phraser: say "даже не знаю, что сказать" when there is nothing to say (V-397)

Review of #108: "поговорили." reads as a summary of a conversation that did
not happen. One exported constant now, so the Stub, the LLMPhraser fallback
and the daemon all say the same thing.

internal/voice/replier.go keeps its own copy — that is the separate replier
seam, not this one.

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

cmd/mavend/actions.go

@@ -40,6 +40,7 @@ import (
 	"context"
 	"log"
 
+	"github.com/kami/maven/internal/phraser"
 	"github.com/kami/maven/internal/router"
 )
 
@@ -58,10 +59,14 @@ func (h *reactiveHandler) actionChat(ctx context.Context, dec router.Decision) s
 	// Conversational: build history from dialogue session (prior user turns)
 	// and let the LLM respond from general knowledge + context.
 	history := h.chatHistory()
+	// The phraser hands back its own fallback text alongside the error, so the
+	// turn survives a dead server and the failure still reaches the log.
 	reply, err := h.phraser.PhraseChat(ctx, dec.Utterance, history)
 	if err != nil {
 		log.Printf("voice: chat: %v", err)
-		return "поговорили."
+	}
+	if reply == "" {
+		return phraser.ChatFallback
 	}
 	return reply
 }

cmd/mavend/actions_fact.go

@@ -7,6 +7,7 @@ import (
 
 	"github.com/kami/maven/internal/ipc"
 	"github.com/kami/maven/internal/router"
+	"github.com/kami/maven/internal/store"
 )
 
 // actionFact handles router.IntentFact: persist a tapped self-fact, index
@@ -15,14 +16,41 @@ func (h *reactiveHandler) actionFact(ctx context.Context, dec router.Decision) s
 	if !dec.Slots.HasKey {
 		return "не разобрала, что записать — попробуй иначе."
 	}
+	// A question is never a fact about him (#470). "какая последняя версия
+	// языка Go?" used to land here, and the value stored was whatever the
+	// model invented for it, at confidence 1.00, indexed for recall under the
+	// question's own text. Two such rows then claimed seven unrelated world
+	// questions through recall and silently disabled world answering.
+	//
+	// The routing error itself is not fixed here — the answer is to answer.
+	// Sending the turn down the query chain is what he asked for anyway, and
+	// it costs a mis-routed capture nothing: an explicit "запиши ..." is not
+	// question-shaped, so it never takes this branch.
+	if router.IsQuestionShaped(dec.Utterance) {
+		log.Printf("voice: fact write refused, utterance is a question: %q (key %q) — answering as a query",
+			dec.Utterance, dec.Slots.Key)
+		q := dec
+		q.Intent = router.IntentQuery
+		// The key the model extracted is its guess at what to store, not a
+		// fact he has. Left in place, queryFactByKey would read it back and
+		// claim the turn before any real source ran.
+		q.Slots.Key, q.Slots.HasKey = "", false
+		q.Slots.Value = ""
+		return h.actionQuery(ctx, q)
+	}
 	now := h.now()
 	req := ipc.WriteFactReq{
-		Ts:         now,
-		Kind:       "self",
-		Key:        dec.Slots.Key,
-		Value:      dec.Slots.Value,
-		Source:     "tap:voice",
-		Confidence: 1.0,
+		Ts:     now,
+		Kind:   "self",
+		Key:    dec.Slots.Key,
+		Value:  dec.Slots.Value,
+		Source: "tap:voice",
+		// Not 1.00 unconditionally any more (#470). A value he said is
+		// evidence; a value the model supplied for words he never said is a
+		// guess, and writing a guess at full confidence is the same mistake
+		// the act path already refuses under "LLM output is not
+		// authorization".
+		Confidence: factConfidence(dec.Utterance, dec.Slots.Value),
 		// Subject: the key doubles as the entity-resolution candidate —
 		// a voice-tapped fact's key is usually the thing/person it's
 		// about ("espresso_machine", "kate"), so queueing it for Nexus
@@ -36,17 +64,25 @@ func (h *reactiveHandler) actionFact(ctx context.Context, dec router.Decision) s
 		log.Printf("voice: write fact: %v", err)
 		return "не получилось сохранить факт."
 	}
-	// Index the fact utterance in long-term memory (best-effort, must not
-	// fail the fact write). Facts aren't in the notes table, so this is the
-	// only recall path for them — "когда я пил воду?" reads back from here.
+	// Index the fact in long-term memory (best-effort, must not fail the fact
+	// write). Facts aren't in the notes table, so this is the only recall path
+	// for them — "когда я пил воду?" reads back from here.
+	//
+	// The indexed text is the fact, not the utterance (#493). queryMemory
+	// returns a fact's stored text verbatim, so what goes in here is what he
+	// 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 vec, err := router.EmbedPassage(ctx, h.embedder, dec.Utterance); err != nil {
+		text := store.FactRecallText(dec.Slots.Key, dec.Slots.Value)
+		if vec, err := router.EmbedPassage(ctx, h.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{
-			"source": "voice",
-			"type":   "fact",
-			"text":   dec.Utterance,
-			"ts":     strconv.FormatInt(now.Unix(), 10),
+			"source":    "voice",
+			"type":      "fact",
+			"text":      text,
+			"utterance": dec.Utterance,
+			"ts":        strconv.FormatInt(now.Unix(), 10),
 		}); err != nil {
 			log.Printf("voice: memory insert fact: %v", err)
 		}

