Record what Kiwix returns and why the gate is not one (V-668)

gofmt on cmd/mavwaked/silero.go came in with a99932b and blocked make test.

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

.gitignore

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

AGENTS.md

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

CLAUDE.md

@@ -47,6 +47,31 @@ free — `worldGap` in `cmd/mavend/worldmodel.go`, which the owner hears instead
 answer. A box with no `workstation` block behaves exactly as it did before the seam: naming
 a gap requires a gap. The offload table in `docs/offload.md` says which caller is which.
 
+**Speech-to-text moved on 2026-08-09** (V-486). `sttSeam` in `cmd/mavend/voicewire.go`
+builds an `stt.Pair` beside `modelSeam`, preferring CrisperWhisper 2.0 turbo on workpc
+with mavsttd as the floor. It takes only the silent half of the rule. A worse
+transcript is still a turn, so `stt.Pair` has no `TranscribeRemote`. The fallback is
+never spoken. CW2 turbo scores **10.4% WER in Russian against 27.5%** for the `ggml-small.bin`
+mavsttd loads, over 200 Golos clips
+(`docs/evals/2026-08-09-crisperwhisper2-russian-wer.md`). It runs in Intended mode, not
+Verbatim, though that corpus cannot separate the two.
+**whisper.cpp cannot load CW2 at all.** It reads its language count off the vocabulary
+size, and CW2's 51897 tokens shift seven special token ids. So it is not a second
+endpoint on mavgpud. It is its own transformers service on port 8081
+(`deploy/cw2/serve.py`), which Maven reaches directly. `stt.HTTPTranscriber`
+posts raw PCM to it with a bearer token, because audio is the most sensitive thing that
+crosses this seam. The switch is `workstation.stt` in
+`deploy/mavend.json`, and deleting the block sends every utterance to mavsttd.
+**mavgpud runs that service as a second child.** That is not an optimisation. CW2 is a
+ROCm process on the same card, so it registers on the KFD like any contender. Under its own
+systemd unit it made mavgpud evict llama-server every few seconds. That took the
+gemma-4-12b arm down for eight minutes on 2026-08-09 before anyone noticed. The card needs
+one owner. Any GPU service added beside this daemon has the same defect, so add it to
+`cmd/mavgpud` and not to systemd. CW2 is on the yield clock and not the idle one. At 1.6GB
+it denies the card to nobody, and unloading it would only send the next voice turn to the
+homesrv floor.
+Text-to-speech has not moved and piper on homesrv is still the only synthesizer.
+
 ## Build & test
 
 CGO daemons (`mavend`, `mavsttd`, `mavttsd`, `mavenclient`) need the vendored toolchain
@@ -230,11 +255,21 @@ re-run it, start a **second** llama-server on a fixed host port — the resident
 `--port 0` inside the container and no host process can reach it.
 
 **The numbers above are the homesrv floor, not the ceiling.** With the workstation up, routing
-completes through `llm.Pair` against gemma-4-12b and scores **84.4% full / 93.5% intent-only at
-p50 329ms** — better than the resident model and about 2.5× faster (`docs/evals/2026-08-02-workstation-gemma4-12b.md`,
-Vikunja #485). The workstation is never assumed up, so both sets of numbers are live. Judge a
+completes through `llm.Pair` against the model mavgpud holds, which is better than the resident
+model and about 2.5× faster. gemma-4-12b scored **84.4% full / 93.5% intent-only at p50 329ms**
+(`docs/evals/2026-08-02-workstation-gemma4-12b.md`, Vikunja #485). The workstation is never
+assumed up, so both sets of numbers are live. Judge a
 routing change against the classifier and the resident model, since those are what always answer.
 
+**The workstation runs gemma-4-E4B since 2026-08-09** (owner's call), and it is a
+step down measured the same day (`docs/evals/2026-08-09-e4b-vs-12b-routing.md`).
+Against a same-session 12B control it scores **83.3% full / 89.6% intent-only,
+destination 19/33 against 23/33, at p50 294ms against 344ms**. So it costs four
+destination cases and buys 50ms. Read destination as the finding: it names nothing
+where the 12B names `recall` or `calendar`, which is safe but walks the whole chain.
+It also has no MTP and cannot be given any here. The only `gemma4-assistant`
+draft on disk is trained against the 12B's hidden states.
+
 **The intended third engine is not a generative model** (owner's call, 05-08-2026, V-546,
 `docs/plans/18-routing-heads-on-e5-small.md`). Routing has a bounded output space, so it is
 classification, and the 118M multilingual-e5-small is already resident. Three heads on one
@@ -308,9 +343,36 @@ the possessive agenda rules claim those cases at stage 0 and name nothing, so no
 label reaches the head. That is the same trade V-660 flagged and it wants the
 owner's call.
 
-**Nothing of this runs in Go.** The weights are `heads.pt` and `out/body_heads/`
-on workpc. Reaching the daemon needs an ONNX export and a caller. The resident
-e5-small must not be replaced by the copy, because recall depends on that file.
+**The heads run in Go and route every turn, since 08-08-2026** (V-664,
+`docs/evals/2026-08-08-routing-heads-in-go.md`). This section used to say
+nothing of it ran. `RouterHeads` in `internal/router/heads.go` loads
+`router_heads.onnx` and reads intent, destination and clarify off one forward
+pass. It is stage 0b: after the grammars, **before** the resident model, and the
+classifier is still behind both. Through the cascade it scores intent **96.9%**
+and destination **75.8%** at p50 27.9ms. That beats the gemma-4-12b cascade,
+84.4% and 72.7%, at a twelfth of its 329ms. The workstation stays the better
+phraser and is no longer the better router.
+
+Three rules around it. The **clarify head decides first**, before the intent
+threshold. It answers a different question. A thin utterance scores low
+intent by construction, so gating it cost 6 of 8 ambiguous cases. The
+**destination head is read on `IntentQuery` only**, since no other intent
+reaches `queryWalk`. And `headsThreshold` is 0.6, the measured knee: every value
+to 0.85 drops right answers and keeps the same two wrong ones.
+
+`voice.embedder.heads_path` is the whole switch. Empty, missing or unloadable
+means the heads are nil and the cascade is byte-for-byte what shipped before
+them. **It must never be pointed at `model_path`.** The resident e5-small must
+not be replaced by the fine-tuned copy. Recall depends on that file scoring
+what it scored.
+
+**The hand-written tokenizer read every long word backwards** until this task
+(`encodeWord`, `onnxembedder.go`). It cost recall@1 7.4 points and recall@3 11.1.
+Nothing caught it because seeds and queries were mangled the same way, so cosine
+survived. The heads found it. They are trained through transformers and read
+through this. The embedder id now carries a tokenizer revision
+(`@384/tok2`), so fixing the tokenizer triggers `ReembedAll` the way swapping the
+model file does. Bump `tokenizerRev` on any change to what it emits.
 
