docs: decide where mavwaked and mavenclient run — not on homesrv (V-463)

They appear in no compose file and run as no host process, and the task
asked whether that is a gap to close or a decision to write down. It is a
decision.

The reason is not hardware. homesrv is a Lenovo laptop and
/proc/asound/cards lists its ACP mic array with capture devices, so
passing /dev/snd into a container would work. It would also listen to an
empty room. A wake-word daemon is worth having where he is standing, and
that is not where the server is.

mavenclient is a client by name and design, mavwaked is the gate in
front of it, and the wire already reaches off-box: ipc.Dial takes
tcp://host:port?token=... through the netaddr seam, with the token
checked before internal/ipc sees the connection. So this needs a machine
and a config line, not protocol work.

The honest consequence is worse than the task suggested, and both docs
now say it: the wake word and the VAD gate are covered by unit tests and
by nothing else. QA session 1 step 2 was reworded to claim only what it
checks, which is push-to-talk through /dash. CLAUDE.md listed all nine
binaries with no column for where they run, which is how this went
unnoticed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

CLAUDE.md

@@ -32,7 +32,11 @@ See `docs/rearchitecture.md` for the target architecture, `docs/design.md` for t
 GPU and the workstation has 16GB of VRAM. So the resident model, STT and TTS become preferred
 remotes with a floor on homesrv. The workstation is never assumed up. Fall back silently when
 it would only do the job better. Name the gap when the 1.7B cannot do it at all. The embedder
-stays on homesrv permanently, because it backs that floor. Read `docs/offload.md` before
+stays on homesrv permanently, because it backs that floor. It is multilingual-e5-small,
+quantized and asymmetric — `EmbedQuery` and `EmbedPassage` apply the `query:`/`passage:`
+prefixes it was trained with, and calling plain `Embed` on a note is a bug. It replaced
+MiniLM and bought ten points of recall@1 and 2.5× the speed; see
+`docs/evals/2026-08-04-recall-e5-small.md`. Read `docs/offload.md` before
 touching a daemon seam or adding a model caller. Vikunja #483 is the umbrella, #484 to #487
 are the work.
 
@@ -73,8 +77,8 @@ Pure-Go packages (`router`, `memory`, `mavweb`, …) run under a plain `go test
 | `mavweb` | HTTP UI + PWA (`/dash`, `/history`, `/trace`, `/notifications`, `/tools`); WebAuthn auth. Connects to mavend's socket. |
 | `mavsttd` | Speech-to-text (whisper.cpp, CGO). |
 | `mavttsd` | Text-to-speech (piper subprocess). |
-| `mavwaked` | Wake-word / VAD gate. |
-| `mavenclient` | Voice loop client (mic → stt → core → tts). |
+| `mavwaked` | Wake-word / VAD gate. **Not on homesrv** — see below. |
+| `mavenclient` | Voice loop client (mic → stt → core → tts). **Not on homesrv** — see below. |
 | `mavpoll` | Telegram long-poll reach. |
 | `mavcaldav` | CalDAV calendar sync. |
 | `mavmaild` | Mail reader (IMAP, read-only). Holds the IMAP password; core never sees it. |
@@ -83,6 +87,15 @@ Daemons are wired socket-to-socket, not linked. `internal/ipc` is the client/ser
 protocol; the config in `deploy/mavend.json` (with `${VAR}` env expansion from gitignored
 `deploy/telegram.env`) sets socket paths, model paths, and the phraser/embedder blocks.
 
+**Seven of the nine run on homesrv. `mavwaked` and `mavenclient` do not, and that is the
+decision, not an oversight** (Vikunja #463, `docs/plans/17-where-the-voice-loop-runs.md`).
+homesrv has a microphone — it is a laptop — but it is in the wrong room, so a wake-word
+daemon there listens to nobody. They belong on a client machine where he is standing.
+`ipc.Dial` already takes `tcp://host:port?token=...` through the netaddr seam, so nothing
+needs building to allow it, but no such machine exists yet. **The consequence: the wake
+word and the VAD gate are covered by unit tests and by nothing else, and no amount of
+sitting at the box changes that.** Push-to-talk through `/dash` is what QA actually covers.
+
 ## The ecosystem: Nexus, Praxis, Hexis
 
 Maven is one of four services. It owns conversation and personal memory. It does not

Makefile

@@ -197,7 +197,7 @@ deps-piper:
 # multilingual-e5-small: an asymmetric retrieval model. It is trained to match
 # a short question against a longer passage, which is what note recall is.
 # The quantized file is the one we download, deploy and measure — see
-# docs/evals/2026-07-31-recall.md.
+# docs/evals/2026-08-04-recall-e5-small.md for what the swap bought.
 EMBEDDER_DIR := $(shell pwd)/models/embedder/multilingual-e5-small
 EMBEDDER_MODEL_URL := https://huggingface.co/Xenova/multilingual-e5-small/resolve/main/onnx/model_quantized.onnx
 EMBEDDER_TOKENIZER_URL := https://huggingface.co/Xenova/multilingual-e5-small/resolve/main/tokenizer.json

cmd/mavend/actions_fact.go

@@ -73,11 +73,11 @@ func (h *reactiveHandler) actionFact(ctx context.Context, dec router.Decision) s
 	// hears; storing the utterance meant recall answered with his own sentence
 	// rather than the value. The utterance stays alongside as provenance —
 	// readable on /trace, never the answer and never embedded.
-	if h.memStore != nil {
+	if h.recall.memStore != nil {
 		text := store.FactRecallText(dec.Slots.Key, dec.Slots.Value)
-		if vec, err := router.EmbedPassage(ctx, h.embedder, text); err != nil {
+		if vec, err := router.EmbedPassage(ctx, h.recall.embedder, text); err != nil {
 			log.Printf("voice: embed fact for memory: %v", err)
-		} else if err := h.memStore.Insert(ctx, "fact:"+dec.Slots.Key+":"+strconv.FormatInt(now.Unix(), 10), vec, map[string]string{
+		} else if err := h.recall.memStore.Insert(ctx, "fact:"+dec.Slots.Key+":"+strconv.FormatInt(now.Unix(), 10), vec, map[string]string{
 			"source":    "voice",
 			"type":      "fact",
 			"text":      text,