cmd/mavend/actions_query.go

@@ -434,10 +434,24 @@ func (h *reactiveHandler) queryMemory(ctx context.Context, t *queryTurn) (string
 		return "", false
 	}
 	text := hit.Meta["text"]
+	// The score cleared the gate and the topic still has to match (#470). A
+	// note about his slow network scored high enough to answer "почему небо
+	// синее?", because the right-note and must-be-silent score ranges overlap
+	// and no threshold sits between them.
+	if !memory.RecallAllowed(t.dec.Utterance, text) {
+		log.Printf("voice: recall %q rejected for %q: a world question and no shared topic word", text, t.dec.Utterance)
+		return "", false
+	}
 	// A note is phrased in Maven's voice; a fact is read back as it was
 	// stored.
 	if hit.Meta["type"] == "note" {
-		if reply, perr := h.phraser.PhraseQuery(ctx, t.dec.Utterance, []string{text}); perr == nil && reply != "" {
+		reply, perr := h.phraser.PhraseQuery(ctx, t.dec.Utterance, []string{text})
+		switch {
+		case perr != nil:
+			// Reading the note back verbatim beats the phraser's own fallback,
+			// which only wraps the same text in "вот что я нашла:".
+			log.Printf("voice: recall phrase: %v", perr)
+		case reply != "":
 			return reply, true
 		}
 	}
@@ -469,6 +483,12 @@ func (h *reactiveHandler) queryNotes(ctx context.Context, t *queryTurn) (string,
 	if !memory.ConfidentScores(noteScores, h.queryMinScore, h.queryMinMargin) {
 		return "", false
 	}
+	// Same topic veto as queryMemory above: the best note must be about what
+	// he asked, not merely the nearest vector in the index.
+	if !memory.RecallAllowed(t.dec.Utterance, notes[0].Text) {
+		log.Printf("voice: note %q rejected for %q: a world question and no shared topic word", notes[0].Text, t.dec.Utterance)
+		return "", false
+	}
 	texts := make([]string, len(notes))
 	for i, n := range notes {
 		texts[i] = n.Text
@@ -701,7 +721,7 @@ func (h *reactiveHandler) queryKiwix(ctx context.Context, t *queryTurn) (string,
 // not be sent to an upstream engine at all. The guard closes both holes with
 // the same test.
 func (h *reactiveHandler) queryPersonal(ctx context.Context, t *queryTurn) (string, bool) {
-	if !isPersonalQuery(t.dec.Utterance) {
+	if !h.isPersonalTurn(ctx, t) {
 		return "", false
 	}
 	log.Printf("voice: %q is about him and his own data did not answer it; not asking the world", t.dec.Utterance)
@@ -726,7 +746,9 @@ var personalMarkers = []*regexp.Regexp{
 	regexp.MustCompile(`(?i)\bdid\s+i\b`),
 }
 
-// isPersonalQuery reports whether the utterance asks about something of his.
+// isPersonalQuery — the offline floor under the boundary. Possession only, and
+// deliberately still narrow: it answers when there is no embedder to ask, and a
+// broad guess made blind is worse than a narrow one.
 func isPersonalQuery(utterance string) bool {
 	if utterance == "" {
 		return false
@@ -739,6 +761,23 @@ func isPersonalQuery(utterance string) bool {
 	return false
 }
 
+// isPersonalTurn — the boundary test. The seeds decide when the embedder is
+// there, which is every deployed box; the possession markers are the floor
+// underneath, for a handler with no embedder or a turn whose vector never got
+// 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 {
+		if personal > world {
+			log.Printf("voice: %q scores personal %.4f vs world %.4f", t.dec.Utterance, personal, world)
+			return true
+		}
+		return false
+	}
+	return isPersonalQuery(t.dec.Utterance)
+}
+
 // queryGeneral — general knowledge, the last source before giving up. It always
 // claims: either a model answers, or Maven names the gap, or she says she does
 // not know.

cmd/mavend/actions_query_personal_test.go

@@ -19,6 +19,10 @@ func TestIsPersonalQuery(t *testing.T) {
 		"when is my meeting",
 		"do i have anything today",
 		"did i take my vitamins",
+		// Speech, but only the forms possession already covers ("did i").
+		// The verb forms the floor cannot see are the seeds' job, scored in
+		// TestONNXPersonalBoundary.
+		"what did i say about backups",
 	} {
 		if !isPersonalQuery(s) {
 			t.Errorf("isPersonalQuery(%q) = false, want true", s)
@@ -33,6 +37,10 @@ func TestIsPersonalQuery(t *testing.T) {
 		"почему небо синее",
 		"столица франции",
 		"how do i boil an egg",
+		// The floor is possession-only by design: a speech verb it cannot see
+		// passes here and is caught by the seeds instead.
+		"что я говорил про бэкапы?",
+		"как я говорил, почему небо синее",
 		"",
 	} {
 		if isPersonalQuery(s) {

cmd/mavend/capture.go

@@ -100,7 +100,7 @@ func (l llmCompleter) Complete(ctx context.Context, system, user string) (string
 // grammar, or a llama-server too old to honour one, gets the plain text it used
 // to get rather than an empty meeting summary.
 func unwrapSummary(raw string) string {
-	s := stripThink(strings.TrimSpace(raw))
+	s := phraser.StripThink(strings.TrimSpace(raw))
 	start := strings.Index(s, "{")
 	end := strings.LastIndex(s, "}")
 	if start < 0 || end <= start {