 `Confidence: 1.0` used to be hardcoded in `llmrouter.go`, so the LLM
 path could never ask for clarification (6/6 refusal cases missed on the fixture) — Vikunja
@@ -448,13 +510,21 @@ they look rather than guess.
 **The personal boundary is the one exception and it is deliberate.** It guesses,
 so naming `SourceWorld` drops it. That is what stops it answering "кто такой
 Линус Торвальдс?" with "не нашла у тебя такой записи", which it did on
-2026-08-07. The cost is that a destination a model wrote can now take the
-boundary off a turn. A question about him that the model calls `world` reaches
-SearXNG, where today the boundary stops it. Only the utterance leaves the box,
-never his notes or history, so this widens what is asked and not what is sent.
-`TestNamingRecallKeepsTheBoundary` pins the other half: naming `SourceRecall`
-keeps the boundary in front of the world. Whether a model may drop it at all is
-the owner's call and has not been made.
+2026-08-07. `TestNamingRecallKeepsTheBoundary` pins the other half: naming
+`SourceRecall` keeps the boundary in front of the world.
+
+**Only a stage 0 grammar may drop it** (owner's call, 09-08-2026, V-666). The
+question of who is allowed to was open until then. Three deciders name a
+destination and two of them infer it: the routing heads and the resident model.
+An inferred `SourceWorld` on a question about him would reach SearXNG, and that
+widens what is asked rather than costing a local answer. So `Decision.SourceAnchored`
+carries the provenance. It is a field and not `Stage == 0`. Stage 0 also means
+confidence 1.0 and an anchored claim band, and one of those could stop implying
+the others. `queryWalk` reads it for the source marked `boundary: true` and for
+no other. So every other guesser still comes off the turn, whoever named the
+destination. `TestOnlyAGrammarMayDropTheBoundary` pins both directions.
+`definitionQueryPattern` claims "кто такой X", so the 2026-08-07 case is still
+anchored and still answered.
 
 What comes out is only the sources that **guess**. Those decide a turn is theirs by
 cosine against frozen seeds, then answer whatever they claimed. They hold no table
@@ -605,11 +675,25 @@ world questions, so she needs to read external sources. What replaces it:
   `wikipedia_ru_all_maxi_2026-02` verbatim** through `kiwix.book_ru`. The RU→EN rewriter
   is the workaround for an English book and is skipped there. Kiwix catalog names come
   from the filename, not the `<name>` field.
+  **That verbatim path sent the whole sentence to a keyword engine until 09-08-2026**
+  (V-668, `docs/evals/2026-08-09-kiwix-topic-retrieval.md`). Kiwix ranks by keyword
+  overlap, so the question words outrank the one word naming the article. "что такое TCP"
+  returned "Перехват TCP-соединения". "кто написал Войну и мир" returned an episode of
+  Doctor Who. `kiwix.Topic` drops the narrative request, the interrogative and a verb
+  behind one. `kiwix.TitlePath` tries the exact article first, since a ZIM is addressable
+  by title and a wrong title is a 404. Four of eight questions reach the right article
+  where they did not, two were already right, and nothing regressed. Both apply on the
+  verbatim path alone. The rewriter already reduces a question, and reducing twice takes
+  the topic off its input.
   `Response.Empty()` is the whole gate and there is no quality threshold in front of it:
   the three signals one could read were measured on 2026-08-05 and none of them separate a
   real question from an invented one. Token overlap would cost "столица Франции" its
   answer, because the answer is Париж and that word is not in the question. See
-  `docs/evals/2026-08-05-search-quality-signals.md` (V-539). **Which query source claimed
+  `docs/evals/2026-08-05-search-quality-signals.md` (V-539). **The embedder is not a
+  fourth signal**, measured 2026-08-09 (V-668). Query-to-passage cosine scores 0.79 to
+  0.91 on answerable questions and 0.75 to 0.84 on unanswerable ones, and the sets
+  overlap. The wrong TCP article scored 0.8653, above five of six unanswerable rows. It
+  measures topic and not whether the passage answers, so no threshold splits them. **Which query source claimed
   a turn is readable on `/chat`** as a badge beside the reply, carried on
   `ipc.ChatReply.Source` and noted by `noteQuerySource` in `cmd/mavend/querysource.go`. It
   rides the context, so `handleText` keeps the one string signature the mic, telegram and

cmd/mavend/actions_query.go

@@ -13,6 +13,7 @@ import (
 	"github.com/kami/maven/internal/crawl"
 	"github.com/kami/maven/internal/decision"
 	"github.com/kami/maven/internal/ipc"
+	"github.com/kami/maven/internal/kiwix"
 	"github.com/kami/maven/internal/memory"
 	"github.com/kami/maven/internal/morning"
 	"github.com/kami/maven/internal/phraser"
@@ -79,6 +80,15 @@ type querySource struct {
 	// not named do not get to try. The lookups still run, because a named
 	// destination is evidence and not a promise.
 	guesses bool
+
+	// boundary — dropping this source widens what leaves the box, so only a
+	// literal pattern may do it (V-666, owner's call of 2026-08-09).
+	//
+	// Every other guesser costs an answer when it is wrongly taken off a turn.
+	// This one costs the rule that a question about him never reaches an
+	// upstream engine. A grammar read the words to name a destination. A model
+	// and a softmax both inferred one, and neither may spend that.
+	boundary bool
 }
 
 // querySources is the ordered chain actionQuery walks; first source to claim
@@ -156,7 +166,7 @@ var querySources = []querySource{
 	// below answers from the world's. A question about him that got this far
 	// has no answer in his data, and no outside source can supply one, so this
 	// stops the walk rather than let the encyclopedia and the model guess.
-	{name: "personal", answer: (*reactiveHandler).queryPersonal, dest: router.SourceRecall, guesses: true},
+	{name: "personal", answer: (*reactiveHandler).queryPersonal, dest: router.SourceRecall, guesses: true, boundary: true},
 	// The world, read live. Owner's ruling of 2026-08-02: a metasearch hit beats
 	// a frozen ZIM, so SearXNG asks before Kiwix does. Nothing of his is at
 	// stake by this point — the boundary above already stopped every question
@@ -198,12 +208,17 @@ var querySources = []querySource{
 // No destination named ⇒ the table exactly as written, which is what shipped
 // before the field existed. That is the floor. The classifier arm names
 // nothing, so a box whose model is down routes queries the way it always did.
-func queryWalk(dest router.Source) (walk, skipped []querySource) {
+// The personal boundary is the one exception, and anchored is what buys it
+// (V-666). A grammar matched a literal pattern to name the destination. The
+// routing heads and the resident model inferred one, and an inferred SourceWorld
+// takes the boundary off a question about him. That widens what is asked
+// upstream rather than costing a local answer, so those two keep it.
+func queryWalk(dest router.Source, anchored bool) (walk, skipped []querySource) {
 	if dest == router.SourceUnknown {
 		return querySources, nil
 	}
 	for _, s := range querySources {
-		if s.guesses && s.dest != dest {
+		if s.guesses && s.dest != dest && (anchored || !s.boundary) {
 			skipped = append(skipped, s)
 			continue
 		}
@@ -219,7 +234,7 @@ func (h *reactiveHandler) actionQuery(ctx context.Context, dec router.Decision)
 	// (V-564). Finish names everyone below the winner.
 	decision.Expect(ctx, decision.StageQuery, querySourceNames())
 	rec := decision.From(ctx)
-	walk, skipped := queryWalk(dec.Source)
+	walk, skipped := queryWalk(dec.Source, dec.SourceAnchored)
 	for _, src := range skipped {
 		rec.Note(decision.Claim{
 			Stage: decision.StageQuery, Claimant: src.name, Outcome: decision.NeverAsked,
@@ -912,6 +927,26 @@ func (h *reactiveHandler) queryKiwix(ctx context.Context, t *queryTurn) (string,
 		}
 	}
 