cmd/mavend/actions_note.go

@@ -20,7 +20,7 @@ func (h *reactiveHandler) actionNote(ctx context.Context, dec router.Decision) s
 	// embed the note text with the same model the classifier uses, persist
 	// via CoreAPI (source=tap:voice). Semantic recall lives in `notes`, not
 	// facts — no predicate reads it (spec's two-memory split).
-	vec, err := router.EmbedPassage(ctx, h.embedder, dec.Utterance)
+	vec, err := router.EmbedPassage(ctx, h.recall.embedder, dec.Utterance)
 	if err != nil {
 		log.Printf("voice: embed note: %v", err)
 		return "не получилось сохранить заметку."
@@ -33,8 +33,8 @@ func (h *reactiveHandler) actionNote(ctx context.Context, dec router.Decision) s
 	}
 	// Insert into long-term memory (best-effort, must not fail the note write).
 	// text/ts in the meta make a Search hit self-describing (see bestRecall).
-	if h.memStore != nil {
-		if err := h.memStore.Insert(ctx, "note:"+strconv.FormatInt(noteID, 10), vec, map[string]string{
+	if h.recall.memStore != nil {
+		if err := h.recall.memStore.Insert(ctx, "note:"+strconv.FormatInt(noteID, 10), vec, map[string]string{
 			"source": "voice",
 			"type":   "note",
 			"text":   dec.Utterance,

cmd/mavend/actions_query.go

@@ -406,7 +406,7 @@ func (h *reactiveHandler) queryWeather(ctx context.Context, t *queryTurn) (strin
 // sources below both need, run once, in the position it always ran in. It
 // only claims the turn when the embedder fails.
 func (h *reactiveHandler) queryEmbed(ctx context.Context, t *queryTurn) (string, bool) {
-	vec, err := router.EmbedQuery(ctx, h.embedder, t.dec.Utterance)
+	vec, err := router.EmbedQuery(ctx, h.recall.embedder, t.dec.Utterance)
 	if err != nil {
 		log.Printf("voice: embed query: %v", err)
 		return "не получилось найти ответ.", true
@@ -426,15 +426,15 @@ func (h *reactiveHandler) queryEmbed(ctx context.Context, t *queryTurn) (string,
 // gate, was the bug — the set of questions Maven answers is unchanged, only
 // which memory gets to answer them.
 func (h *reactiveHandler) queryMemory(ctx context.Context, t *queryTurn) (string, bool) {
-	if h.memStore == nil {
+	if h.recall.memStore == nil {
 		return "", false
 	}
-	hits, herr := h.memStore.Search(ctx, t.vec, 3)
+	hits, herr := h.recall.memStore.Search(ctx, t.vec, 3)
 	if herr != nil {
 		log.Printf("voice: memory search: %v", herr)
 		return "", false
 	}
-	hit, ok := bestRecall(hits, h.queryMinScore, h.queryMinMargin)
+	hit, ok := bestRecall(hits, h.recall.minScore, h.recall.minMargin)
 	if !ok {
 		return "", false
 	}
@@ -479,7 +479,7 @@ func (h *reactiveHandler) queryNotes(ctx context.Context, t *queryTurn) (string,
 	for i, n := range notes {
 		noteScores[i] = n.Score
 	}
-	if !memory.ConfidentScores(noteScores, h.queryMinScore, h.queryMinMargin) {
+	if !memory.ConfidentScores(noteScores, h.recall.minScore, h.recall.minMargin) {
 		return "", false
 	}
 	// Same topic veto as queryMemory above: the best note must be about what
@@ -766,8 +766,8 @@ func isPersonalQuery(utterance string) bool {
 // computed. Same shape as the cascade: the better test leads, the offline one
 // always answers.
 func (h *reactiveHandler) isPersonalTurn(ctx context.Context, t *queryTurn) bool {
-	h.boundary.load(ctx, h.embedder)
-	if personal, world, ok := h.boundary.score(t.vec); ok {
+	h.recall.boundary.load(ctx, h.recall.embedder)
+	if personal, world, ok := h.recall.boundary.score(t.vec); ok {
 		if personal > world {
 			log.Printf("voice: %q scores personal %.4f vs world %.4f", t.dec.Utterance, personal, world)
 			return true

cmd/mavend/clarify_test.go

@@ -299,7 +299,7 @@ func TestClarifyExpiryIsAnnouncedAndWordsStillRoute(t *testing.T) {
 	ctx := withDialogueID(context.Background(), dialogueIDFor(sourceText, ""))
 	h, _, now := newClarifyHandler(t)
 	emb := router.NewHashEmbedder(1024)
-	h.embedder = emb
+	h.recall.embedder = emb
 	h.router = buildRouter(emb, h.matcher, 0.55, nil)
 
 	if _, asked := h.askClarify(ctx, clarifyDec(router.IntentReminder, router.Slots{Text: "напомни"}, "напомни")); !asked {

cmd/mavend/fact_subject_test.go

@@ -28,11 +28,10 @@ func TestApplyAction_FactCapture_QueuesEntityResolution(t *testing.T) {
 
 	h := &reactiveHandler{
 		api:       api,
-		embedder:  emb,
+		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
 		router:    rtr,
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
-		memStore:  memory.NewInMemoryStore(),
 		dataStore: st,
 	}
 

cmd/mavend/factgate_test.go

@@ -19,11 +19,10 @@ func newFactGateHandler(t *testing.T, now time.Time) (*reactiveHandler, ipc.Core
 	emb := router.NewHashEmbedder(1024)
 	h := &reactiveHandler{
 		api:       api,
-		embedder:  emb,
+		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
 		router:    buildRouter(emb, tool.NewMatcher(api), 0.55, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
-		memStore:  memory.NewInMemoryStore(),
 		dataStore: st,
 	}
 	return h, api
@@ -44,7 +43,7 @@ func TestActionFact_QuestionIsNotWritten(t *testing.T) {
 	if _, err := api.LatestFact(ctx, "go_version"); err == nil {
 		t.Fatal("a question was stored as a fact about him")
 	}
-	hits, err := h.memStore.Search(ctx, mustEmbedPassage(t, h, "какая последняя версия языка Go?"), 3)
+	hits, err := h.recall.memStore.Search(ctx, mustEmbedPassage(t, h, "какая последняя версия языка Go?"), 3)
 	if err != nil {
 		t.Fatalf("memory search: %v", err)
 	}
@@ -81,7 +80,7 @@ func TestActionFact_ExplicitCaptureStillWrites(t *testing.T) {
 	// #493: what recall reads back is the fact, not the sentence he said.
 	// queryMemory returns a fact's text verbatim, so the utterance sitting here
 	// meant "запиши что я пил воду" was the answer to "когда я пил воду?".
-	hits, err := h.memStore.Search(ctx, mustEmbedPassage(t, h, "вода"), 3)
+	hits, err := h.recall.memStore.Search(ctx, mustEmbedPassage(t, h, "вода"), 3)
 	if err != nil {
 		t.Fatalf("memory search: %v", err)
 	}
@@ -117,7 +116,7 @@ func TestFactConfidence(t *testing.T) {
 