cmd/mavend/factgate.go

@@ -0,0 +1,82 @@
+package main
+
+import (
+	"log"
+	"strings"
+	"unicode"
+)
+
+// ungroundedConfidence — what a self fact is worth when its value appears
+// nowhere in what he said. Below `query_min_score` is not the point (recall
+// gates on vector distance, not on this number); the point is that
+// `/history` and every future reader can tell a value he said from a value
+// the model supplied.
+const ungroundedConfidence = 0.6
+
+// factConfidence scores a self fact by whether its value is grounded in the
+// utterance it came from. Grounded stays 1.00, which is what a tapped fact
+// has always been worth. Ungrounded drops, and says so in the log.
+//
+// An empty value is grounded by definition: the key alone carries the fact
+// ("поужинал"), and there is nothing for the model to have invented.
+func factConfidence(utterance, value string) float64 {
+	if strings.TrimSpace(value) == "" {
+		return 1.0
+	}
+	if valueGrounded(utterance, value) {
+		return 1.0
+	}
+	log.Printf("voice: fact value %q is not in %q — writing at confidence %.2f",
+		value, utterance, ungroundedConfidence)
+	return ungroundedConfidence
+}
+
+// valueGrounded reports whether every word of value traces back to a word he
+// actually said. The comparison is on a 4-rune prefix, so the model's
+// normalization survives ("пил воду" → "вода") while an invented value
+// ("1.20" for a question about Go) does not.
+func valueGrounded(utterance, value string) bool {
+	said := factTokens(utterance)
+	words := factTokens(value)
+	if len(words) == 0 {
+		return true
+	}
+	for _, w := range words {
+		if !anyTokenMatches(said, w) {
+			return false
+		}
+	}
+	return true
+}
+
+func anyTokenMatches(said []string, w string) bool {
+	for _, s := range said {
+		if s == w || sameStem(s, w) {
+			return true
+		}
+	}
+	return false
+}
+
+// sameStem is inflection tolerance and nothing more: it compares all but the
+// last rune of the shorter word, and never fewer than three. Russian marks
+// case on the ending, so "пил воду" and the stored "вода" are the same word he
+// said, while "1.20" and "версия" are not. A word of three runes or fewer must
+// match outright, where a shorter prefix would match half the language.
+func sameStem(a, b string) bool {
+	ar, br := []rune(a), []rune(b)
+	shorter := min(len(ar), len(br))
+	n := shorter - 1
+	if n < 3 || len(ar) < n || len(br) < n {
+		return false
+	}
+	return string(ar[:n]) == string(br[:n])
+}
+
+// factTokens lowercases and splits on everything that is not a letter or a
+// digit, the same shape planTokens uses in the router.
+func factTokens(s string) []string {
+	return strings.FieldsFunc(strings.ToLower(s), func(r rune) bool {
+		return !unicode.IsLetter(r) && !unicode.IsDigit(r)
+	})
+}

cmd/mavend/factgate_test.go

@@ -0,0 +1,125 @@
+package main
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/ipc"
+	"github.com/kami/maven/internal/memory"
+	"github.com/kami/maven/internal/router"
+	"github.com/kami/maven/internal/tool"
+	"github.com/kami/maven/internal/voice"
+)
+
+func newFactGateHandler(t *testing.T, now time.Time) (*reactiveHandler, ipc.CoreAPI) {
+	t.Helper()
+	st := newTestStore(t)
+	api := ipc.NewStoreAPI(st)
+	emb := router.NewHashEmbedder(1024)
+	h := &reactiveHandler{
+		api:       api,
+		embedder:  emb,
+		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
+}
+
+// The write half of #470: a question routed to IntentFact must not become a
+// fact about him, and must not leave a vector behind for recall to serve.
+func TestActionFact_QuestionIsNotWritten(t *testing.T) {
+	ctx := context.Background()
+	h, api := newFactGateHandler(t, time.Now())
+
+	reply := h.actionFact(ctx, router.Decision{
+		Intent:    router.IntentFact,
+		Utterance: "какая последняя версия языка Go?",
+		Slots:     router.Slots{Key: "go_version", HasKey: true, Value: `"1.20"`},
+	})
+
+	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)
+	if err != nil {
+		t.Fatalf("memory search: %v", err)
+	}
+	if len(hits) != 0 {
+		t.Fatalf("the question was indexed for recall: %+v", hits)
+	}
+	// It went down the query chain instead. Nothing is configured to answer a
+	// world question in this harness, so "не знаю." is the honest outcome —
+	// what matters is that the turn was answered, not stored.
+	if reply == "" {
+		t.Fatal("the turn was neither stored nor answered")
+	}
+}
+
+// The capture that must survive the gate: an explicit instruction to record,
+// even though it contains an interrogative.
+func TestActionFact_ExplicitCaptureStillWrites(t *testing.T) {
+	ctx := context.Background()
+	h, api := newFactGateHandler(t, time.Now())
+
+	h.actionFact(ctx, router.Decision{
+		Intent:    router.IntentFact,
+		Utterance: "запиши что я пил воду",
+		Slots:     router.Slots{Key: "water", HasKey: true, Value: `"вода"`},
+	})
+
+	f, err := api.LatestFact(ctx, "water")
+	if err != nil {
+		t.Fatalf("an explicit capture was refused: %v", err)
+	}
+	if f.Confidence != 1.0 {
+		t.Errorf("confidence = %v, want 1.0 for a value he said", f.Confidence)
+	}
+	// #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)
+	if err != nil {
+		t.Fatalf("memory search: %v", err)
+	}
+	if len(hits) != 1 {
+		t.Fatalf("the fact was not indexed once: %+v", hits)
+	}
+	if got := hits[0].Meta["text"]; got != "water — вода" {
+		t.Errorf("indexed text = %q, want the fact", got)
+	}
+	if got := hits[0].Meta["utterance"]; got != "запиши что я пил воду" {
+		t.Errorf("utterance provenance = %q, want it kept alongside", got)
+	}
+}
+
+func TestFactConfidence(t *testing.T) {
+	cases := []struct {
+		utterance, value string
+		want             float64
+	}{
+		{"запиши что я пил воду", `"вода"`, 1.0},
+		{"я выпил кофе", `"кофе"`, 1.0},
+		{"поужинал", "", 1.0},
+		{"отметь что я полил кактус", `"полил кактус"`, 1.0},
+		{"какая последняя версия языка Go", `"1.20"`, ungroundedConfidence},
+		{"кто премьер Японии", `"Тонио Озаки"`, ungroundedConfidence},
+	}
+	for _, c := range cases {
+		if got := factConfidence(c.utterance, c.value); got != c.want {
+			t.Errorf("factConfidence(%q, %q) = %v, want %v", c.utterance, c.value, got, c.want)
+		}
+	}
+}
+
+func mustEmbedPassage(t *testing.T, h *reactiveHandler, text string) []float32 {
+	t.Helper()
+	vec, err := router.EmbedQuery(context.Background(), h.embedder, text)
+	if err != nil {
+		t.Fatalf("embed %q: %v", text, err)
+	}
+	return vec
+}