+	// The topic, not the sentence (V-668). Kiwix ranks by keyword overlap, so
+	// the question words outrank the one word that names the article: measured
+	// on 2026-08-09, "что такое TCP" returns "Перехват TCP-соединения" and
+	// "TCP" returns TCP. Only the verbatim path needs this. The rewriter
+	// already reduces a question to English keywords, and reducing twice would
+	// take the topic off the input it reads.
+	if verbatim {
+		if topic := kiwix.Topic(pattern); topic != "" {
+			// The article named exactly, before any ranking runs. A ZIM is
+			// addressable by title and a wrong title is a 404, so this either
+			// answers or costs one request that says nothing.
+			page, err := h.kiwix.client.Article(ctxK, kiwix.TitlePath(book, topic), h.kiwix.runes)
+			if err == nil && page.Text != "" {
+				log.Printf("voice: kiwix: %q in %q → title hit %q", topic, book, page.Title)
+				return h.kiwixReply(ctx, t, page.Title, page.Text)
+			}
+			pattern = topic
+		}
+	}
+
 	hits, err := h.kiwix.client.Search(ctxK, pattern, book, h.kiwix.max)
 	if err != nil {
 		log.Printf("voice: kiwix: search %q: %v", pattern, err)
@@ -944,14 +979,18 @@ func (h *reactiveHandler) queryKiwix(ctx context.Context, t *queryTurn) (string,
 		}
 		page = crawl.Page{Title: top.Title, Text: top.Snippet}
 	}
-	// Handed over the same way a note or a page is: context for the question he
-	// asked, not something to recite.
-	snippet := top.Title + "\n" + crawl.TrimRunes(page.Text, h.kiwix.runes)
+	return h.kiwixReply(ctx, t, top.Title, page.Text)
+}
+
+// kiwixReply hands one article over the same way a note or a page is handed
+// over: context for the question he asked, not something to recite.
+func (h *reactiveHandler) kiwixReply(ctx context.Context, t *queryTurn, title, text string) (string, bool) {
+	snippet := title + "\n" + crawl.TrimRunes(text, h.kiwix.runes)
 	reply := h.phraseSource(ctx, "kiwix", t.dec.Utterance, []string{snippet})
 	if reply == "" {
 		// No phraser, or it failed. Read back the best hit rather than pretend
 		// the search did not happen.
-		return readBack(top.Title + " — " + page.Text), true
+		return readBack(title + " — " + text), true
 	}
 	return reply, true
 }

cmd/mavend/chat_degrade_test.go

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

cmd/mavend/clarify_test.go

@@ -317,7 +317,7 @@ func TestClarifyExpiryIsAnnouncedAndWordsStillRoute(t *testing.T) {
 	h, _, now := newClarifyHandler(t)
 	emb := router.NewHashEmbedder(1024)
 	h.recall.embedder = emb
-	h.router = buildRouter(emb, h.matcher, 0.55, nil)
+	h.router = buildRouter(emb, h.matcher, 0.55, nil, nil)
 
 	if _, asked := h.askClarify(ctx, clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {
 		t.Fatal("expected a question")
@@ -671,7 +671,7 @@ func TestUnresolvedActSaysItDoesNotKnowTheCommand(t *testing.T) {
 func newRoutingClarifyHandler(t *testing.T) (*reactiveHandler, *store.Store) {
 	t.Helper()
 	h, st, _ := newClarifyHandler(t)
-	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil)
+	h.router = buildRouter(router.NewHashEmbedder(1024), h.matcher, 0.55, nil, nil)
 	h.recall = recallWiring{embedder: router.NewHashEmbedder(1024), memStore: memory.NewInMemoryStore()}
 	return h, st
 }

cmd/mavend/decisiontrace_test.go

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

cmd/mavend/dialogue_contract_test.go

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

cmd/mavend/fact_subject_test.go

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

cmd/mavend/factgate_test.go

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

cmd/mavend/notefragment_test.go

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

cmd/mavend/querywalk_test.go

@@ -10,7 +10,7 @@ import (
 // whose model is down names nothing, and naming nothing has to walk the chain
 // the way it walked before the field existed.
 func TestNoDestinationWalksTheWholeChain(t *testing.T) {
-	walk, skipped := queryWalk(router.SourceUnknown)
+	walk, skipped := queryWalk(router.SourceUnknown, false)
 	if len(skipped) != 0 {
 		t.Errorf("skipped %d sources with no destination named, want none", len(skipped))
 	}
@@ -32,15 +32,16 @@ func TestANamedDestinationSilencesTheOtherGuessers(t *testing.T) {
 		dest      router.Source
 		utterance string
 		silenced  string
+		anchored  bool // a stage 0 grammar named the destination
 	}{
-		{router.SourceWorld, "что такое TCP?", "weather"},
-		{router.SourceWorld, "сколько будет 17 на 23?", "weather"},
-		{router.SourceWorld, "кто такой Линус Торвальдс?", "personal"},
-		{router.SourceRecall, "какой у меня любимый язык?", "feeds"},
-		{router.SourceCalendar, "что в календаре на завтра?", "weather"},
+		{router.SourceWorld, "что такое TCP?", "weather", true},
+		{router.SourceWorld, "сколько будет 17 на 23?", "weather", true},
+		{router.SourceWorld, "кто такой Линус Торвальдс?", "personal", true},
+		{router.SourceRecall, "какой у меня любимый язык?", "feeds", false},
+		{router.SourceCalendar, "что в календаре на завтра?", "weather", true},
 	}
 	for _, c := range cases {
-		walk, skipped := queryWalk(c.dest)
+		walk, skipped := queryWalk(c.dest, c.anchored)
 		if inWalk(walk, c.silenced) {
 			t.Errorf("%q named %q: %q is still asked", c.utterance, c.dest, c.silenced)
 		}
@@ -56,7 +57,7 @@ func TestANamedDestinationSilencesTheOtherGuessers(t *testing.T) {
 // data first, then the world", and a destination a model wrote must not be able
 // to reverse it.
 func TestNamingTheWorldStillReadsHisDataFirst(t *testing.T) {
-	walk, _ := queryWalk(router.SourceWorld)
+	walk, _ := queryWalk(router.SourceWorld, true)
 	for _, look := range []string{"fact-by-key", "embed", "memory", "notes"} {
 		if !inWalk(walk, look) {
 			t.Errorf("%q was dropped; only the sources that guess may be dropped", look)
@@ -74,7 +75,7 @@ func TestNamingTheWorldStillReadsHisDataFirst(t *testing.T) {
 // makes "какой у меня любимый язык?" answer "не нашла у тебя такой записи"
 // rather than reaching SearXNG once nothing local had it.
 func TestNamingRecallKeepsTheBoundary(t *testing.T) {
-	walk, _ := queryWalk(router.SourceRecall)
+	walk, _ := queryWalk(router.SourceRecall, true)
 	if !inWalk(walk, "personal") {
 		t.Fatal("the personal boundary was skipped on a turn named for his own data")
 	}
@@ -83,12 +84,33 @@ func TestNamingRecallKeepsTheBoundary(t *testing.T) {
 	}
 }
 