 func mustEmbedPassage(t *testing.T, h *reactiveHandler, text string) []float32 {
 	t.Helper()
-	vec, err := router.EmbedQuery(context.Background(), h.embedder, text)
+	vec, err := router.EmbedQuery(context.Background(), h.recall.embedder, text)
 	if err != nil {
 		t.Fatalf("embed %q: %v", text, err)
 	}

cmd/mavend/feeds_test.go

@@ -29,12 +29,12 @@ func buildFeedHandler(t *testing.T, feedsOn bool, notes ...ipc.Note) *reactiveHa
 		}
 	}
 	return &reactiveHandler{
-		api:      ipc.NewStoreAPI(st),
-		replier:  voice.NewStubReplier(),
-		phraser:  phraser.NewStub(),
-		now:      func() time.Time { return now },
-		feedsOn:  feedsOn,
-		embedder: nil,
+		api:     ipc.NewStoreAPI(st),
+		replier: voice.NewStubReplier(),
+		phraser: phraser.NewStub(),
+		now:     func() time.Time { return now },
+		feedsOn: feedsOn,
+		recall:  recallWiring{embedder: nil},
 	}
 }
 

cmd/mavend/personalboundary_test.go

@@ -70,7 +70,7 @@ func TestONNXPersonalBoundary(t *testing.T) {
 		{"how do i boil an egg", false},
 	}
 
-	h := &reactiveHandler{embedder: emb}
+	h := &reactiveHandler{recall: recallWiring{embedder: emb}}
 	ctx := context.Background()
 	wrong := 0
 	for _, c := range cases {
@@ -80,7 +80,7 @@ func TestONNXPersonalBoundary(t *testing.T) {
 		}
 		turn := &queryTurn{dec: router.Decision{Utterance: c.utterance}, vec: vec}
 		got := h.isPersonalTurn(ctx, turn)
-		p, w, ok := h.boundary.score(vec)
+		p, w, ok := h.recall.boundary.score(vec)
 		if !ok {
 			t.Fatal("seeds did not load with a working embedder")
 		}

cmd/mavend/query_recall_test.go

@@ -85,15 +85,17 @@ func buildRecallHandler(t *testing.T, question string, mems []recallCase) (*reac
 
 	phr := &recordingPhraser{Stub: phraser.NewStub()}
 	h := &reactiveHandler{
-		api:             ipc.NewStoreAPI(st),
-		embedder:        emb,
+		api: ipc.NewStoreAPI(st),
+		recall: recallWiring{
+			embedder:  emb,
+			memStore:  mem,
+			minScore:  0.55,
+			minMargin: 0.008,
+		},
 		replier:         voice.NewStubReplier(),
 		phraser:         phr,
 		now:             func() time.Time { return now },
-		memStore:        mem,
 		dataStore:       st,
-		queryMinScore:   0.55,
-		queryMinMargin:  0.008,
 		weatherProvider: nil,
 	}
 	return h, phr

cmd/mavend/reactive_notes_test.go

@@ -26,11 +26,10 @@ func TestReactiveNotesReminders(t *testing.T) {
 
 	h := &reactiveHandler{
 		api:       api,
-		embedder:  emb,
+		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
 		router:    rtr,
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
-		memStore:  memory.NewInMemoryStore(),
 		dataStore: st,
 	}
 
@@ -101,11 +100,10 @@ func TestSpokenTaskCaptureFilesATask(t *testing.T) {
 	matcher := tool.NewMatcher(api)
 	h := &reactiveHandler{
 		api:       api,
-		embedder:  emb,
+		recall:    recallWiring{embedder: emb, memStore: memory.NewInMemoryStore()},
 		router:    buildRouter(emb, matcher, 0.55, nil),
 		replier:   voice.NewStubReplier(),
 		now:       func() time.Time { return now },
-		memStore:  memory.NewInMemoryStore(),
 		dataStore: st,
 	}
 

cmd/mavend/recall.go

@@ -1,6 +1,9 @@
 package main
 
-import "github.com/kami/maven/internal/memory"
+import (
+	"github.com/kami/maven/internal/memory"
+	"github.com/kami/maven/internal/router"
+)
 
 // bestRecall is the read side of the long-term memory store: the top hit when
 // it clears the confidence gate. The index holds BOTH notes and facts, and
@@ -23,3 +26,34 @@ func bestRecall(results []memory.Result, minScore, minMargin float64) (memory.Re
 	}
 	return results[0], true
 }
+
+// recallWiring — the recall subsystem's dependencies, held as one group on
+// reactiveHandler (Vikunja #433). It is the worked example for the wiring
+// decision in docs/handler-wiring.md: cohesive groups of fields, not thirty
+// loose ones, so a handler names what it needs and the package can be split
+// later without exporting the whole struct.
+//
+// The zero value is usable and means "no recall": no embedder, no vector
+// store, and a gate that is never consulted because nothing is ever searched.
+type recallWiring struct {
+	// embedder — reused for note write/query (same model as the classifier).
+	embedder router.Embedder
+
+	// memStore — the vector index over notes and facts.
+	memStore memory.Store
+
+	// boundary — the embedded seed sets behind the personal boundary
+	// (personalboundary.go). Zero value is usable and loads on first query;
+	// with no embedder it never loads and the boundary uses personalMarkers.
+	boundary personalBoundary
+
+	// minScore — the note-recall confidence gate. Top cosine below this ⇒
+	// "I don't know" instead of a guess. Tuned for the ONNX embedder; a knob,
+	// not load-bearing math (same posture as the presence thresholds). Set by
+	// wireVoice from VoiceConfig; default 0.55.
+	minScore float64
+
+	// minMargin — the second half of that gate: how far the top hit must beat
+	// the runner-up. 0 ⇒ margin off.
+	minMargin float64
+}

cmd/mavend/simulator_test.go

@@ -428,20 +428,22 @@ func newSimWorld(t *testing.T, sc scenario) *simWorld {
 	rtr := buildRouter(emb, matcher, config.DefaultRouterThreshold, router.NewLLMRouter(scripted))
 
 	w.handler = &reactiveHandler{
-		stt:                simTranscriber{},
-		tts:                simSynthesizer{},
-		router:             rtr,
-		embedder:           emb,
+		stt:    simTranscriber{},
+		tts:    simSynthesizer{},
+		router: rtr,
+		recall: recallWiring{
+			embedder:  emb,
+			memStore:  st.VectorMemory(),
+			minScore:  config.DefaultQueryMinScore,
+			minMargin: config.DefaultQueryMinMargin,
+		},
 		api:                api,
 		matcher:            matcher,
 		tools:              tool.NewExecutor(api, 5*time.Second),
 		phraser:            phraser.NewStub(),
 		replier:            newLLMReplier(scripted, nil),
 		now:                clock.Now,
-		memStore:           st.VectorMemory(),
 		dataStore:          st,
-		queryMinScore:      config.DefaultQueryMinScore,
-		queryMinMargin:     config.DefaultQueryMinMargin,
 		timeParser:         router.StubDateTimeParser{},
 		dialogueSessions:   dialogue.NewSessionStore(time.Hour),
 		clarifyStore:       dialogue.NewClarifyStore(time.Hour),

