	if err != nil {
+		return res, fmt.Errorf("reembed: read memory vectors: %w", err)
+	}
+	for rows.Next() {
+		var id, metaJSON string
+		if err := rows.Scan(&id, &metaJSON); err != nil {
+			rows.Close()
+			return res, fmt.Errorf("reembed: memory row: %w", err)
+		}
+		meta := map[string]string{}
+		if err := json.Unmarshal([]byte(metaJSON), &meta); err != nil {
+			rows.Close()
+			return res, fmt.Errorf("reembed: meta for %q: %w", id, err)
+		}
+		if meta["text"] == "" {
+			res.NoText++
+			continue
+		}
+		vecs = append(vecs, vecRow{id: id, text: meta["text"], kind: meta["type"]})
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return res, fmt.Errorf("reembed: memory vectors: %w", err)
+	}
+
+	for _, v := range vecs {
+		vec, err := embed(ctx, v.text)
+		if err != nil {
+			return res, fmt.Errorf("reembed: embed %q: %w", v.id, err)
+		}
+		if _, err := tx.ExecContext(ctx,
+			`UPDATE memory_vectors SET vec = ? WHERE id = ?`, encodeVec(vec), v.id); err != nil {
+			return res, fmt.Errorf("reembed: write %q: %w", v.id, err)
+		}
+		if v.kind == "fact" {
+			res.Facts++
+		} else {
+			res.MemNotes++
+		}
+	}
+
+	// Same transaction as the rewrite, on purpose: the marker can only exist if
+	// every vector above was written.
+	if _, err := tx.ExecContext(ctx,
+		`INSERT INTO meta (key, value) VALUES (?,?)
+		 ON CONFLICT(key) DO UPDATE SET value = excluded.value`,
+		metaKeyEmbedderID, currentID); err != nil {
+		return res, fmt.Errorf("reembed: write marker: %w", err)
+	}
+	if err := tx.Commit(); err != nil {
+		return res, fmt.Errorf("reembed: commit: %w", err)
+	}
+	res.Took = time.Since(start)
+	return res, nil
+}