cmd/mavend/main.go

@@ -251,6 +251,7 @@ func run(args []string) error {
 				Listen:       cfg.Phraser.Listen,
 				NGpuLayers:   cfg.Phraser.NGpuLayers,
 				NCtx:         cfg.Phraser.NCtx,
+				CacheRAMMiB:  cacheRAMMiB(cfg.Phraser.CacheRAMMiB),
 				Timeout:      time.Duration(cfg.Phraser.Timeout),
 				LLMNudges:    cfg.Phraser.LLMNudges,
 				ContextBlock: contextBlockFn(cfg, time.Now),
@@ -525,6 +526,7 @@ func run(args []string) error {
 					Listen:       cfg.Phraser.Listen,
 					NGpuLayers:   cfg.Phraser.NGpuLayers,
 					NCtx:         cfg.Phraser.NCtx,
+					CacheRAMMiB:  cacheRAMMiB(cfg.Phraser.CacheRAMMiB),
 					Timeout:      time.Duration(cfg.Phraser.Timeout),
 					LLMNudges:    cfg.Phraser.LLMNudges,
 					ContextBlock: contextBlockFn(cfg, time.Now),
@@ -788,6 +790,22 @@ func personaFacts(cfg *config.Config) persona.Facts {
 	return f
 }
 
+// cacheRAMMiB resolves phraser.cache_ram_mib into the phraser's field. Unset
+// means 512 MiB and not "whatever the server does", because the server's own
+// default is 8 GiB of prompt cache and that is what put 7.9 GB of RSS and half
+// a gigabyte of swap on homesrv for a 1.1 GB model. A negative value is the
+// deliberate opt-out: no flag is passed, the server's default applies, and the
+// operator owns the consequence.
+func cacheRAMMiB(configured int) int {
+	if configured == 0 {
+		return 512
+	}
+	if configured < 0 {
+		return 0
+	}
+	return configured
+}
+
 // contextBlockFn returns the per-turn renderer of the shared context block.
 // Per turn, not once at startup, because the block states the current time.
 func contextBlockFn(cfg *config.Config, now func() time.Time) func() string {