cmd/mavend/voice.go

@@ -55,7 +55,6 @@ import (
 	"github.com/kami/maven/internal/crawl"
 	"github.com/kami/maven/internal/dialogue"
 	"github.com/kami/maven/internal/ipc"
-	"github.com/kami/maven/internal/memory"
 	"github.com/kami/maven/internal/phraser"
 	"github.com/kami/maven/internal/router"
 	"github.com/kami/maven/internal/store"
@@ -72,14 +71,16 @@ import (
 // safe (the wired stt/tts/router/api all are); called from per-conn
 // goroutines on the voice.Server.
 type reactiveHandler struct {
-	stt      stt.Transcriber
-	tts      tts.Synthesizer
-	router   *router.Router
-	embedder router.Embedder // reused for note write/query (same model as the classifier)
-	// boundary — the embedded seed sets behind the personal boundary
-	// (personalboundary.go). Zero value is usable and loads on first query;
-	// with no embedder it never loads and the boundary uses personalMarkers.
-	boundary personalBoundary
+	stt    stt.Transcriber
+	tts    tts.Synthesizer
+	router *router.Router
+
+	// recall — the note-and-fact recall subsystem: the embedder, the vector
+	// store it writes into, the personal boundary, and the two numbers that
+	// gate an answer. Grouped rather than spread across the handler because a
+	// handler that recalls needs all five and a handler that does not needs
+	// none of them (Vikunja #433, docs/handler-wiring.md).
+	recall recallWiring
 	// api — the CoreAPI the handler reads and writes through. Wired with the
 	// bare store adapter and UPGRADED by main once the daemonAPI exists; see
 	// upgradeAPI.
@@ -123,18 +124,8 @@ type reactiveHandler struct {
 	weatherProvider weather.Provider
 	weatherLocation string // default location for weather queries
 
-	memStore  memory.Store
 	dataStore *store.Store // direct store access for event extraction + pattern detection
 
-	// queryMinScore — the note-recall confidence gate. Top cosine below this ⇒
-	// "I don't know" instead of a guess. Tuned for the ONNX embedder; a knob, not
-	// load-bearing math (same posture as the presence thresholds). Set by
-	// wireVoice from VoiceConfig; default 0.55.
-	queryMinScore float64
-	// queryMinMargin — the second half of that gate: how far the top hit must
-	// beat the runner-up. 0 ⇒ margin off.
-	queryMinMargin float64
-
 	// timeParser — used as a fallback for stage-0 reminder grammar matches
 	// (where the extractor didn't run). Shared with the router's extractor.
 	// The production dateparser will replace StubDateTimeParser here too.

cmd/mavend/voicewire.go

@@ -262,19 +262,18 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 
 	// ----- the handler (the reactive path; closes over stt / tts / router / coreAPI / memory) -----
 	h := &reactiveHandler{
-		stt:      transcriber,
-		tts:      synthesizer,
-		router:   rtr,
-		embedder: emb,
-		api:      coreAPI,
-		tools:    exec,
-		matcher:  matcher,
-		replier:  replier,
-		phraser:  phr,
-		now:      time.Now,
-		feedsOn:  cfg.Feeds != nil,
-		home:     w.home,
-		netscan:  w.netscan,
+		stt:     transcriber,
+		tts:     synthesizer,
+		router:  rtr,
+		api:     coreAPI,
+		tools:   exec,
+		matcher: matcher,
+		replier: replier,
+		phraser: phr,
+		now:     time.Now,
+		feedsOn: cfg.Feeds != nil,
+		home:    w.home,
+		netscan: w.netscan,
 		// nil unless `crawl.on_demand` is on: reading a page he names is a
 		// capability, and capabilities are off unless configured.
 		crawler: onDemandCrawler(cfg),
@@ -283,18 +282,21 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		search: wireSearch(cfg),
 		// nil unless a `kiwix` block names a server. Same swap-aware client the
 		// router and replier use, so the rewriter follows a model swap.
-		kiwix:            wireKiwix(cfg, llmClient),
-		weatherProvider:  weatherProvider,
-		weatherLocation:  weatherLocation,
-		memStore:         memStore,
+		kiwix:           wireKiwix(cfg, llmClient),
+		weatherProvider: weatherProvider,
+		weatherLocation: weatherLocation,
+		recall: recallWiring{
+			embedder:  emb,
+			memStore:  memStore,
+			minScore:  cfg.Voice.QueryMinScore,
+			minMargin: cfg.Voice.QueryMinMargin,
+		},
 		dataStore:        dataStore,
 		dialogueSessions: dialogueSessions,
 		clarifyStore:     clarifyStore,
 		// 0 here (unset config) ⇒ the dialogue default.
 		clarifyMaxAttempts: cfg.Voice.ClarifyMaxAttempts,
 		extractor:          router.Extractor{Time: timeParser, Acts: matcher, Facts: router.DefaultFactParser{}},
-		queryMinScore:      cfg.Voice.QueryMinScore,
-		queryMinMargin:     cfg.Voice.QueryMinMargin,
 		timeParser:         timeParser,
 		ecosystem:          eco,
 	}