internal/store/backfill_test.go

@@ -0,0 +1,171 @@
+package store
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+// markerVec is a recognisable vector: nothing in these tests writes it except
+// the backfill, so finding it proves the row really was rewritten.
+var markerVec = []float32{9, 9, 9}
+
+func newEmbedder(calls *int) EmbedFunc {
+	return func(_ context.Context, _ string) ([]float32, error) {
+		*calls++
+		return markerVec, nil
+	}
+}
+
+// seedOldVectors puts one note (notes table + unified index) and one fact
+// (unified index only) in the DB, both carrying obviously-old vectors.
+func seedOldVectors(t *testing.T, s *Store) {
+	t.Helper()
+	ctx := context.Background()
+	old := []float32{0.1, 0.2, 0.3}
+	id, err := s.WriteNote(ctx, time.Now(), "молоко в холодильнике", old, "voice")
+	if err != nil {
+		t.Fatalf("WriteNote: %v", err)
+	}
+	mem := s.VectorMemory()
+	if err := mem.Insert(ctx, "note:1", old, map[string]string{
+		"type": "note", "text": "молоко в холодильнике",
+	}); err != nil {
+		t.Fatalf("Insert note vector: %v", err)
+	}
+	if err := mem.Insert(ctx, "fact:water:1", old, map[string]string{
+		"type": "fact", "text": "я пил воду",
+	}); err != nil {
+		t.Fatalf("Insert fact vector: %v", err)
+	}
+	_ = id
+}
+
+func noteVec(t *testing.T, s *Store) []float32 {
+	t.Helper()
+	var blob []byte
+	if err := s.db.QueryRow(`SELECT embedding FROM notes LIMIT 1`).Scan(&blob); err != nil {
+		t.Fatalf("read note embedding: %v", err)
+	}
+	return blobToFloats(blob)
+}
+
+func memVec(t *testing.T, s *Store, id string) []float32 {
+	t.Helper()
+	var blob []byte
+	if err := s.db.QueryRow(`SELECT vec FROM memory_vectors WHERE id = ?`, id).Scan(&blob); err != nil {
+		t.Fatalf("read memory vector %s: %v", id, err)
+	}
+	return decodeVec(blob)
+}
+
+func sameVec(a, b []float32) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// The deployed case: old vectors everywhere, no marker. Every vector in both
+// places must be rewritten and the marker recorded.
+func TestReembedAllRewritesEveryVector(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedOldVectors(t, s)
+
+	calls := 0
+	res, err := s.ReembedAll(ctx, "multilingual-e5-small@384", newEmbedder(&calls))
+	if err != nil {
+		t.Fatalf("ReembedAll: %v", err)
+	}
+	if res.Skipped {
+		t.Fatal("first run should not skip")
+	}
+	if res.Notes != 1 || res.MemNotes != 1 || res.Facts != 1 {
+		t.Fatalf("counts: notes=%d memNotes=%d facts=%d", res.Notes, res.MemNotes, res.Facts)
+	}
+	if calls != 3 {
+		t.Fatalf("embedder called %d times, want 3", calls)
+	}
+	if !sameVec(noteVec(t, s), markerVec) {
+		t.Fatalf("notes table not rewritten: %v", noteVec(t, s))
+	}
+	if !sameVec(memVec(t, s, "note:1"), markerVec) {
+		t.Fatal("unified index note row not rewritten")
+	}
+	if !sameVec(memVec(t, s, "fact:water:1"), markerVec) {
+		t.Fatal("unified index fact row not rewritten")
+	}
+	got, err := s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		t.Fatalf("Meta: %v", err)
+	}
+	if got != "multilingual-e5-small@384" {
+		t.Fatalf("marker = %q", got)
+	}
+}
+
+// Re-running must do nothing at all — not a second pass over the same rows.
+func TestReembedAllSecondRunIsNoop(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedOldVectors(t, s)
+
+	calls := 0
+	if _, err := s.ReembedAll(ctx, "e5@384", newEmbedder(&calls)); err != nil {
+		t.Fatalf("first run: %v", err)
+	}
+	first := calls
+
+	res, err := s.ReembedAll(ctx, "e5@384", newEmbedder(&calls))
+	if err != nil {
+		t.Fatalf("second run: %v", err)
+	}
+	if !res.Skipped {
+		t.Fatal("second run should report Skipped")
+	}
+	if calls != first {
+		t.Fatalf("second run embedded %d more rows, want 0", calls-first)
+	}
+}
+
+// A failure partway must leave the DB exactly as it was: no marker, and the old
+// vectors still in place (one transaction, rolled back).
+func TestReembedAllPartialFailureLeavesMarkerUnset(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	seedOldVectors(t, s)
+	before := noteVec(t, s)
+
+	calls := 0
+	boom := func(_ context.Context, _ string) ([]float32, error) {
+		calls++
+		if calls == 2 {
+			return nil, errors.New("onnx blew up")
+		}
+		return markerVec, nil
+	}
+	if _, err := s.ReembedAll(ctx, "e5@384", boom); err == nil {
+		t.Fatal("expected an error")
+	}
+	got, err := s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		t.Fatalf("Meta: %v", err)
+	}
+	if got != "" {
+		t.Fatalf("marker was set to %q after a failed run", got)
+	}
+	if !sameVec(noteVec(t, s), before) {
+		t.Fatal("a failed run left a partially rewritten notes table")
+	}
+	// And the mismatch warning must still fire, so the user knows to re-run.
+	if _, mismatch, err := s.CheckEmbedder(ctx, "e5@384"); err != nil || !mismatch {
+		t.Fatalf("CheckEmbedder after failed backfill: mismatch=%v err=%v", mismatch, err)
+	}
+}

internal/store/delivery.go

@@ -13,11 +13,15 @@ import (
 // unknown = a pending row found stale at startup: the process that started it
 // is gone, and the send may or may not have reached the external channel.
 // Never auto-resolved into sent or failed — that would be guessing.
+// dropped = the routing table deliberately suppressed this one (a care nudge
+// while you're away). Nothing was sent and nothing went wrong; the row exists
+// so "she dropped it" and "the rule never fired" don't look the same later.
 const (
 	DeliveryPending = "pending"
 	DeliverySent    = "sent"
 	DeliveryFailed  = "failed"
 	DeliveryUnknown = "unknown"
+	DeliveryDropped = "dropped"
 )
 