cmd/mavend/personalboundary.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"context"
+	"log"
+	"math"
+	"sync"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// The personal boundary decides one thing: is this question about him. It used
+// to decide it by matching possession words, and that was the whole defect
+// behind Vikunja #495. "что я говорил про бэкапы?" is his data by definition —
+// nothing outside the box has ever heard him say anything — and it carried no
+// possession word, so it walked past the boundary into SearXNG and came back
+// answered out of a Habr article about somebody else's backups.
+//
+// The first fix was one more marker class, `я говорил|сказал|писал|…`, plus a
+// carve-out so "как я говорил, почему небо синее" stayed a world question. Both
+// halves are a lexicon, and a lexicon is the wrong instrument here: Russian
+// gives every verb a dozen surface forms, the preamble list has no end, and
+// every utterance the list misses is one that reaches the world. It also drifts
+// silently — a missing verb looks exactly like no bug.
+//
+// So the boundary asks the embedder instead. Two frozen seed sets — questions
+// about him, questions about the world — are embedded once, and the turn's own
+// query vector, already computed by queryEmbed upstream, is scored against
+// both. Nearest side wins. Word order, verb form and unseen phrasing stop
+// mattering, which is exactly what a lexicon could not do.
+//
+// Measured 03-08-2026 against multilingual-e5-small on 19 held-out utterances,
+// none of them a seed: 19 right (TestONNXPersonalBoundary). A 20th, "as i said,
+// what is the population of india", missed by +0.008 during the first pass and
+// is a world seed now, which is why it is not in the held-out set. True
+// positives clear the world side by +0.014 to +0.089 and the nearest true
+// negative sits at -0.005, so the gate is the sign of the difference and
+// nothing tighter: the margins are too thin to justify a threshold, and the
+// asymmetry favours claiming anyway. A false claim costs one honest "не знаю";
+// a false pass sends his life to an upstream engine.
+//
+// The embedder is the one model CLAUDE.md pins to homesrv permanently, and it
+// is what makes this affordable: no llama-server call, no network, one cosine
+// per seed against a vector the turn already has.
+
+// personalSeeds — questions about him. Frozen: they are scoring data, so
+// editing one moves the boundary and must be re-measured, not eyeballed. Cover
+// both classes the boundary owns, possession and first-person speech, in both
+// languages.
+var personalSeeds = []string{
+	"что я говорил про это",
+	"я тебе рассказывал об этом?",
+	"что я записал про врача",
+	"я упоминал эту тему?",
+	"что у меня сегодня",
+	"когда моя встреча",
+	"what did i say about this",
+	"did i mention this to you",
+}
+
+// worldSeeds — questions the world can answer, including the two shapes that
+// look personal and are not: a first-person preamble on a world question ("как
+// я говорил, ..."), and first person without possession ("что я могу
+// посмотреть вечером"). Refusing those is the opposite mistake and the older
+// comment on personalMarkers already named it.
+var worldSeeds = []string{
+	"почему небо синее",
+	"какая столица франции",
+	"как сварить борщ",
+	"кто написал эту книгу",
+	"what is the capital of france",
+	"how do i boil an egg",
+	"как я говорил, почему небо синее",
+	"as i said, why is the sky blue",
+	"as i said, what is the population of india",
+	"что я могу посмотреть вечером",
+	"что мне почитать про историю",
+	"что я должен знать про питон",
+	"what can i watch tonight",
+}
+
+// personalBoundary holds the embedded seeds. Zero value is usable and means
+// "not loaded yet"; a handler built without an embedder never loads and the
+// boundary falls back to personalMarkers.
+type personalBoundary struct {
+	once     sync.Once
+	personal [][]float32
+	world    [][]float32
+	loaded   bool
+}
+
+// load embeds both seed sets, once per process. Seeds are embedded on the QUERY
+// side, like the utterance they are compared with — a question against a
+// question. Mixing sides would measure the e5 prefix, not the meaning.
+func (b *personalBoundary) load(ctx context.Context, emb router.Embedder) {
+	b.once.Do(func() {
+		if emb == nil {
+			return
+		}
+		embedAll := func(ss []string) [][]float32 {
+			out := make([][]float32, 0, len(ss))
+			for _, s := range ss {
+				v, err := router.EmbedQuery(ctx, emb, s)
+				if err != nil {
+					log.Printf("voice: personal boundary seeds unavailable (%v); falling back to possession markers", err)
+					return nil
+				}
+				out = append(out, v)
+			}
+			return out
+		}
+		p, w := embedAll(personalSeeds), embedAll(worldSeeds)
+		if p == nil || w == nil {
+			return
+		}
+		b.personal, b.world, b.loaded = p, w, true
+	})
+}
+
+// score returns the best similarity to each side. ok is false when the seeds
+// are not loaded, which is the caller's signal to use the markers instead.
+func (b *personalBoundary) score(vec []float32) (personal, world float64, ok bool) {
+	if !b.loaded || len(vec) == 0 {
+		return 0, 0, false
+	}
+	best := func(seeds [][]float32) float64 {
+		m := -1.0
+		for _, s := range seeds {
+			if c := cosine(vec, s); c > m {
+				m = c
+			}
+		}
+		return m
+	}
+	return best(b.personal), best(b.world), true
+}
+
+// cosine — same math as internal/router and internal/memory, small enough that
+// importing one of them for it would be the larger coupling.
+func cosine(a, b []float32) float64 {
+	if len(a) != len(b) {
+		return 0
+	}
+	var dot, na, nb float64
+	for i := range a {
+		dot += float64(a[i]) * float64(b[i])
+		na += float64(a[i]) * float64(a[i])
+		nb += float64(b[i]) * float64(b[i])
+	}
+	if na == 0 || nb == 0 {
+		return 0
+	}
+	return dot / (math.Sqrt(na) * math.Sqrt(nb))
+}

cmd/mavend/personalboundary_test.go

@@ -0,0 +1,94 @@
+package main
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// A handler with no embedder never loads the seeds, so the boundary falls back
+// to the possession markers. That is the offline floor and it must keep working
+// — an embedder that fails to load must not open the boundary.
+func TestBoundaryFallsBackToMarkersWithNoEmbedder(t *testing.T) {
+	h := personalHandler()
+	if !h.isPersonalTurn(context.Background(), &queryTurn{
+		dec: router.Decision{Utterance: "во сколько у меня встреча"},
+	}) {
+		t.Error("no embedder: a possession question must still be personal")
+	}
+	if h.isPersonalTurn(context.Background(), &queryTurn{
+		dec: router.Decision{Utterance: "почему небо синее"},
+	}) {
+		t.Error("no embedder: a world question must still pass")
+	}
+}
+
+// TestONNXPersonalBoundary — the number that matters, scored against the
+// embedder homesrv actually runs. Opt-in via MAVEN_ONNX_LIB, exactly like
+// TestONNXRecall in internal/memory/recalleval.
+//
+// Every case here is held out: none of these strings is a seed. The #495
+// regression is the first row — "что я говорил про бэкапы?" reached SearXNG and
+// was answered from a Habr article, and no possession word appears in it.
+func TestONNXPersonalBoundary(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")
+	}
+	dir := filepath.Join("../..", "models/embedder/multilingual-e5-small")
+	emb, err := router.NewONNXEmbedder(filepath.Join(dir, "model_quantized.onnx"), filepath.Join(dir, "tokenizer.json"), lib)
+	if err != nil {
+		t.Skipf("onnx embedder unavailable: %v", err)
+	}
+	defer emb.Close()
+
+	cases := []struct {
+		utterance string
+		personal  bool
+	}{
+		{"что я говорил про бэкапы?", true},
+		{"что я сказал вчера про отпуск", true},
+		{"я писал что-нибудь про сервер", true},
+		{"я упоминал про конференцию?", true},
+		{"что я отмечал по поводу переезда", true},
+		{"я рассказывал тебе про новую работу?", true},
+		{"во сколько у меня встреча", true},
+		{"когда мой следующий отпуск", true},
+		{"what did i say about backups", true},
+		{"did i tell you about the doctor", true},
+		{"как я говорил, почему небо синее", false},
+		{"как уже я говорил, какая столица франции", false},
+		{"почему трава зелёная", false},
+		{"столица франции", false},
+		{"как мне сварить борщ", false},
+		{"что мне посмотреть вечером", false},
+		{"я хочу узнать про рим", false},
+		{"кто такой гагарин", false},
+		{"how do i boil an egg", false},
+	}
+
+	h := &reactiveHandler{embedder: emb}
+	ctx := context.Background()
+	wrong := 0
+	for _, c := range cases {
+		vec, err := router.EmbedQuery(ctx, emb, c.utterance)
+		if err != nil {
+			t.Fatalf("embed %q: %v", c.utterance, err)
+		}
+		turn := &queryTurn{dec: router.Decision{Utterance: c.utterance}, vec: vec}
+		got := h.isPersonalTurn(ctx, turn)
+		p, w, ok := h.boundary.score(vec)
+		if !ok {
+			t.Fatal("seeds did not load with a working embedder")
+		}
+		if got != c.personal {
+			wrong++
+			t.Errorf("%q: personal=%v want %v (personal %.4f world %.4f)", c.utterance, got, c.personal, p, w)
+		}
+		t.Logf("personal=%-5v personal %.4f world %.4f delta %+.4f  %s", got, p, w, p-w, c.utterance)
+	}
+	t.Logf("personal boundary: %d/%d held-out utterances correct", len(cases)-wrong, len(cases))
+}