docs/evals/2026-08-04-recall-e5-small.md

@@ -0,0 +1,88 @@
+# Note recall after the e5-small swap — 04-08-2026
+
+Closes Vikunja #371, which asked for the swap and for this re-measurement. The embedder is no
+longer paraphrase-multilingual-MiniLM-L12-v2. It is **multilingual-e5-small**, quantized, with the
+`query:` / `passage:` prefixes it was trained with (`internal/router/onnxembedder.go`,
+`EmbedQuery` / `EmbedPassage`). `deploy/mavend.json` loads
+`models/embedder/multilingual-e5-small/model_quantized.onnx`, which is the same file `make
+download-embedder` fetches and the same file this run measured.
+
+- Fixture + scorer: `internal/memory/recalleval/` — 32 cases now, not 30
+- Reproduce: `make eval-recall`
+- Commit: `b6abb19`
+- Gate as deployed: `query_min_score` 0.55, `query_min_margin` 0.008
+
+The fixture grew since 31-07, so the case counts are not comparable row for row. The percentages
+are.
+
+## Results
+
+| | 31-07 MiniLM (onnx) | 04-08 e5-small (onnx) |
+|---|---|---|
+| **recall@1** | 60.0% (15/25) | **70.4% (19/27)** |
+| recall@3 | 80.0% (20/25) | **85.2% (23/27)** |
+| **answered after the gate** | 48.0% (12/25) | **63.0% (17/27)** |
+| **false recall** | 1/5 (20%) | **0/5** |
+| wrong note on top / tie on top | 10 / 0 | 8 / 0 |
+| ranked first, then silenced by the gate | 3 | 2 |
+| `hard` cases passed | 2/11 | 5/12 |
+| RU / EN passed | 13/24 / 3/6 | 18/26 / 4/6 |
+| latency p50 / p95 / max | 59ms / 148ms / 194ms | **23ms / 41ms / 62ms** |
+
+The hash ratchet CI runs is unchanged in kind and still answers nothing after the gate: recall@1
+37.0%, recall@3 74.1%, 0/27 answered, 0/5 false. It is lexical and exists so CI has a deterministic
+floor. Never compare a hash number to an ONNX one.
+
+## Findings
+
+### 1. The swap paid on every axis at once, including latency
+
+Ten points of recall@1, fifteen points of *answered*, the one false recall gone, and it is 2.5×
+faster because the quantized e5-small is 118MB against the 470MB fp32 file the old config loaded.
+Finding 3 of the 31-07 eval predicted the recall half and said nothing about speed; the speed came
+from fixing the second half of that finding, which was that the deployed path loaded a different
+file than the download target.
+
+The concrete case that eval named is fixed. "из-за чего кончилось место" no longer returns the
+guitar-chords filler note. It now returns a homelab note, `n2` at 0.884, and the wanted note is
+still not in the top 3 — so the query moved from absurd to merely wrong. That is the shape of what
+is left.
+
+### 2. The score distributions still overlap. The margin is what separates them
+
+This is the part of #371's premise that did not come true. Right-note-first top-1 scores run
+0.791 / 0.857 / 0.890 (min / median / max). Must-stay-silent top-1 scores run 0.795 / 0.815 /
+0.835. The silent cases sit *inside* the answering range, so no value of `query_min_score` keeps
+every real recall and rejects every false one — the same verdict as 31-07, at a higher and tighter
+band of scores.
+
+What separates them is the second-place gap. Margin top1-top2 for a right first hit: median 0.024.
+For a must-be-silent case: median 0.002, max 0.019. A false recall is a note that beats its
+neighbours by nothing, because nothing in the store is about the question. The sweep:
+
+| margin | answered | false recall |
+|---|---|---|
+| 0.000 | 18/27 (67%) | 3/5 |
+| 0.005 | 17/27 (63%) | 1/5 |
+| **0.008 (deployed)** | **17/27 (63%)** | **0/5** |
+| 0.010 | 15/27 (56%) | 0/5 |
+| 0.015 | 12/27 (44%) | 0/5 |
+
+0.008 is the knee: it is the smallest margin that silences all five, and the next step up costs two
+real answers for nothing. The score gate contributes almost nothing on its own — every value from
+0.00 to 0.70 answers the same 18 and admits the same 3 — so `query_min_score` is now close to inert
+and the margin is the live dial. Leave both where they are; #412 is where a further sweep belongs.
+
+### 3. What is left is a retrieval problem, not a gate problem
+
+Eight cases put the wrong note on top, and the failures cluster: `hard` 5/12, `preference` 5/9,
+`homelab` 8/13. Four of the eight have the right note in the top 3, so a reranker would collect
+them; the other four do not, so nothing downstream can. Two more rank first and are silenced by the
+margin — `en-hard-024` at 0.826 with margin 0.023, and `ru-home-026` at 0.846 with margin 0.001,
+which is a genuine near-tie against a second note that is also plausible.
+
+Preference queries are the weakest class in a way that is not about the model. "когда запускать
+резервное копирование" and "как мне присылать оповещения" both return a fact, not the note that
+states the preference. Facts and notes are searched in one pass since #373, so a confidently-scored
+fact wins a question that a note answers better. That is a ranking policy question and it belongs
+in its own task, not in a threshold.

docs/handler-wiring.md

@@ -0,0 +1,62 @@
+# How reactiveHandler is wired
+
+*Last verified: 2026-08-04 @ b6abb19. Living doc: correct it in place, do not append.*
+
+The decision Vikunja #433 asked for, and the rule that follows from it.
+
+## The decision
+
+**Group the fields into cohesive wiring structs. No container, no generator, no
+framework.** `reactiveHandler` (`cmd/mavend/voice.go`) stays the one type the voice
+server talks to. What changes is that a capability arrives as one named group, not as
+four more loose fields on a struct that already had thirty.
+
+The pattern was already in the file before this was written down: `searchWiring`,
+`kiwixWiring`, `homeWiring`, `netWiring` and `ecosystemWiring` are all this shape, each
+`nil` when the capability is off. #433 makes it the rule rather than a habit, and adds
+the case the habit had missed — a group that is not a capability toggle.
+
+## The worked example
+
+`recallWiring` (`cmd/mavend/recall.go`) holds the five things the recall path needs:
+the embedder, the vector store, the personal boundary, and the score and margin that
+gate an answer. They used to sit in three separate places on the handler with the two
+gate numbers a hundred lines away from the store they gate.
+
+Its zero value means "no recall", which is why it is a value and not a pointer. The
+`*Wiring` types that model an optional capability stay pointers, because `nil` is how
+"not configured" is spelled and a zero-valued search client would be a client pointed at
+nothing.
+
+## Why not the alternatives
+
+**Narrow consumer-side interfaces at each handler** is the more idiomatic Go answer and
+it is not rejected, only deferred. It is the right move at the point a handler is pulled
+into its own package, because that is when the import direction starts to matter. Doing
+it first would mean writing an interface per handler against a struct nobody can pass
+anywhere, which is churn bought against a package split that has not happened.
+
+**A container or a wire-style generator** is rejected outright. This is one binary with
+one composition root (`wireVoice` in `cmd/mavend/voicewire.go`). Generated wiring would
+add a build step and a layer of indirection to solve a problem that is currently one
+composite literal long, and it would make the "is this capability configured" question
+harder to answer by reading, which is the question this file is mostly about.
+
+## What this unlocks
+
+The reason `cmd/mavend/` cannot split into `mavend/actions/` today is that every action
+handler is a method on a struct with thirty unexported fields: moving handlers to a
+subdirectory means exporting all of them or inventing an interface to pass through. That
+was the answer given on the tick.go and voice.go splits, and it is still true. Grouping
+is the step that makes it false later — a handler that takes `recallWiring` and nothing
+else can move without the other twenty-five fields following it.
+
+## The rule
+
+**A wiring change does not ride a feature PR.** The voice.go and tick.go splits were
+safe to merge because the moved code diffed identical, line for line. A regrouping that
+touches the confirm gate or the act allowlist is its own change, reviewed on its own, or
+it is not reviewable at all.
+
+New capability, new group. A capability that adds four fields to `reactiveHandler`
+instead of one struct is the thing this decision exists to stop.