 // BeginDeliveryAttempt durably records intent to send BEFORE the external
@@ -43,10 +47,11 @@ func (s *Store) BeginDeliveryAttempt(ctx context.Context, kind, rule string, rem
 }
 
 // CompleteDeliveryAttempt records the sink's outcome for a prior
-// BeginDeliveryAttempt. status is "sent" or "failed" — never "pending" or
-// "unknown" (those are set only by Begin and reconciliation respectively).
+// BeginDeliveryAttempt. status is "sent", "failed" or "dropped" — never
+// "pending" or "unknown" (those are set only by Begin and reconciliation
+// respectively).
 func (s *Store) CompleteDeliveryAttempt(ctx context.Context, id int64, status string, now time.Time) error {
-	if status != DeliverySent && status != DeliveryFailed {
+	if status != DeliverySent && status != DeliveryFailed && status != DeliveryDropped {
 		return fmt.Errorf("store: invalid delivery completion status %q", status)
 	}
 	_, err := s.db.ExecContext(ctx,

internal/store/delivery_test.go

@@ -0,0 +1,34 @@
+package store
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+// TestDroppedDeliveryAttemptRoundTrips — Vikunja #370. A suppressed nudge is
+// recorded as 'dropped'. The status column has a CHECK constraint, so this
+// only works if migration #12 widened it; a fake outbox in a unit test would
+// not catch that.
+func TestDroppedDeliveryAttemptRoundTrips(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+	now := time.Now()
+
+	id, err := s.BeginDeliveryAttempt(ctx, "nudge", "water", 0, "drop", "abc123", now)
+	if err != nil {
+		t.Fatalf("BeginDeliveryAttempt: %v", err)
+	}
+	if err := s.CompleteDeliveryAttempt(ctx, id, DeliveryDropped, now); err != nil {
+		t.Fatalf("CompleteDeliveryAttempt: %v", err)
+	}
+
+	var status string
+	err = s.db.QueryRowContext(ctx, `SELECT status FROM delivery_attempts WHERE id = ?`, id).Scan(&status)
+	if err != nil {
+		t.Fatalf("read back: %v", err)
+	}
+	if status != DeliveryDropped {
+		t.Fatalf("status: want %q, got %q", DeliveryDropped, status)
+	}
+}

internal/store/dialogue.go

@@ -0,0 +1,74 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"time"
+)
+
+// DialogueSessionRow — one saved follow-up session. Data is the session
+// encoded by the dialogue package; the store does not look inside it.
+type DialogueSessionRow struct {
+	ID      string
+	Data    []byte
+	Ts      time.Time
+	TTL     time.Duration
+	Expires time.Time
+}
+
+// SaveDialogueSession — write (or replace) the session for one dialogue id.
+// One row per id: a newer turn overwrites the older state.
+func (s *Store) SaveDialogueSession(ctx context.Context, id string, data []byte, ts time.Time, ttl time.Duration) error {
+	expires := ts.Add(ttl)
+	_, err := s.db.ExecContext(ctx, `
+		INSERT INTO dialogue_sessions (id, data, ts, ttl_ms, expires_ts) VALUES (?, ?, ?, ?, ?)
+		ON CONFLICT(id) DO UPDATE SET data       = excluded.data,
+		                              ts         = excluded.ts,
+		                              ttl_ms     = excluded.ttl_ms,
+		                              expires_ts = excluded.expires_ts`,
+		id, data, ts.UnixMilli(), ttl.Milliseconds(), expires.UnixMilli())
+	if err != nil {
+		return fmt.Errorf("save dialogue session: %w", err)
+	}
+	return nil
+}
+
+// DeleteDialogueSession — drop one session (ended, or expired).
+func (s *Store) DeleteDialogueSession(ctx context.Context, id string) error {
+	if _, err := s.db.ExecContext(ctx, `DELETE FROM dialogue_sessions WHERE id = ?`, id); err != nil {
+		return fmt.Errorf("delete dialogue session: %w", err)
+	}
+	return nil
+}
+
+// LoadDialogueSessions — return the sessions still alive at `now` and delete
+// the ones that already ran out. An expired session is dead: it never comes
+// back after a restart.
+func (s *Store) LoadDialogueSessions(ctx context.Context, now time.Time) ([]DialogueSessionRow, error) {
+	if _, err := s.db.ExecContext(ctx,
+		`DELETE FROM dialogue_sessions WHERE expires_ts <= ?`, now.UnixMilli()); err != nil {
+		return nil, fmt.Errorf("prune dialogue sessions: %w", err)
+	}
+	rows, err := s.db.QueryContext(ctx,
+		`SELECT id, data, ts, ttl_ms, expires_ts FROM dialogue_sessions ORDER BY id`)
+	if err != nil {
+		return nil, fmt.Errorf("load dialogue sessions: %w", err)
+	}
+	defer rows.Close()
+	var out []DialogueSessionRow
+	for rows.Next() {
+		var r DialogueSessionRow
+		var tsMilli, ttlMilli, expMilli int64
+		if err := rows.Scan(&r.ID, &r.Data, &tsMilli, &ttlMilli, &expMilli); err != nil {
+			return nil, fmt.Errorf("scan dialogue session: %w", err)
+		}
+		r.Ts = time.UnixMilli(tsMilli).UTC()
+		r.TTL = time.Duration(ttlMilli) * time.Millisecond
+		r.Expires = time.UnixMilli(expMilli).UTC()
+		out = append(out, r)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("load dialogue sessions: %w", err)
+	}
+	return out, nil
+}