cmd/mavend/replier_llm.go

@@ -2,122 +2,33 @@ package main
 
 import (
 	"context"
-	"encoding/json"
-	"strings"
-	"time"
 
-	"github.com/kami/maven/internal/llm"
-	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/phraser"
 	"github.com/kami/maven/internal/router"
 	"github.com/kami/maven/internal/voice"
 )
 
-// completer is the LLM seam for the replier (subset of router.Completer).
-// *llm.Client satisfies it.
-type completer interface {
-	Complete(ctx context.Context, r llm.Req) (string, error)
-}
-
-// llmReplier phrases reactive confirmations with the resident model
-// (Qwen3-1.7B). Stub is the
-// floor on any error (offline-safe). Maven speaks as "she", feminine RU.
+// llmReplier is the daemon-side wiring around phraser.Replier: it owns the
+// deterministic floor, and nothing else. The phrasing itself, the prompt and the
+// output parsing live in internal/phraser so the eval can score them (#396).
 type llmReplier struct {
-	c    completer
+	p    *phraser.Replier
 	stub *voice.StubReplier
-
-	// block renders the shared context block per turn (who he is, the time).
-	// nil ⇒ the prompt stands alone.
-	block func() string
 }
 
-func newLLMReplier(c completer, block func() string) *llmReplier {
-	return &llmReplier{c: c, stub: voice.NewStubReplier(), block: block}
+func newLLMReplier(c phraser.Completer, block func() string) *llmReplier {
+	return &llmReplier{p: phraser.NewReplier(c, block), stub: voice.NewStubReplier()}
 }
 
-const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Владелец — мужчина, говоришь с ним на "ты", в единственном числе; никогда не "вы"/"ваш" и не "он"/"его". Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), по-русски, спокойно и без официальных формулировок. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
-Пример: {"response": "Записала, что ты выпил стакан воды.", "mood": "neutral"}
-Никогда не пиши "..." в поле response.`
-
+// Reply never fails: a clarify, a model error and an unusable generation all
+// answer from the stub, which is what keeps a turn from breaking on the model.
 func (r *llmReplier) Reply(d router.Decision) string {
 	if d.Clarify {
 		return r.stub.Reply(d)
 	}
-	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
-	defer cancel()
-	out, err := r.c.Complete(ctx, llm.Req{
-		System:        persona.Prepend(r.block, replySystem),
-		User:          replyContext(d),
-		Grammar:       phraser.ResponseGrammar,
-		MaxTokens:     512,
-		RepeatPenalty: 1.3,
-	})
-	if err != nil {
+	out, err := r.p.PhraseReply(context.Background(), d)
+	if err != nil || out == "" {
 		return r.stub.Reply(d)
 	}
-	out = stripThink(out)
-	if response, _ := parseResponseMood(out); response != "" {
-		return response
-	}
-	// fallback: try plain-text parsing
-	if out = firstSentence(out); out != "" {
-		return out
-	}
-	return r.stub.Reply(d)
-}
-
-// firstSentence trims the model's output to a single clean confirmation: first
-// line, first sentence, whitespace-normalized — the last-line defense against a
-// small model that rambles past the first period despite the prompt + stop.
-// stripThink removes the <think> block that Thinking-variant models emit.
-func stripThink(s string) string {
-	if i := strings.LastIndex(s, "</think>"); i >= 0 {
-		s = strings.TrimSpace(s[i+8:])
-	}
-	return s
-}
-
-func firstSentence(s string) string {
-	s = strings.TrimSpace(s)
-	if i := strings.IndexByte(s, '\n'); i >= 0 {
-		s = s[:i]
-	}
-	// keep up to and including the first sentence-ending punctuation.
-	if i := strings.IndexAny(s, ".!?"); i >= 0 {
-		s = s[:i+1]
-	}
-	return strings.TrimSpace(s)
-}
-
-// parseResponseMood extracts {"response","mood"} from LLM output, tolerant
-// of thinking tokens and extra text before/after the JSON block.
-func parseResponseMood(raw string) (response, mood string) {
-	cleaned := strings.TrimSpace(raw)
-	start := strings.Index(cleaned, "{")
-	end := strings.LastIndex(cleaned, "}")
-	if start < 0 || end < 0 || end <= start {
-		return "", ""
-	}
-	var parsed struct {
-		Response string `json:"response"`
-		Mood     string `json:"mood"`
-	}
-	if err := json.Unmarshal([]byte(cleaned[start:end+1]), &parsed); err != nil {
-		return "", ""
-	}
-	return parsed.Response, parsed.Mood
-}
-
-// replyContext renders the decision into a compact RU description for the model.
-func replyContext(d router.Decision) string {
-	switch d.Intent {
-	case router.IntentFact:
-		return "записала факт: " + d.Slots.Key + " " + d.Slots.Value
-	case router.IntentNote:
-		return "сохранила заметку: " + d.Slots.Text
-	case router.IntentReminder:
-		return "поставила напоминание: " + d.Slots.Text
-	default:
-		return string(d.Intent) + ": " + d.Slots.Text
-	}
+	return out
 }

cmd/mavend/replier_llm_test.go

@@ -5,28 +5,22 @@ import (
 	"testing"
 