docs/plans/10-speaker-recognition.md

@@ -103,9 +103,14 @@ query path.
 ```
 
 `Recognizes()` requires both `enabled` and a `model_path`, so a half-filled block reads as off
-rather than as a capability that fails every turn. With `enabled` and no model the daemon still
-attaches the three methods — profiles can be created, listed and deleted — and logs that
-recognition is blocked.
+rather than as a capability that fails every turn.
+
+That gate was written, documented, and then never called. It is called now, and the behaviour it
+describes changed with it. `enabled` with no `model_path` used to attach all three methods and log
+that enrolment was on. Today `newSpeakerWiring` returns nil, so the methods are absent, and the log
+says why: there is nothing to embed with, so enrol, list and forget would all be no-ops. That is
+the one config shape where the operator most needs to be told otherwise, and it was the shape that
+lied.
 
 ## Still open
 

docs/plans/15-board-surface.md

@@ -0,0 +1,119 @@
+# Plan: The work board surface
+
+**The decision Vikunja #431 asked for. Written 04-08-2026.**
+
+**Verdict: build it, in a smaller shape than the task imagined.** The board is worth
+moving out of the file. The intake form belongs on the `/tasks` page, not on the voice
+path. The argument is not built, now or later.
+
+## Why it is worth building
+
+The reason is the one the task gives and it holds: company rules forbid pointing Claude
+at work repos, and Maven is the one assistant on the box that work material may reach.
+No telemetry, no cloud model, no third-party account. That is not a preference here, it
+is the whole permission.
+
+The build is also small, because most of it landed already:
+
+| Piece | Where | State |
+|---|---|---|
+| task rows, dedupe, status lifecycle | `internal/store/migrations.go:149` and migration #15 | done |
+| capture from speech, urgency stripped | `router.ParseTaskCapture`, `router.TaskCaptureGrammar` | done |
+| recite the list on request | `router.IsTaskListQuery` | done |
+| a page to read and change the board | `/tasks` in `cmd/mavweb` | done |
+| counting a shape without judging it | `internal/memory/behavior.go` | done, as precedent |
+| a proposal he reads when he chooses | `/routines`, the proposed-routine queue | done, as precedent |
+
+`tasks` already carries `status` (candidate, open, done, dropped), `due_ts`, `weight`,
+`source`, `evidence`, `ext_id` and `resolved_by`. Three things are missing. It has no
+definition of done and no blocked-on. There is no way to edit a task after capture:
+`SetTaskStatus` moves the status and nothing writes text, date or weight again. And
+there is no grouping he controls, because order is computed by `tasks.Rank` alone.
+
+## Where the form lives, and why not voice
+
+The task asks the form to refuse a capture with no definition of done. That refusal
+cannot live on the voice path, for two reasons.
+
+**The parked-state mechanism is binary.** `resolveConfirm` in `cmd/mavend/confirm.go`
+answers yes or no against a slot with a 90-second life. Filling four fields over four
+turns is slot filling, which is a different mechanism and a new one. Nothing in the
+daemon does it today.
+
+**The definition of done is the worst possible field to dictate.** It is the one string
+that has to be exact, because its whole purpose is to be unarguable later. Whisper
+transcribing a sentence of Russian work vocabulary is where exactness goes to die, and
+the capture path already had to strip a question mark that whisper invented.
+
+So: voice captures a line and recites the list. The page is where a line becomes an
+item with a definition of done, a blocked-on and a date. A captured line lands as
+`candidate` and stays there until it is filled in, which is what `candidate` was for.
+
+The refusal the task wants survives, moved: the page will not promote a candidate to
+`open` without a definition of done, the same way `ParseTaskCapture` will not file a
+marker with nothing after it. And the field must close on either outcome, so "it already
+works" counts as complete. A definition of done that only one result satisfies is a wish.
+
+## Does the stage-0 trick stretch
+
+The task asks this before any shape is committed to. It was checked. The answer is
+partly.
+
+`TaskCaptureGrammar` matches every utterance and lets `ParseTaskCapture` decide inside
+`Build`, keeping the intent at `note` and leaving the frozen seven-intent contract alone.
+That trick stretches to **recite** and to **status change**: both are a marker plus a
+referent, both are a lookup, and a status change is a small closed verb set over a list
+he can see. It does not stretch to **intake**, because intake is not one utterance, and
+it does not need to, because intake moved to the page.
+
+One cost to name. Each such grammar matches everything and runs its parser on every
+turn, ahead of the resident model. Two more of them is fine. A dozen would make stage 0
+a second router with no evaluation behind it, and at that point the frozen enum is the
+smaller problem.
+
+## What is not built: the argument
+
+Not now and not later behind a flag. The task is right about why, and
+`internal/memory/behavior.go` already argued it for habits: a 1.7B asked whether evidence
+proves anything will agree fluently and launder a guess into a decision. A wrong claim
+about his work, stated confidently, is the most expensive kind of wrong Maven can be.
+
+The line is the same line behaviour memory drew. She may **count**:
+
+- no state change in eleven days
+- blocked on a person, with no date
+- four of nine waiting on two people
+
+Those are queries over rows. She may not assess whether a build proves anything, whether
+a blocker is real, or whether a task should be dropped.
+
+## Persona
+
+A progress tracker is a nag by default, and "not a nag" is hard. The line is already
+drawn twice in the codebase and it is drawn the same way here:
+
+- A date he set becomes a reminder. He set it, so it is not her raising it.
+- A stall becomes a proposal he reads when he chooses, on a page, like `/routines`.
+- Ask what is on the board and she recites. She never opens with it.
+
+The day plan is the place to watch. `tickLoop.dayPlan` reads calendar events, pending
+reminders and checklist facts, and it does not read tasks. Adding the board to the
+morning nudge is exactly the move that turns this into a nag, so the board goes on the
+page and into the answer when asked, and not into the unprompted morning message.
+
+## The build, as tasks
+
+1. Two columns on `tasks`: definition of done, and blocked-on. Blocked-on resolves
+   through Nexus like any other person reference, because identity lives in Nexus.
+2. An edit path. Today a task is write-once except for its status, so the form has
+   nothing to save into.
+3. `/tasks` grows the form: promote candidate to open only with a definition of done,
+   set a date, set blocked-on. A date set here writes a reminder.
+4. A status-change grammar at stage 0, following `TaskCaptureGrammar`.