internal/store/meta.go

@@ -0,0 +1,97 @@
+package store
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+)
+
+// metaKeyEmbedderID names the embedder that wrote the stored vectors.
+//
+// Why one value for the whole DB and not a column on every vector row: the
+// vectors are only ever rewritten all at once (one backfill re-embeds every
+// note and fact together), so a per-row marker would hold the same string in
+// every row and cost a column on two tables for nothing.
+const metaKeyEmbedderID = "embedder_id"
+
+// Meta reads a single value from the meta table. Missing key ⇒ empty string.
+func (s *Store) Meta(ctx context.Context, key string) (string, error) {
+	var v string
+	err := s.db.QueryRowContext(ctx, `SELECT value FROM meta WHERE key = ?`, key).Scan(&v)
+	if errors.Is(err, sql.ErrNoRows) {
+		return "", nil
+	}
+	if err != nil {
+		return "", fmt.Errorf("read meta %s: %w", key, err)
+	}
+	return v, nil
+}
+
+// SetMeta writes (or overwrites) a single meta value.
+func (s *Store) SetMeta(ctx context.Context, key, value string) error {
+	_, err := s.db.ExecContext(ctx,
+		`INSERT INTO meta (key, value) VALUES (?,?)
+		 ON CONFLICT(key) DO UPDATE SET value = excluded.value`, key, value)
+	if err != nil {
+		return fmt.Errorf("write meta %s: %w", key, err)
+	}
+	return nil
+}
+
+// EmbedderUnknown is the stored id reported for a DB that already holds
+// vectors but never recorded who wrote them.
+const EmbedderUnknown = "unknown (written before this marker existed)"
+
+// CheckEmbedder compares the embedder now configured against the one that
+// wrote the stored vectors. Returns the stored id and whether it differs.
+//
+// Vectors from two different models live in different spaces, so cosine
+// between them is noise rather than a low score — and both of our models are
+// 384-dimensional, so nothing else catches it.
+//
+// Three cases, and the middle one is the one that actually matters:
+//
+//   - marker present ⇒ compare the two ids.
+//   - marker absent but vectors already stored ⇒ this is a DB from before the
+//     marker, so we cannot know who wrote them. Report a mismatch. This is the
+//     real case on the deployed box: those vectors came from the old embedder,
+//     and claiming them for the current one would hide the exact problem the
+//     marker was added to catch.
+//   - marker absent and no vectors ⇒ fresh DB, claim it, nothing to fix.
+//
+// On a mismatch the fix is ReembedAll (backfill.go), run explicitly with
+// `mavend -reembed`. Nothing is re-embedded here: that work is minutes of CPU
+// on the laptop and must not stall a normal start.
+func (s *Store) CheckEmbedder(ctx context.Context, currentID string) (stored string, mismatch bool, err error) {
+	stored, err = s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		return "", false, err
+	}
+	if stored != "" {
+		return stored, stored != currentID, nil
+	}
+	n, err := s.countVectors(ctx)
+	if err != nil {
+		return "", false, err
+	}
+	if n > 0 {
+		return EmbedderUnknown, true, nil
+	}
+	return currentID, false, s.SetMeta(ctx, metaKeyEmbedderID, currentID)
+}
+
+// countVectors — how many stored rows carry an embedding. Used only to tell a
+// fresh DB apart from one that predates the marker.
+func (s *Store) countVectors(ctx context.Context) (int, error) {
+	var notes, vecs int
+	if err := s.db.QueryRowContext(ctx,
+		`SELECT count(*) FROM notes WHERE embedding IS NOT NULL`).Scan(&notes); err != nil {
+		return 0, fmt.Errorf("count note vectors: %w", err)
+	}
+	if err := s.db.QueryRowContext(ctx,
+		`SELECT count(*) FROM memory_vectors`).Scan(&vecs); err != nil {
+		return 0, fmt.Errorf("count memory vectors: %w", err)
+	}
+	return notes + vecs, nil
+}