 	"github.com/kami/maven/internal/llm"
-	"github.com/kami/maven/internal/phraser"
 	"github.com/kami/maven/internal/router"
 	"github.com/kami/maven/internal/voice"
 )
 
-type mockCompleter struct {
+// The phrasing itself is tested in internal/phraser. What is left here is the
+// only thing the daemon adds: the stub floor, on the three ways a reply can
+// fail to arrive.
+type stubCompleter struct {
 	out string
 	err error
 }
 
-func (m mockCompleter) Complete(_ context.Context, _ llm.Req) (string, error) { return m.out, m.err }
+func (s stubCompleter) Complete(_ context.Context, _ llm.Req) (string, error) { return s.out, s.err }
 
-func TestLLMReplierReturnsLLMReply(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: `{"response":"записала, кофе закончился","mood":"neutral"}`}, nil)
-	got := r.Reply(router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}})
-	if got != "записала, кофе закончился" {
-		t.Errorf("got %q, want %q", got, "записала, кофе закончился")
-	}
-}
-
-func TestLLMReplierFallsBackToPlainText(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: "записала, кофе закончился"}, nil)
+func TestLLMReplierPassesTheModelReplyThrough(t *testing.T) {
+	r := newLLMReplier(stubCompleter{out: `{"response":"записала, кофе закончился","mood":"neutral"}`}, nil)
 	got := r.Reply(router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}})
 	if got != "записала, кофе закончился" {
 		t.Errorf("got %q, want %q", got, "записала, кофе закончился")
@@ -34,54 +28,30 @@ func TestLLMReplierFallsBackToPlainText(t *testing.T) {
 }
 
 func TestLLMReplierFallsBackToStubOnError(t *testing.T) {
-	r := newLLMReplier(mockCompleter{err: errTestLLMDown}, nil)
-	noteDec := router.Decision{Intent: router.IntentNote}
-	got := r.Reply(noteDec)
-	want := voice.NewStubReplier().Reply(noteDec)
-	if got != want {
-		t.Errorf("on llm error: got %q, want stub %q", got, want)
-	}
+	r := newLLMReplier(stubCompleter{err: errReplierTest}, nil)
+	assertStub(t, r, router.Decision{Intent: router.IntentNote}, "llm error")
 }
 
 func TestLLMReplierFallsBackToStubOnEmpty(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: ""}, nil)
-	noteDec := router.Decision{Intent: router.IntentNote}
-	got := r.Reply(noteDec)
-	want := voice.NewStubReplier().Reply(noteDec)
-	if got != want {
-		t.Errorf("on empty llm: got %q, want stub %q", got, want)
-	}
+	r := newLLMReplier(stubCompleter{out: ""}, nil)
+	assertStub(t, r, router.Decision{Intent: router.IntentNote}, "empty llm")
 }
 
 func TestLLMReplierClarifyUsesStub(t *testing.T) {
-	r := newLLMReplier(mockCompleter{out: "я всё поняла"}, nil)
-	clarifyDec := router.Decision{Clarify: true}
-	got := r.Reply(clarifyDec)
-	want := voice.NewStubReplier().Reply(clarifyDec)
+	r := newLLMReplier(stubCompleter{out: "я всё поняла"}, nil)
+	assertStub(t, r, router.Decision{Clarify: true}, "clarify")
+}
+
+func assertStub(t *testing.T, r *llmReplier, d router.Decision, what string) {
+	t.Helper()
+	got, want := r.Reply(d), voice.NewStubReplier().Reply(d)
 	if got != want {
-		t.Errorf("on clarify: got %q, want stub %q", got, want)
+		t.Errorf("on %s: got %q, want stub %q", what, got, want)
 	}
 }
 
-var errTestLLMDown = errTest("llm down")
+var errReplierTest = errTest("llm down")
 
 type errTest string
 
 func (e errTest) Error() string { return string(e) }
-
-// grammarRecorder captures the request so the grammar can be asserted on.
-type grammarRecorder struct{ req llm.Req }
-
-func (g *grammarRecorder) Complete(_ context.Context, r llm.Req) (string, error) {
-	g.req = r
-	return `{"response":"записала","mood":"neutral"}`, nil
-}
-
-func TestLLMReplierCarriesTheResponseGrammar(t *testing.T) {
-	rec := &grammarRecorder{}
-	r := newLLMReplier(rec, nil)
-	r.Reply(router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}})
-	if rec.req.Grammar != phraser.ResponseGrammar {
-		t.Errorf("grammar = %q, want phraser.ResponseGrammar", rec.req.Grammar)
-	}
-}

cmd/mavend/voice.go

@@ -76,6 +76,10 @@ type reactiveHandler struct {
 	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
 	// 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.

cmd/mavend/voicewire.go

@@ -146,6 +146,7 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		emb = router.NewHashEmbedder(1024)
 	}
 	w.embedder = emb
+	repairFactVectors(dataStore, emb)
 	checkStoredEmbedder(dataStore, emb)
 
 	// ----- tool executor (the enabled act allowlist, store-backed) -----
@@ -473,6 +474,36 @@ func seedTools(api ipc.CoreAPI, tools []config.ToolConfig) {
 	log.Printf("voice: seeded %d act tools from config", n)
 }
 