+// The owner's call of 2026-08-09 (V-666): only a stage 0 grammar may take the
+// personal boundary off a turn. The routing heads and the resident model both
+// name a destination by inference, and an inferred SourceWorld would send a
+// question about him upstream. Every other guesser still goes.
+func TestOnlyAGrammarMayDropTheBoundary(t *testing.T) {
+	walk, skipped := queryWalk(router.SourceWorld, false)
+	if !inWalk(walk, "personal") {
+		t.Error("an inferred destination took the boundary off the turn")
+	}
+	if !inWalk(skipped, "weather") {
+		t.Error("weather is still asked; the rule covers the boundary alone")
+	}
+	if posOf(walk, "personal") > posOf(walk, "search") {
+		t.Error("the boundary no longer sits in front of the world")
+	}
+	if anchored, _ := queryWalk(router.SourceWorld, true); inWalk(anchored, "personal") {
+		t.Error(`a grammar named the world and the boundary stayed: ` +
+			`"кто такой Линус Торвальдс?" is answered "не нашла у тебя такой записи" again`)
+	}
+}
+
 // Whatever the destination, the walk is a subsequence of the table. Every
 // comment on that table argues an order between two sources, and none of those
 // reasons is about this field.
 func TestTheWalkNeverReordersTheTable(t *testing.T) {
 	for _, dest := range append([]router.Source{router.SourceUnknown}, router.Sources...) {
-		walk, skipped := queryWalk(dest)
+		walk, skipped := queryWalk(dest, true)
 		if len(walk)+len(skipped) != len(querySources) {
 			t.Errorf("%q: %d walked + %d skipped, want %d", dest, len(walk), len(skipped), len(querySources))
 		}

cmd/mavend/reactive_notes_test.go

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

cmd/mavend/simulator_test.go

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

cmd/mavend/sttseam_test.go

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

cmd/mavend/voicewire.go

@@ -36,7 +36,9 @@ type voiceWiring struct {
 	sessions  *voice.Sessions
 	voiceSink delivery.Sink
 	embedder  router.Embedder
-	handler   *reactiveHandler // the reactive handler for IPC Chat
+	// heads — the routing heads, nil unless embedder.heads_path is set.
+	heads   *router.RouterHeads
+	handler *reactiveHandler // the reactive handler for IPC Chat
 	// worker clients (set when configured as Remote): closed on shutdown so
 	// mavsttd / mavttsd don't keep a stale conn into a restarting daemon.
 	sttClient *worker.Client
@@ -53,7 +55,11 @@ type voiceWiring struct {
 	// unless a `workstation` block names an address. Held here only so the
 	// prober is stopped on shutdown; callers were handed it at build time.
 	pair *llm.Pair
-	mcp  *mcpWiring
+	// sttPair — CrisperWhisper 2.0 on the workstation with mavsttd as the
+	// floor, nil unless the `workstation.stt` block names an address. Held for
+	// the same reason as pair: to stop its prober on shutdown.
+	sttPair *stt.Pair
+	mcp     *mcpWiring
 	// home — the Home Assistant client, nil unless the `smarthome` block is
 	// enabled (Vikunja #256). Its devices land in the same allowlist as every
 	// other act, so nothing else here has to know about it.
@@ -72,6 +78,9 @@ func (w *voiceWiring) close() {
 	if w.embedder != nil {
 		_ = w.embedder.Close()
 	}
+	if w.heads != nil {
+		_ = w.heads.Close()
+	}
 	if w.server != nil {
 		_ = w.server.Close()
 	}
@@ -84,6 +93,9 @@ func (w *voiceWiring) close() {
 	if w.pair != nil {
 		w.pair.Stop()
 	}
+	if w.sttPair != nil {
+		w.sttPair.Stop()
+	}
 	w.mcp.close()
 }
 
@@ -112,6 +124,7 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	} else {
 		transcriber = stt.NewStub()
 	}
+	transcriber, w.sttPair = sttSeam(cfg, transcriber)
 	w.transcriber = transcriber
 
 	// ----- tts (Stub in-process OR Remote) -----
@@ -147,6 +160,24 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		emb = router.NewHashEmbedder(1024)
 	}
 	w.embedder = emb
+
+	// ----- router: routing heads (only when configured, and never fatal) -----
+	// A missing or broken weights file logs and leaves w.heads nil, which is
+	// byte-for-byte the cascade that shipped before V-664. Refusing to start
+	// over a routing accelerator would trade a working box for a better one.
+	if cfg.Voice.Embedder != nil && cfg.Voice.Embedder.HeadsPath != "" {
+		h, err := router.NewRouterHeads(
+			cfg.Voice.Embedder.HeadsPath,
+			cfg.Voice.Embedder.TokenizerPath,
+		)
+		if err != nil {
+			log.Printf("voice: routing heads unavailable, cascade unchanged: %v", err)
+		} else {
+			log.Printf("voice: routing heads loaded from %s", cfg.Voice.Embedder.HeadsPath)
+			w.heads = h
+		}
+	}
+
 	repairFactVectors(dataStore, emb)
 	checkStoredEmbedder(dataStore, emb)
 	// Retention is enforced on write, which is not enough on its own: a box that
@@ -223,7 +254,8 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	// against the classifier's 50.0%, at about 1s a turn instead of 30ms (see
 	// config.VoiceConfig.LLMRouter). The classifier always stays wired as the
 	// fallback, so a model error never breaks a turn.
-	rtr := buildRouter(emb, matcher, threshold, pickLLMRouter(cfg.Voice.UseLLMRouter(), hot))
+	rtr := buildRouter(emb, matcher, threshold,
+		pickLLMRouter(cfg.Voice.UseLLMRouter(), hot), w.heads)
 
 	// ----- sessions registry (shared with voicesink) -----
 	sessions := voice.NewSessions()
@@ -366,6 +398,44 @@ func modelSeam(cfg *config.Config, resident *llm.Client) (router.Completer, *llm
 	return pair, pair
 }
 