internal/store/meta_test.go

@@ -0,0 +1,100 @@
+package store
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+// A fresh DB has no marker yet, so the current embedder is recorded and
+// nothing is flagged.
+func TestCheckEmbedderFreshDBRecords(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+
+	stored, mismatch, err := s.CheckEmbedder(ctx, "multilingual-e5-small@384")
+	if err != nil {
+		t.Fatalf("CheckEmbedder: %v", err)
+	}
+	if mismatch {
+		t.Fatal("fresh DB reported a mismatch")
+	}
+	if stored != "multilingual-e5-small@384" {
+		t.Fatalf("stored = %q", stored)
+	}
+	got, err := s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		t.Fatalf("Meta: %v", err)
+	}
+	if got != "multilingual-e5-small@384" {
+		t.Fatalf("marker not persisted, got %q", got)
+	}
+}
+
+// The deployed box: notes were written by the old embedder, before the marker
+// existed. Claiming them for the current one would hide exactly the problem
+// the marker is for, so an unmarked DB that already holds vectors is a
+// mismatch.
+func TestCheckEmbedderUnmarkedDBWithVectorsIsMismatch(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+
+	if _, err := s.WriteNote(ctx, time.Now(), "молоко в холодильнике", []float32{0.1, 0.2}, "voice"); err != nil {
+		t.Fatalf("WriteNote: %v", err)
+	}
+
+	stored, mismatch, err := s.CheckEmbedder(ctx, "multilingual-e5-small@384")
+	if err != nil {
+		t.Fatalf("CheckEmbedder: %v", err)
+	}
+	if !mismatch {
+		t.Fatal("an unmarked DB with stored vectors should report a mismatch")
+	}
+	if stored != EmbedderUnknown {
+		t.Fatalf("stored = %q, want %q", stored, EmbedderUnknown)
+	}
+	// It must NOT claim the DB — that would silence the warning on restart.
+	got, err := s.Meta(ctx, metaKeyEmbedderID)
+	if err != nil {
+		t.Fatalf("Meta: %v", err)
+	}
+	if got != "" {
+		t.Fatalf("marker written despite unknown provenance: %q", got)
+	}
+}
+
+// Both models are 384-dim, so this is the only thing that catches the swap.
+func TestCheckEmbedderDifferentModelMismatch(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+
+	if err := s.SetMeta(ctx, metaKeyEmbedderID, "paraphrase-multilingual-MiniLM-L12-v2@384"); err != nil {
+		t.Fatalf("SetMeta: %v", err)
+	}
+	stored, mismatch, err := s.CheckEmbedder(ctx, "multilingual-e5-small@384")
+	if err != nil {
+		t.Fatalf("CheckEmbedder: %v", err)
+	}
+	if !mismatch {
+		t.Fatal("different embedder not detected")
+	}
+	if stored != "paraphrase-multilingual-MiniLM-L12-v2@384" {
+		t.Fatalf("stored = %q", stored)
+	}
+}
+
+// The same embedder must never raise a false alarm, including on re-check.
+func TestCheckEmbedderSameModelNoAlarm(t *testing.T) {
+	s := newTestStore(t)
+	ctx := context.Background()
+
+	for i := 0; i < 2; i++ {
+		_, mismatch, err := s.CheckEmbedder(ctx, "multilingual-e5-small@384")
+		if err != nil {
+			t.Fatalf("CheckEmbedder: %v", err)
+		}
+		if mismatch {
+			t.Fatalf("false alarm on pass %d", i)
+		}
+	}
+}