+5. Counted stall shapes on `/tasks`, phrased as counts. No assessment.
+
+Note for whoever picks up 3: `/tasks` accepts its POST without the step-up gate, while
+`/routines` and `/tools` require a passkey. That was deliberate for capture. Adding an
+edit path is the moment to re-argue it, not to inherit it silently.
+
+Each is separable and each is worth stopping after.

docs/plans/16-ambient-calendar.md

@@ -0,0 +1,87 @@
+# Plan: What the ambient calendar path should be
+
+**The decision Vikunja #432 asked for. Written 04-08-2026.**
+
+**Verdict: keep the endpoint, change the contract.** The relay app sends structured
+fields, not a notification blob. The free-text parser stays as the degraded path, because
+there is a real case where the phone cannot produce structure. Delete the endpoint only
+if the answer to the one open question below is no.
+
+## First, the task's premise is out of date
+
+#432 states as confirmed that every ambient event lands on the day the notification was
+posted, because there is no date parsing at all. That was true when the task was filed
+and it is not true now.
+
+`4e4c917` added `dayWords` and `dayOffset` (`internal/calendar/ambient.go:53`), so
+"завтра в 15:00" now dates to tomorrow. The same commit added `ambientPastGrace`, which
+refuses an event landing more than two hours before the notification, on the reasoning
+that the day was inferred and a stale inference is wrong rather than late. `45a5e37`
+(#482, this week) fixed a second dating bug the task did not know about: the wall clock
+was resolved against the notification's own zone, so every ambient meeting on a non-UTC
+box landed off by the deploy's UTC offset.
+
+What is still missing is an explicit date. "5 августа в 15:00" and "12.08 15:00" carry no
+day word, so they date to today and the past-grace check drops them. That is a smaller
+defect than the one filed, and it fails safe rather than storing a wrong meeting.
+
+The task's third question also has an answer, and the answer is yes. `internal/morning/plan.go:174`
+prefixes an uncertain item with "похоже, " and `cmd/mavend/actions_query.go:334` carries
+`Confidence < 1.0` into the calendar recital. Both readers hedge.
+
+## The open question, and it is the owner's
+
+**Can the phone read Android's calendar provider, or only the notification text?**
+
+Everything follows from this and nothing in this repo can answer it.
+
+A `NotificationListenerService` sees a title and a body. It cannot know a meeting's real
+start, end or organiser, because those are not in the notification. So if the relay is
+limited to the notification stream, free-text parsing on this side is not a choice, it is
+the only thing available, and #432's suggestion that the phone send structured JSON
+cannot be honoured.
+
+If the app may instead read `CalendarContract`, it has the actual event rows, and the
+whole parser stops being necessary. That is the better shape by a wide margin: a real
+start and end, a real title, an explicit date, no clock-reading heuristic and no
+past-grace guard, because nothing is inferred.
+
+Reading the phone's calendar provider does not break the constraint the design was built
+around. The refusal in `internal/calendar/ambient.go:10` is about holding a work
+credential **on the homelab**, which is what ties the box's blast radius to the employer.
+The phone already holds that session. Nothing new lands on homesrv either way.
+
+**Assumption, and it needs his answer:** a managed work profile may block a third-party
+app from reading work calendar rows. If it does, the notification stream is all there is.
+
+## The decision
+
+**Keep the endpoint.** Deleting it costs the parser, the tests and the wg-facing token,
+and buys nothing while the question above is open. It is off unless `-ambient-token` is
+set, so an unbuilt relay carries no surface today.
+
+**Make structured the primary shape.** `/api/ambient` should accept an event with an
+explicit start, end and title, and store it without parsing anything. Confidence stays
+below 1.0 and the source stays `ambient:notif`, because the provenance claim is unchanged:
+this is the phone telling Maven what it sees, not Maven reading a calendar.
+
+**Keep the free-text shape as the degraded path.** It is what a notification-only relay
+can send, and it is already written and tested.
+
+**Delete it instead if** the relay is not going to be built. That is the one answer that
+closes this without code, and it is his to give.
+
+## What not to do
+
+Do not add date parsing to the free-text path yet. That is the patch #432 explicitly
+refuses to accept as closure, and it is the wrong order: if the relay can send a date, no
+date parser is needed, and if it cannot, the parser is guessing at a date from text that
+was never meant to carry one.
+
+## Follow-on, unrelated to the decision
+
+`cmd/mavweb/ambient.go:33` records a known gap worth keeping visible: an ambient meeting
+writes `calendar_event_*` and never `calendar_busy`, so it is good enough to recite and
+not good enough to suppress a nudge. That is backwards. Suppressing a nudge is the
+lower-risk use of a low-confidence signal, and reciting one is the higher-risk use. It
+needs an expiry on the busy level, so it is its own task either way.

docs/plans/17-where-the-voice-loop-runs.md

@@ -0,0 +1,60 @@
+# Plan: Where mavwaked and mavenclient run
+
+**The decision Vikunja #463 asked for. Written 04-08-2026.**
+
+**Verdict: not in compose on homesrv. They run on a client machine in the room he is in.**
+The transport for that already exists and nothing needs building to allow it. What needs
+building is a way to check the wake path at all, which is a separate task.
+
+## The state that prompted this
+
+`docker-compose.yml` runs mavend, mavsttd, mavttsd, mavweb and mavpoll. `mavwaked` and
+`mavenclient` appear in no compose file and run as no host process. Both build under
+`make build`. So the wake word and the voice-activity gate are untested by construction:
+QA session 1 step 2 covers push-to-talk from `/dash` only, and #287 (voice session
+quality) can never be more than half-answered while this holds.
+
+## The reason is not hardware
+
+homesrv has a microphone. It is a Lenovo IdeaPad 5 Pro, and `/proc/asound/cards` lists
+the ACP digital mic array with capture devices at `/dev/snd/pcmC1D0c` and
+`/dev/snd/pcmC2D0c`. Adding `/dev/snd` to compose and joining the `audio` group would
+work.
+
+It would also be pointless. A wake-word daemon is worth having in the room he is standing
+in. homesrv is a server, so its microphone hears the room the server is in, which is not
+where anybody talks to Maven. Wiring audio into a container to listen to an empty room is
+work spent on a capability nobody can use.
+
+## The reason it belongs off-box
+
+`mavenclient` is a client by name and by design: microphone, then STT, then core, then
+TTS. It is the one binary in the tree meant to run somewhere else. `mavwaked` is the gate
+in front of it, so it goes wherever the microphone goes.
+
+Maven already has components that are not containers on homesrv. `mavupdate` is a