+// sttSeam builds the transcription seam the voice path and the meeting
+// recorder share. It is modelSeam for audio and follows the same rule.
+//
+// With no `workstation.stt` block it hands back the floor untouched, which is
+// today's deploy exactly. With one, it is an stt.Pair preferring CrisperWhisper
+// 2.0 on workpc, which scores 10.4% WER in Russian against the floor's 27.5%
+// (docs/evals/2026-08-09-crisperwhisper2-russian-wer.md).
+//
+// Only the silent half of the degradation rule applies here. A worse transcript
+// is still a turn, so there is nothing to name a gap about and the fallback is
+// never spoken. That is why stt.Pair has no TranscribeRemote.
+func sttSeam(cfg *config.Config, floor stt.Transcriber) (stt.Transcriber, *stt.Pair) {
+	if cfg.Workstation == nil || cfg.Workstation.Stt == nil {
+		return floor, nil
+	}
+	s := cfg.Workstation.Stt
+	lang := ""
+	if cfg.Voice != nil {
+		lang = cfg.Voice.Lang
+		if cfg.Voice.Stt != nil && cfg.Voice.Stt.Lang != "" {
+			lang = cfg.Voice.Stt.Lang
+		}
+	}
+	pair := stt.NewPair(
+		stt.NewHTTPTranscriber(s.URL, s.Token, lang, time.Duration(s.Timeout)),
+		floor,
+		s.Health,
+		time.Duration(s.Probe),
+	)
+	pair.Start(context.Background())
+	if s.Token == "" {
+		log.Print("voice: the workstation transcriber has no token, so anything on the LAN can post audio to it")
+	}
+	log.Printf("voice: workstation transcriber at %s, probed every %s, mavsttd as the floor",
+		s.URL, time.Duration(s.Probe))
+	return pair, pair
+}
+
 func pickLLMRouter(enabled bool, c router.Completer) *router.LLMRouter {
 	if !enabled {
 		return nil
@@ -390,7 +460,8 @@ func pickLLMRouter(enabled bool, c router.Completer) *router.LLMRouter {
 //     intent from seedDir (models/seeds/<intent>.txt) — see seedClassifier
 //     below for the current intent list and file names.
 //   - Threshold is from voice.router_threshold config (default 0.55).
-func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64, llmR *router.LLMRouter) *router.Router {
+func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
+	llmR *router.LLMRouter, heads *router.RouterHeads) *router.Router {
 	cls := router.NewClassifier(emb)
 	seedClassifier(cls)
 	grammars := router.DefaultGrammars(acts)
@@ -442,6 +513,7 @@ func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
 		},
 		Threshold: threshold,
 		LLM:       llmR,
+		Heads:     heads,
 	})
 }
 

cmd/mavgpud/gpu.go

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

cmd/mavgpud/gpu_test.go

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

cmd/mavgpud/main.go

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

cmd/mavgpud/runner.go

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

cmd/mavgpud/runner_test.go

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

cmd/mavwaked/main.go

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

cmd/mavwaked/silero.go

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

cmd/mavwaked/silero_test.go

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

cmd/mavwaked/vad.go

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

deploy/cw2/serve.py

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

deploy/mavend.json

@@ -78,10 +78,30 @@
     "Addressed by LAN address, not container name: mavgpud runs on another",
     "machine and there is no shared docker network to name it on."
   ],
+  "//workstation.stt": [
+    "CrisperWhisper 2.0 turbo on the same machine, a second service on port",
+    "8081 and not a second endpoint on mavgpud. whisper.cpp cannot load CW2 at",
+    "all: it derives its language count from the vocabulary size, and CW2's",
+    "51897 tokens shift seven special token ids. So it runs under transformers",
+    "there and mavsttd stays whisper.cpp here.",
+    "Worth the second service: CW2 turbo scores 10.4% WER in Russian against",
+    "27.5% for the ggml-small.bin mavsttd loads, measured on 200 Golos clips",
+    "in docs/evals/2026-08-09-crisperwhisper2-russian-wer.md.",
+    "Deleting this block sends every utterance to mavsttd, which is what the",
+    "box did before it existed. A worse transcript is still a turn, so the",
+    "fallback is silent and Kami is never told which machine heard him.",
+    "The token is what stops anything on the LAN posting audio to that port."
+  ],
   "workstation": {
     "url": "http://192.168.1.105:8080",
     "probe": "15s",
-    "timeout": "90s"
+    "timeout": "90s",
+    "stt": {
+      "url": "http://192.168.1.105:8081/transcribe",
+      "token": "${MAVEN_STT_TOKEN}",
+      "probe": "15s",
+      "timeout": "10s"
+    }
   },
 
   "//search": [
@@ -231,7 +251,8 @@
     "embedder": {
       "model_path": "/opt/maven/models/embedder/multilingual-e5-small/model_quantized.onnx",
       "tokenizer_path": "/opt/maven/models/embedder/multilingual-e5-small/tokenizer.json",
-      "lib_path": "/opt/maven/lib/libonnxruntime.so"
+      "lib_path": "/opt/maven/lib/libonnxruntime.so",
+      "heads_path": "/opt/maven/models/embedder/router-heads/router_heads.onnx"
     },
     "llm_router": true,
     "query_min_score": 0.55,

deploy/mavgpud.json

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

deploy/mavgpud.service

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

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

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

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

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

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

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

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

@@ -0,0 +1,103 @@
+# Kiwix answered the wrong question, and the fix was not a relevance gate
+
+Date: 2026-08-09. Task: V-668. Box: homesrv, workstation off.
+Book: `wikipedia_ru_all_maxi_2026-02` on `127.0.0.1:8034`.
+
+## What started it
+
+Two turns on 2026-08-09 came back wrong from the offline encyclopedia.
+"почему небо голубое" was answered off the song "Город золотой". "что такое
+TCP?" was answered off "Перехват TCP-соединения". Both were phrased
+confidently, because `queryKiwix` claims a turn whenever the search returns
+anything and `len(hits) == 0` is its only gate.
+
+The plan was a relevance gate. multilingual-e5-small is asymmetric and trained
+for exactly this, `query:` against `passage:`, and the query vector is already
+held on the turn. The 2026-08-05 measurement that killed a search-quality gate
+killed three lexical signals. It says in its own words that it never probed
+Kiwix.
+
+## The gate does not exist
+
+Fourteen Russian questions, eight the encyclopedia can answer and six it
+cannot. Each question was searched, the top article read, and the cosine of
+`EmbedQuery(question)` against `EmbedPassage(article)` recorded.
+
+| set | n | min | mean | max |
+|---|---|---|---|---|
+| answerable | 8 | 0.7934 | 0.8400 | 0.9087 |
+| not answerable | 6 | 0.7480 | 0.7852 | 0.8367 |
+
+Two of the six unanswerable score above the weakest answerable one. That alone
+would be a poor threshold. The log killed it outright: seven of the eight
+answerable questions got a **wrong** article back, and those wrong articles
+scored high. The TCP hijacking article scored 0.8653, above five of the six
+unanswerable rows.
+
+The finding is that this cosine measures topic and not answerhood. A page about
+hijacking TCP sessions is about TCP. No threshold separates it from a page that
+defines TCP, and one that tried would take the definition with it.
+
+## The defect is retrieval
+
+`internal/kiwix/client.go` has said it since it was written: ranking is keyword
+based, "why is the sky blue" finds a TV episode. `queryKiwix` sends the whole
+sentence. The English path has a rewriter that reduces a question to keywords
+with a model call. The Russian path reads the book verbatim (V-508) and had
+nothing. So the question words compete with the one word that names the article.
+
+Dropping the question words changes the answer:
+
+| sent | first hit |
+|---|---|
+| `кто написал Войну и мир` | Радуйся, мир (Доктор Кто) |
+| `Война и мир` | Война и мир |
+| `что такое TCP` | Перехват TCP-соединения |
+| `TCP` | TCP |
+
+A ZIM is also addressable by title, which nothing here used. `/A/Франция`,
+`/A/TCP` and `/A/Небо` are 200. `/A/Трюмбальная_нидроскопия` is 404. So an
+exact title is safe to try first: it either answers or costs one request that
+says nothing.
+
+## What shipped, measured
+
+`kiwix.Topic` drops the narrative request, the interrogative and a verb sitting
+behind one. It keeps everything else, because a word it cannot classify is more
+likely the topic than noise. `kiwix.TitlePath` tries the exact article before
+any ranking runs. Both apply on the verbatim path only, since reducing twice
+would take the topic off the rewriter's input.
+
+| question | before | after |
+|---|---|---|
+| что такое TCP? | Перехват TCP-соединения | **TCP** (by title) |
+| что такое фотосинтез | C4-фотосинтез | **Фотосинтез** |
+| кто такой Линус Торвальдс? | Tux | **Торвальдс, Линус** (by title) |
+| кто написал Войну и мир | Радуйся, мир (Доктор Кто) | **Война и мир** |
+| что такое чёрная дыра | Чёрная дыра | Чёрная дыра |
+| почему небо голубое | Город золотой | Под небом голубым… (фильм) |
+| почему трава зелёная | Сено | Зелень |
+| столица Франции | Список столиц Олимпийских игр | Список столиц Олимпийских игр |
+
+Four questions reach the right article where they did not. Two were already
+right and stay right. Nothing regressed.
+
+## What is still wrong
+
+Two of the eight are still not answered, and both are the same shape. The
+question names no article. "почему небо голубое" is answered by Rayleigh
+scattering and "столица Франции" by the lead of Франция. Neither title is in
+the question. Keyword retrieval cannot bridge that and neither can a
+threshold. The candidates are a semantic index over titles, or asking the
+resident model for the article title rather than for keywords.
+
+`Response.Empty()` is still the whole gate. A wrong article that the search
+does return is still spoken. What this change buys is that the article is
+usually right, not that a wrong one is caught.
+
+## Not measured here
+
+The English path, which still goes through the rewriter and was not touched.
+SearXNG, where the same question about answerhood is open and the 2026-08-05
+result stands. The cascade end to end, since the workstation is off and the
+phrasing arm is the resident model.