internal/store/migrations.go

@@ -74,6 +74,44 @@ ALTER TABLE reminders ADD COLUMN next_fire_ts INTEGER;`, // #2
 	`CREATE INDEX IF NOT EXISTS idx_nudges_snoozed ON nudges (outcome_ts) WHERE outcome = 'snoozed';`, // #8 — SnoozedUntil runs every tick; keep it off a full scan (Vikunja #364)
 	`ALTER TABLE proposed_routines ADD COLUMN accepted_ts INTEGER;
 	ALTER TABLE proposed_routines ADD COLUMN last_fired_ts INTEGER;`, // #9 — accepted routines keep firing (Vikunja #366): the tick loop needs to know when a routine was accepted and when it last nudged
+
+	`CREATE TABLE IF NOT EXISTS dialogue_sessions (
+		id         TEXT PRIMARY KEY,
+		data       BLOB NOT NULL,
+		ts         INTEGER NOT NULL,
+		ttl_ms     INTEGER NOT NULL,
+		expires_ts INTEGER NOT NULL
+	);
+	CREATE INDEX IF NOT EXISTS idx_dialogue_sessions_expires ON dialogue_sessions (expires_ts);`, // #10 — the follow-up session survives a restart (Vikunja #363); small, TTL-pruned table, not a history log
+
+	`CREATE TABLE IF NOT EXISTS meta (
+		key   TEXT PRIMARY KEY,
+		value TEXT NOT NULL
+	);`, // #11 — small key/value table for facts about the DB itself; first key is embedder_id (Vikunja #378)
+
+	// #12 — a suppressed nudge gets a 'dropped' row (Vikunja #370). sqlite
+	// can't widen a CHECK constraint in place, so the table is rebuilt; the
+	// index goes with the old table and is recreated. The columns are listed
+	// out rather than `SELECT *` — copying by position would silently shuffle
+	// every row if the old table's column order ever differed from this one.
+	`CREATE TABLE delivery_attempts_v12 (
+		id           INTEGER PRIMARY KEY AUTOINCREMENT,
+		kind         TEXT NOT NULL CHECK (kind IN ('nudge','reminder')),
+		rule         TEXT NOT NULL DEFAULT '',
+		reminder_id  INTEGER NOT NULL DEFAULT 0,
+		channel      TEXT NOT NULL,
+		body_hash    TEXT NOT NULL,
+		status       TEXT NOT NULL DEFAULT 'pending' CHECK (status IN ('pending','sent','failed','unknown','dropped')),
+		created_ts   INTEGER NOT NULL,
+		completed_ts INTEGER
+	);
+	INSERT INTO delivery_attempts_v12
+		(id, kind, rule, reminder_id, channel, body_hash, status, created_ts, completed_ts)
+		SELECT id, kind, rule, reminder_id, channel, body_hash, status, created_ts, completed_ts
+		FROM delivery_attempts;
+	DROP TABLE delivery_attempts;
+	ALTER TABLE delivery_attempts_v12 RENAME TO delivery_attempts;
+	CREATE INDEX IF NOT EXISTS idx_delivery_attempts_status ON delivery_attempts (status);`,
 }
 
 // migrate applies every migration with a number greater than the DB's current

kill-maven.sh

@@ -1,8 +1,10 @@
 #!/usr/bin/env bash
 # Unified script to stop all Maven services.
 # Usage: ./kill-maven.sh
-#   - Graceful SIGTERM is attempted first.
-#   - If any process lingers, force with SIGKILL.
+#   - Docker deploy: `docker compose stop` (see why below).
+#   - Bare-metal / dev run: graceful SIGTERM first, SIGKILL if anything lingers.
+# Exits non-zero if it cannot confirm everything is stopped. It must never say
+# "stopped" unless it checked.
 