+host-side tool. The resident model, STT and TTS prefer the workstation and fall back to
+homesrv (`docs/offload.md`). Off-box is a shape this system already has.
+
+**And the wire already supports it.** `ipc.Dial` takes a netaddr seam address:
+a bare path is the unix socket, and `tcp://host:port?token=...` reaches a core on another
+host, with the token checked in `internal/netaddr` before `internal/ipc` sees the
+connection. So a client machine reaching mavend over wg or the LAN needs no new protocol
+work. It needs mavend to listen on TCP, which `deploy/mavend.json` does not currently ask
+for.
+
+## What this means for the QA plan
+
+QA session 1 step 2 should say what it actually covers, which is push-to-talk through
+`/dash`. It should not read as though it covers the voice loop. The wake path is checked
+on the client machine or it is not checked, and today there is no client machine.
+
+That is the honest state, and it is worse than the task suggests: this is not a
+configuration gap that a compose entry closes. Until a machine with a microphone runs
+`mavwaked` and `mavenclient` against a TCP-listening mavend, `internal/wake` and
+`cmd/mavwaked`'s VAD are covered by their unit tests and by nothing else.
+
+## What was wrong in CLAUDE.md
+
+The daemon table lists all nine binaries with no column for where they run, which is how
+this went unnoticed for as long as it did. It now says which two are not on the box.

docs/qa.md

@@ -1,6 +1,6 @@
 # QA plan: checking Maven properly
 
-*Last verified: 2026-08-02 @ 20aa2d5. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-04 @ 58635f1. Living doc: correct it in place, do not append.*
 
 Written 2026-08-01, after the 35-PR stack landed and the box came back up.
 Refreshed 2026-08-02 against the live list, after PRs #85-#90.
@@ -105,9 +105,9 @@ turns look misaligned when they are not.
    **Passes** (02-08-2026, five turns): `я рада`, `поняла`, `помогла`,
    `проверила`, `записала`, `грустна`, `ты` throughout, no pet names.
 2. Press push-to-talk on `/dash`. Say `привет`. Confirm a spoken reply comes
-   back. This is the only check that covers mic to STT to core to TTS to
-   speaker as one path. It is also the path the eleven-day outage most likely
-   broke.
+   back. This covers browser mic to STT to core to TTS as one path. It does
+   **not** cover the wake word or the voice-activity gate, and no step here
+   does — see below.
 3. Say `тихий режим`. Expect `тихий режим включён. буду реже напоминать.` **Passes.**
 4. Say `выключи тихий режим`. Expect `тихий режим выключен.` Negation must win. **Passes.**
 5. Say `в комнате тихо`. Quiet mode must NOT flip. Confirm on `/history` that no
@@ -129,11 +129,18 @@ turns look misaligned when they are not.
    and stitch unrelated topics. Asked whether he should move flats, she opened
    with the weather. That is 287, and it is a phrasing problem, not a loop problem.
 
-**The wake path cannot be checked as deployed.** `mavwaked` and `mavenclient`
-appear in no compose file and run as no host process. Step 2 covers only
-push-to-talk, from `/dash` through mavsttd and mavttsd. Wake word and VAD
-are untested by construction. Decide whether they belong in compose or on a
-client machine, and say which in the deploy docs. Tracked as **463**.
+**The wake path cannot be checked here, and that is now the decision rather
+than a gap.** `mavwaked` and `mavenclient` appear in no compose file and run as
+no host process. They are not going to. They belong on a client machine in the
+room he is standing in, because homesrv's microphone is real and in the wrong
+room — **463**, written up in `docs/plans/17-where-the-voice-loop-runs.md`.
+
+So the wake word and the VAD gate are covered by their unit tests and by
+nothing else, and no session at this box changes that. Checking them needs a
+machine with a microphone running both binaries against a TCP-listening mavend.
+`ipc.Dial` already speaks `tcp://host:port?token=...`, so the work is a machine
+and a config line, not protocol work. Until then, **287** can only be
+half-answered, and step 2 above is push-to-talk, not the voice loop.
 
 **319's single-token bug is fixed** (01-08-2026). Single-word Russian utterances no longer come
 back as `не совсем поняла — можешь переформулировать?`. `привет` and `поужинал`
@@ -471,6 +478,25 @@ at any address (**478**). `allow_private` does work, measured both ways.
 `CaptureStart`. There is no `cmd/mavheard`, no mavweb route, and `mavenclient`
 never calls it (**480**). Two of its QA steps are also stale.
 
+**Four stale QA steps were rewritten on 04-08-2026** under **480**, against the
+code rather than against what the plans said. All four failed the same way: the
+daemon was right and the step described an older daemon.
+
+| Step | Said | Says now |
+|---|---|---|
+| 253/3 | boots with the methods unknown | refuses to boot, `config.go:1651` |
+| 253/10 | no `:transcript` note by default | true only with a summary present |
+| 255/5 | `speaker: enrolment on, recognition BLOCKED` | that line is gone, the capability stays off |
+| 252/3 | `vision: stored image <id-prefix>` | `vision: stored <id>`, `vision.go:199` |
+
+Two of them are worth reading past the correction. 253/10 was false in exactly
+the scenario 253/16 creates, because `writeNotes` saves the transcript whenever
+the summary is empty so a dead llama-server does not lose the meeting. And 255/5
+changed because `Recognizes()` was written as the gate, documented as one, and
+never called — calling it turned `enabled` with no model from a half-working
+capability into a refusal. Enrolling into a store nothing can match against is
+not a working half.
+
 **257, netscan.** Steps 2, 3 and 9 pass at unit level. Step 1 fails. Steps 4 to 8
 need the block enabled. Step 10 is Bluetooth and stays skipped.
 

docs/rearchitecture.md

@@ -1,6 +1,6 @@
 # Maven — Re-architecture (Qwen3 resident model, revised 2026-07-18)
 
-*Last verified: 2026-08-02 @ 7079a24. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-04 @ b6abb19. Living doc: correct it in place, do not append.*
 
 > Supersedes the classifier-first routing model. Agreed in a design session
 > after diagnosing that homesrv deploys with a **stub phraser** (no LLM
@@ -40,8 +40,10 @@ utterance
   are deferred until the main feature set is complete.
 - **Embedder demoted from router to tool** — it now backs `memory.search`
   (RAG) and gives the router a cheap "similar past notes/intents" hint. The
-  router no longer depends on it clearing a threshold. Upgrade MiniLM → bge-m3
-  for better RU retrieval later (model swap, not architecture).
+  router no longer depends on it clearing a threshold. The MiniLM upgrade is
+  done: it is multilingual-e5-small, asymmetric, with the `query:`/`passage:`
+  prefixes (Vikunja #371, `docs/evals/2026-08-04-recall-e5-small.md`). A
+  further swap is a model swap, not architecture (Vikunja #412).
 
 ### Router output
 - Constrained structured JSON action `{tool, args, escalate}` — NOT free-form