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

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

docs/offload.md

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

internal/config/voice.go

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

internal/config/workstation.go

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

internal/kiwix/client.go

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

internal/kiwix/live_test.go

@@ -0,0 +1,65 @@
+package kiwix
+
+import (
+	"context"
+	"os"
+	"testing"
+	"time"
+)
+
+// TestLiveTopicBeatsTheSentence — the measurement V-668 turned on, kept as a
+// test so the claim can be re-run rather than believed.
+//
+// It prints the article the old path returned and the article the new one
+// returns, for the same question. It asserts nothing about which is better,
+// because "is this the right article" is a human's call. It fails only if the
+// two paths agree on every case, which would mean the change does nothing.
+//
+//	MAVEN_KIWIX_URL=http://127.0.0.1:8034 make t PKG=./internal/kiwix/ RUN=TestLive V=1
+func TestLiveTopicBeatsTheSentence(t *testing.T) {
+	base := os.Getenv("MAVEN_KIWIX_URL")
+	if base == "" {
+		t.Skip("MAVEN_KIWIX_URL unset — point it at the kiwix-server host port")
+	}
+	const book = "wikipedia_ru_all_maxi_2026-02"
+	c := New(base)
+	questions := []string{
+		"что такое TCP?",
+		"что такое фотосинтез",
+		"кто такой Линус Торвальдс?",
+		"кто написал Войну и мир",
+		"что такое чёрная дыра",
+		"почему небо голубое",
+		"почему трава зелёная",
+		"столица Франции",
+	}
+	moved := 0
+	for _, q := range questions {
+		ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+		before := firstTitle(ctx, c, q, book)
+		topic := Topic(q)
+		after := ""
+		if page, err := c.Article(ctx, TitlePath(book, topic), 400); err == nil && page.Text != "" {
+			after = page.Title + " (by title)"
+		} else {
+			after = firstTitle(ctx, c, topic, book)
+		}
+		cancel()
+		if before != after {
+			moved++
+		}
+		t.Logf("%-30s  before=%-34q after=%q", q, before, after)
+	}
+	t.Logf("%d of %d questions reach a different article", moved, len(questions))
+	if moved == 0 {
+		t.Error("the topic path returns exactly what the sentence path returned")
+	}
+}
+
+func firstTitle(ctx context.Context, c *Client, pattern, book string) string {
+	hits, err := c.Search(ctx, pattern, book, 3)
+	if err != nil || len(hits) == 0 {
+		return "(nothing)"
+	}
+	return hits[0].Title
+}

internal/kiwix/topic.go

@@ -0,0 +1,98 @@
+package kiwix
+
+import (
+	"strings"
+	"unicode"
+
+	"github.com/kami/maven/internal/lexicon"
+	"github.com/kami/maven/internal/morph"
+)
+
+// Topic reduces a question to the thing it is about, because Kiwix ranks by
+// keyword overlap and a whole sentence buries the keyword that matters.
+//
+// This package's own doc says it: "why is the sky blue" finds a TV episode.
+// Measured against the Russian ZIM on 2026-08-09, the sentence and the topic
+// return different articles for the same question. "кто написал Войну и мир"
+// returns "Радуйся, мир (Доктор Кто)"; "Войну и мир" returns the novel first.
+// "что такое TCP" returns "Перехват TCP-соединения"; "TCP" returns TCP. The
+// English path had a rewriter doing this with a model call. The Russian path
+// reads the book verbatim (V-508) and had nothing.
+//
+// It drops three things off the front and stops: the narrative request, the
+// interrogative, and a verb sitting between them and the noun. Everything else
+// is kept, because a word this cannot classify is more likely the topic than
+// noise. An empty return means the utterance was question words alone, and the
+// caller searches the sentence as before.
+func Topic(utterance string) string {
+	words := strings.Fields(strings.TrimSpace(utterance))
+	cut := 0
+	for cut < len(words) {
+		w := strings.Trim(strings.ToLower(words[cut]), ".,!?…:;\"'«»")
+		if w == "" {
+			cut++
+			continue
+		}
+		switch {
+		case inList(lexicon.NarrativeRequests(), w),
+			inList(lexicon.Interrogatives(), w),
+			inList(lexicon.FirstPerson(), w),
+			// "что ТАКОЕ x", "кто ТАКОЙ x" — the copula that only ever follows
+			// an interrogative, and never a topic on its own.
+			cut > 0 && isCopula(w),
+			// "расскажи ПРО x", "о x". One-letter and two-letter prepositions
+			// are not a closed class worth a lexicon set of their own.
+			cut > 0 && isLeadingPreposition(w),
+			// "кто НАПИСАЛ Войну и мир". A verb here is the question's own
+			// verb, not part of the title. Only after something was already
+			// dropped, so "написал отчёт" as a topic survives intact.
+			cut > 0 && morph.IsVerbForm(w):
+			cut++
+		default:
+			// The question mark is the sentence's, not the title's, and Kiwix
+			// carries it into the keyword match.
+			topic := strings.TrimRight(strings.Join(words[cut:], " "), " .,!?…:;\"'«»")
+			if !hasLetter(topic) {
+				return ""
+			}
+			return topic
+		}
+	}
+	return ""
+}
+
+func isCopula(w string) bool {
+	switch w {
+	case "такое", "такой", "такая", "такие", "is", "are", "was", "were":
+		return true
+	}
+	return false
+}
+
+func isLeadingPreposition(w string) bool {
+	switch w {
+	case "про", "о", "об", "обо", "по", "about", "of", "on":
+		return true
+	}
+	return false
+}
+
+func inList(list []string, w string) bool {
+	for _, x := range list {
+		if x == w {
+			return true
+		}
+	}
+	return false
+}
+
+// hasLetter is the guard against a topic that reduced to punctuation or digits
+// alone, which no ZIM title matches.
+func hasLetter(s string) bool {
+	for _, r := range s {
+		if unicode.IsLetter(r) {
+			return true
+		}
+	}
+	return false
+}