 set -euo pipefail
 
@@ -24,6 +26,73 @@ else
     LLM='llama-server.*\.gguf'
 fi
 
+COMPOSE_FILE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/docker-compose.yml"
+
+# --- containerised deploy ------------------------------------------------
+# docker-compose.yml does not set `pid: host`, so each container has its own
+# PID namespace: pkill on the host sees nothing inside them. This script used
+# to print "all stopped" while every daemon was still happily running. Stop the
+# containers through compose instead — that actually reaches them.
+#
+# running_containers prints the ids of the project's running containers, or
+# nothing. Empty output plus a non-zero return means "could not ask docker",
+# which is different from "nothing is running" and is handled below.
+running_containers() {
+    docker compose -f "$COMPOSE_FILE" ps -q --status running 2>/dev/null
+}
+
+DOCKER_OK=0
+CONTAINERS=""
+if command -v docker >/dev/null 2>&1 && [ -f "$COMPOSE_FILE" ]; then
+    if CONTAINERS="$(running_containers)"; then
+        DOCKER_OK=1
+    fi
+fi
+
+if [ "$DOCKER_OK" = 1 ] && [ -n "$CONTAINERS" ]; then
+    echo "--- Maven is running in containers: stopping via docker compose ---"
+    if ! docker compose -f "$COMPOSE_FILE" stop; then
+        echo "ERROR: 'docker compose stop' failed. Containers may still be running." >&2
+        exit 1
+    fi
+    echo "--- Verifying containers are gone ---"
+    LEFT="$(running_containers || true)"
+    if [ -n "$LEFT" ]; then
+        echo "ERROR: containers still running after stop:" >&2
+        docker compose -f "$COMPOSE_FILE" ps >&2 || true
+        exit 1
+    fi
+    echo "All containers stopped."
+    exit 0
+fi
+
+# --- bare-metal / dev run -----------------------------------------------
+# pgrep -f matches whole command lines, so a shell that merely mentions
+# "mavend" (this script's own parent, for one) shows up. Drop ourselves and our
+# parent, otherwise the SIGKILL sweep can take out the terminal you ran this in.
+host_pids() {
+    pgrep -f "$PAT|$LLM" | grep -v -e "^$$\$" -e "^$PPID\$" | paste -sd, - || true
+}
+HOST_PIDS=$(host_pids)
+
+if [ -z "$HOST_PIDS" ]; then
+    # Nothing on the host. Whether that means "already down" depends on whether
+    # we managed to ask docker, and the two must not read the same.
+    if [ "$DOCKER_OK" = 1 ]; then
+        # Docker answered and named no running containers, and there is nothing
+        # on the host either. That is a real answer: Maven is already stopped.
+        echo "Nothing to stop: no Maven processes and no running containers."
+        exit 0
+    fi
+    # We could not ask docker, so Maven may be alive in a container we cannot
+    # see. Saying "stopped" here is the exact false success this script had.
+    echo "ERROR: no Maven processes on this host, and docker could not be asked." >&2
+    echo "       If this is the container deploy it may still be running:" >&2
+    echo "       docker compose -f $COMPOSE_FILE stop" >&2
+    echo "       Nothing was stopped. Check by hand before assuming Maven is down." >&2
+    exit 1
+fi
+
 echo "--- Sending graceful SIGTERM to Maven services ---"
 pkill -TERM -f "$PAT" || true
 # mavend's Pdeathsig SIGKILLs its llama-server on exit, but sweep strays too
@@ -32,12 +101,18 @@ pkill -TERM -f "$LLM" || true
 
 echo "--- Verifying processes are gone ---"
 sleep 1
-PIDS=$(pgrep -d ',' -f "$PAT|$LLM") || PIDS=""
+PIDS=$(host_pids)
 if [ -n "$PIDS" ]; then
     echo "Warning: some processes still alive. PIDs: $PIDS"
     echo "--- Force killing with SIGKILL ---"
     echo "$PIDS" | tr ',' '\n' | xargs -r kill -9
+    sleep 1
+    LEFT=$(host_pids)
+    if [ -n "$LEFT" ]; then
+        echo "ERROR: still alive after SIGKILL. PIDs: $LEFT" >&2
+        exit 1
+    fi
     echo "Done (SIGKILL)."
 else
     echo "All services gracefully stopped."
-fi
+fi