internal/kiwix/topic_test.go

@@ -0,0 +1,51 @@
+package kiwix
+
+import "testing"
+
+// The cases the 2026-08-09 measurement turned on, plus the ones a topic must
+// not damage. Each left column returned a wrong article when it was sent whole.
+func TestTopicKeepsTheThingTheQuestionIsAbout(t *testing.T) {
+	cases := []struct{ utterance, want string }{
+		{"что такое TCP?", "TCP"},
+		{"что такое фотосинтез", "фотосинтез"},
+		{"кто такой Линус Торвальдс?", "Линус Торвальдс"},
+		{"кто написал Войну и мир", "Войну и мир"},
+		{"расскажи про битву при Ватерлоо", "битву при Ватерлоо"},
+		{"what is photosynthesis", "photosynthesis"},
+		// No question word, so there is nothing to drop. The topic is the
+		// whole utterance and the search is what it was before.
+		{"столица Франции", "столица Франции"},
+		{"почему небо голубое", "небо голубое"},
+	}
+	for _, c := range cases {
+		if got := Topic(c.utterance); got != c.want {
+			t.Errorf("Topic(%q) = %q, want %q", c.utterance, got, c.want)
+		}
+	}
+}
+
+// A verb only goes when a question word already went. Otherwise "написал
+// отчёт" loses the verb that names what he means.
+func TestTopicDropsAVerbOnlyBehindAQuestionWord(t *testing.T) {
+	if got := Topic("написал отчёт"); got != "написал отчёт" {
+		t.Errorf("Topic dropped a leading verb with no question word: %q", got)
+	}
+}
+
+// Question words alone reduce to nothing, and the caller reads that as "no
+// topic" and searches the sentence rather than searching the empty string.
+func TestTopicIsEmptyWhenNothingIsLeft(t *testing.T) {
+	for _, q := range []string{"что такое?", "кто?", "почему", "???"} {
+		if got := Topic(q); got != "" {
+			t.Errorf("Topic(%q) = %q, want empty", q, got)
+		}
+	}
+}
+
+func TestTitlePathEscapesAndUnderscores(t *testing.T) {
+	got := TitlePath("wikipedia_ru_all_maxi_2026-02", "Чёрная дыра")
+	want := "/content/wikipedia_ru_all_maxi_2026-02/A/%D0%A7%D1%91%D1%80%D0%BD%D0%B0%D1%8F_%D0%B4%D1%8B%D1%80%D0%B0"
+	if got != want {
+		t.Errorf("TitlePath = %q, want %q", got, want)
+	}
+}

internal/router/decisiontrace.go

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

internal/router/embedderid_test.go

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

internal/router/eval/heads_test.go

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

internal/router/heads.go

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

internal/router/intent.go

@@ -111,6 +111,21 @@ type Decision struct {
 	// before this field existed. See source.go for why it is twelve values.
 	Source Source
 
+	// SourceAnchored — a stage 0 grammar named that destination, matching a
+	// literal pattern to do it. Only the router sets this, and only there.
+	//
+	// It exists because one thing downstream is not reversible by evidence
+	// (V-666). Naming a destination normally takes guessing sources off a turn,
+	// and one of those is the personal boundary, which is what stops a question
+	// about him from reaching the world. A grammar that read "что такое X" may
+	// take it off. A model or a softmax may not, because a wrong destination
+	// there widens what leaves the box rather than costing an answer.
+	//
+	// Read Stage instead and the two decisions get coupled: stage 0 also means
+	// confidence 1.0 and an anchored claim band, and a later cascade change
+	// could make one true where the other is not.
+	SourceAnchored bool
+
 	// Continued — this decision was rebuilt from the previous turn rather
 	// than routed, because the utterance was an ellipsis ("а завтра?").
 	// Handlers use it to know that Slots.Text is the PREVIOUS turn's topic

internal/router/onnxembedder.go

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

internal/router/router.go

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

internal/stt/http.go

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

internal/stt/http_test.go

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

internal/stt/pair.go

@@ -0,0 +1,157 @@
+package stt
+
+import (
+	"context"
+	"errors"
+	"log"
+	"net/http"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// Pair — a preferred transcriber on the workstation, with mavsttd as the floor.
+//
+// Same arrangement as llm.Pair and for the same reason. The microphone is at
+// workpc, the card there has 16GB, and CrisperWhisper 2.0 turbo scores 10.4%
+// WER in Russian against 27.5% for the ggml-small.bin homesrv loads
+// (docs/evals/2026-08-09-crisperwhisper2-russian-wer.md). The workstation is
+// never assumed up: it sleeps, and the card is often held by a training run.
+//
+// Speech-to-text has only the silent half of the degradation rule. A worse
+// transcript is still a turn, and there is nothing to name a gap about, so
+// Transcribe always falls back. That is the whole difference from llm.Pair,
+// which also carries CompleteRemote for callers that must refuse instead.
+type Pair struct {
+	remote Transcriber
+	floor  Transcriber
+
+	// up — the cached admission answer, written only by the prober and read by
+	// every turn. A voice turn must never wait on a machine that may be asleep.
+	up atomic.Bool
+
+	health   string
+	interval time.Duration
+	http     *http.Client
+	stop     chan struct{}
+	stopOnce sync.Once
+}
+
+const (
+	probeTimeout         = 2 * time.Second
+	defaultProbeInterval = 15 * time.Second
+)
+
+// ErrNoFloor — a Pair was built with no local transcriber to fall back to. A
+// configuration mistake: the floor is what makes the remote optional.
+var ErrNoFloor = errors.New("stt: no floor transcriber")
+
+// NewPair builds the two-transcriber arrangement. remote may be nil, which is
+// the unconfigured deploy: every turn goes to the floor and nothing probes.
+func NewPair(remote, floor Transcriber, health string, interval time.Duration) *Pair {
+	if interval <= 0 {
+		// The config normalises this, so a zero here is a caller that built the
+		// Pair directly. Panicking in a ticker is the wrong way to say so.
+		interval = defaultProbeInterval
+	}
+	return &Pair{
+		remote:   remote,
+		floor:    floor,
+		health:   health,
+		interval: interval,
+		http:     &http.Client{Timeout: probeTimeout},
+		stop:     make(chan struct{}),
+	}
+}
+
+// Start begins probing. The first probe runs before the first tick, so a
+// workstation that is already up serves the first utterance rather than the
+// second. Safe with a nil remote.
+func (p *Pair) Start(ctx context.Context) {
+	if p.remote == nil || p.health == "" {
+		return
+	}
+	go func() {
+		p.probe(ctx)
+		t := time.NewTicker(p.interval)
+		defer t.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case <-p.stop:
+				return
+			case <-t.C:
+				p.probe(ctx)
+			}
+		}
+	}()
+}
+
+// Stop ends the prober. Idempotent and safe from two goroutines.
+func (p *Pair) Stop() {
+	p.stopOnce.Do(func() { close(p.stop) })
+}
+
+// Available reports whether the workstation will transcribe right now.
+func (p *Pair) Available() bool {
+	return p.remote != nil && p.up.Load()
+}
+
+func (p *Pair) probe(ctx context.Context) {
+	ctx, cancel := context.WithTimeout(ctx, probeTimeout)
+	defer cancel()
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, p.health, nil)
+	if err != nil {
+		p.set(false)
+		return
+	}
+	resp, err := p.http.Do(req)
+	if err != nil {
+		p.set(false)
+		return
+	}
+	defer resp.Body.Close()
+	p.set(resp.StatusCode == http.StatusOK)
+}
+
+// set records the admission answer and logs only transitions. A machine that
+// sleeps nightly would otherwise write one line per interval forever.
+func (p *Pair) set(up bool) {
+	if p.up.Swap(up) == up {
+		return
+	}
+	if up {
+		log.Printf("stt: workstation transcriber available at %s", p.health)
+	} else {
+		log.Print("stt: workstation transcriber unavailable, falling back to mavsttd")
+	}
+}
+
+// Transcribe sends the audio to the workstation when it will take work, and to
+// mavsttd otherwise. A remote that fails mid-request falls back too, because
+// the admission answer is a cache and can be one interval out of date.
+//
+// Killing the remote mid-session must not drop the turn. That is the whole
+// point of the floor, and it is what TestPairFallsBackWhenRemoteFails pins.
+func (p *Pair) Transcribe(ctx context.Context, a audio.Audio) (string, float64, error) {
+	if p.floor == nil {
+		return "", 0, ErrNoFloor
+	}
+	if p.Available() {
+		text, conf, err := p.remote.Transcribe(ctx, a)
+		if err == nil {
+			log.Print("stt: transcribed on the workstation")
+			return text, conf, nil
+		}
+		// The cached answer was wrong. Correct it now rather than sending the
+		// next utterance into the same hole, then fall back.
+		p.set(false)
+		log.Printf("stt: workstation failed mid-request, falling back: %v", err)
+	}
+	return p.floor.Transcribe(ctx, a)
+}
+
+var _ Transcriber = (*Pair)(nil)

internal/stt/pair_test.go

@@ -0,0 +1,143 @@
+package stt
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// scripted — a Transcriber that answers with a fixed text, or fails.
+type scripted struct {
+	text  string
+	err   error
+	calls atomic.Int32
+}
+
+func (s *scripted) Transcribe(_ context.Context, _ audio.Audio) (string, float64, error) {
+	s.calls.Add(1)
+	if s.err != nil {
+		return "", 0, s.err
+	}
+	return s.text, 0.9, nil
+}
+
+func sample() audio.Audio {
+	return audio.Audio{Format: audio.PCM16kMono, Bytes: make([]byte, 3200)}
+}
+
+// up builds a Pair whose admission answer is already true, without probing.
+func up(remote, floor Transcriber) *Pair {
+	p := NewPair(remote, floor, "", time.Minute)
+	p.up.Store(true)
+	return p
+}
+
+func TestPairPrefersTheWorkstation(t *testing.T) {
+	t.Parallel()
+	remote := &scripted{text: "с рабочей станции"}
+	floor := &scripted{text: "с homesrv"}
+	text, _, err := up(remote, floor).Transcribe(context.Background(), sample())
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if text != "с рабочей станции" {
+		t.Fatalf("want the remote transcript, got %q", text)
+	}
+	if floor.calls.Load() != 0 {
+		t.Fatalf("floor was called %d times, want 0", floor.calls.Load())
+	}
+}
+
+// The turn is what matters. A remote that dies mid-session must cost a worse
+// transcript and nothing else. This is the V-486 bar.
+func TestPairFallsBackWhenRemoteFails(t *testing.T) {
+	t.Parallel()
+	remote := &scripted{err: errors.New("connection refused")}
+	floor := &scripted{text: "с homesrv"}
+	p := up(remote, floor)
+
+	text, conf, err := p.Transcribe(context.Background(), sample())
+	if err != nil {
+		t.Fatalf("a failed remote must not fail the turn: %v", err)
+	}
+	if text != "с homesrv" {
+		t.Fatalf("want the floor transcript, got %q", text)
+	}
+	if conf != 0.9 {
+		t.Fatalf("want the floor confidence, got %v", conf)
+	}
+	if p.Available() {
+		t.Fatal("a failed request must correct the cached admission answer")
+	}
+
+	// The next utterance goes straight to the floor rather than into the
+	// same hole.
+	if _, _, err := p.Transcribe(context.Background(), sample()); err != nil {
+		t.Fatalf("second turn: %v", err)
+	}
+	if remote.calls.Load() != 1 {
+		t.Fatalf("remote called %d times, want 1", remote.calls.Load())
+	}
+}
+
+func TestPairWithNoRemoteIsTheFloor(t *testing.T) {
+	t.Parallel()
+	floor := &scripted{text: "с homesrv"}
+	p := NewPair(nil, floor, "", time.Minute)
+	p.Start(context.Background()) // no health url, so this is a no-op
+	if p.Available() {
+		t.Fatal("an unconfigured remote is never available")
+	}
+	text, _, err := p.Transcribe(context.Background(), sample())
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if text != "с homesrv" {
+		t.Fatalf("want the floor transcript, got %q", text)
+	}
+}
+
+func TestPairWithNoFloorRefuses(t *testing.T) {
+	t.Parallel()
+	_, _, err := NewPair(nil, nil, "", time.Minute).Transcribe(context.Background(), sample())
+	if !errors.Is(err, ErrNoFloor) {
+		t.Fatalf("want ErrNoFloor, got %v", err)
+	}
+}
+
+func TestPairProbeReadsHealth(t *testing.T) {
+	t.Parallel()
+	var ok atomic.Bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		if !ok.Load() {
+			w.WriteHeader(http.StatusServiceUnavailable)
+			return
+		}
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer srv.Close()
+
+	p := NewPair(&scripted{text: "remote"}, &scripted{text: "floor"}, srv.URL, time.Minute)
+	p.probe(context.Background())
+	if p.Available() {
+		t.Fatal("a 503 means the card is busy, so the workstation is not available")
+	}
+	ok.Store(true)
+	p.probe(context.Background())
+	if !p.Available() {
+		t.Fatal("a 200 means the workstation will take work")
+	}
+}
+
+func TestPairStopIsIdempotent(t *testing.T) {
+	t.Parallel()
+	p := NewPair(nil, &scripted{}, "", time.Minute)
+	p.Stop()
+	p.Stop()
+}