+// repairFactVectors brings stored fact vectors in line with the facts they name
+// (#493), once per box, before the embedder marker is even looked at.
+//
+// Automatic and not a flag, unlike -reembed: only voice-tapped facts are in
+// this index, so the work is tens of embeddings rather than the thousands of
+// notes that made the backfill a deliberate act. And the box that needs it is
+// broken in a way nobody can see — recall answers with the wrong text and
+// nothing logs an error — so waiting for an operator to know to run it is how
+// the defect survived four restarts in the first place.
+func repairFactVectors(dataStore *store.Store, emb router.Embedder) {
+	if dataStore == nil {
+		return
+	}
+	res, err := dataStore.RepairFactVectors(context.Background(),
+		// EmbedPassage, the stored side, same as every other writer of these
+		// vectors.
+		func(ctx context.Context, text string) ([]float32, error) {
+			return router.EmbedPassage(ctx, emb, text)
+		})
+	if err != nil {
+		log.Printf("voice: fact vector repair failed, no marker written and nothing half-done — retried next start: %v", err)
+		return
+	}
+	if res.Skipped || res.Rewritten+res.Dropped == 0 {
+		return
+	}
+	log.Printf("voice: fact vector repair — %d re-embedded from the fact they name, %d dropped as voided or superseded, %d already right, took %s (#493)",
+		res.Rewritten, res.Dropped, res.Kept, res.Took.Round(time.Millisecond))
+}
+
 // reembedOnStart is the -reembed flag (set in run()). Opt-in on purpose: see
 // runReembed.
 var reembedOnStart bool

cmd/mavend/worldmodel.go

@@ -54,7 +54,11 @@ func (h *reactiveHandler) phraseSource(ctx context.Context, name, utterance stri
 		log.Printf("voice: %s: no world model, reading the source back instead", name)
 		return ""
 	case err != nil:
+		// The resident phraser answers this call with its fallback text and the
+		// error together. Drop the text: these callers hold the passage itself
+		// and read it back better than "вот что я нашла: <passage>" does.
 		log.Printf("voice: %s: phrase: %v", name, err)
+		return ""
 	}
 	return reply
 }

deploy/mavend.json

@@ -20,6 +20,7 @@
     "bin_path": "llama-server",
     "n_gpu_layers": 99,
     "n_ctx": 4096,
+    "cache_ram_mib": 512,
     "timeout": "60s",
     "llm_nudges": false
   },

docs/evals/2026-08-03-llama-prompt-cache.md

@@ -0,0 +1,84 @@
+# Where the resident model's 7.9GB of RSS goes (2026-08-03, homesrv)
+
+Measured for Vikunja #499. The deployed llama-server held 7.9GB RSS for a 1.1GB
+model file. Half a gigabyte of it was in swap, on a box that also runs
+whisper.cpp, piper and the embedder.
+
+## Method
+
+`maven-mavend-1` was stopped for the measurement, with the owner's approval.
+Its own binary then ran on the host with the exact deployed command line. That
+binary is `/opt/maven/bin/llama-server`, version `1 (4c65955)`, a Vulkan build.
+
+```sh
+llama-server -m /mnt/hdd1/llms/qwen3/Qwen3-1.7B-UD-Q4_K_XL.gguf \
+  --host 127.0.0.1 --port 18099 -c 4096 -ngl 99 --no-webui
+```
+
+RSS was read from `/proc/<pid>/status` after load and after each of 8 distinct
+1521-token prompts. `smaps` of the deployed process was read first, from inside
+the container, since the host user cannot read another user's maps.
+
+## The cause: the prompt cache, not the weights and not the offload
+
+The startup log says it outright:
+
+```text
+srv load_model: prompt cache is enabled, size limit: 8192 MiB
+srv llama_server: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
+```
+
+The server saves the full KV state of every idle slot it evicts. It keeps up to
+8GiB of those states in host RAM (llama.cpp PR 16391). One saved prompt of 1521
+tokens costs 166.377 MiB. That is 112 kiB per token, exactly Qwen3-1.7B's KV
+footprint (28 layers x 2 x 1024 dims x 2 bytes).
+
+RSS at rest, and per distinct prompt:
+
+| Prompts served | RSS, default | RSS, `--cache-ram 512` |
+|---|---|---|
+| 0 (just loaded) | 443 MB | 411 MB |
+| 1 | 445 MB | 411 MB |
+| 4 | 958 MB | 929 MB |
+| 8 | 1641 MB | 932 MB |
+
+Uncapped, RSS climbs about 170MB per distinct prompt and does not stop until
+the 8GiB limit. Capped at 512 MiB it plateaus at 932MB from the fourth prompt
+on, with the cache holding steady at `3 prompts, 499.132 MiB` and evicting.
+
+The 7.9GB on the running daemon was that climb, weeks of it. Its `smaps` showed
+one 6.03GB anonymous mapping at 5.32GB resident plus a 1.45GB mapping at 1.27GB
+resident, and only 30MB of file-backed RSS.
+
+## The task's leading guess was wrong
+
+`-ngl 99` on the Vega iGPU costs almost no process RSS. A freshly loaded server
+has 95MB of anonymous RSS in total. RADV allocates device memory through the
+kernel, outside the process, and the log sees 8202 MiB free on `Vulkan0`. The
+weights are mmapped and file-backed, so they are evictable and do not pin RSS. The logit buffer is not visible in the numbers above at all.
+
+## Decision
+
+`--cache-ram 512` is now the default, wired as `phraser.cache_ram_mib` and set
+in `deploy/mavend.json`. 512 MiB caps total RSS near 1GB, an eighth of what the
+box carried. It still holds three of the 1521-token probes above. Maven's real
+routing and phrasing prompts are much shorter, so it holds more of those than
+the table suggests. `-c 4096` is untouched, as #499
+required. A negative `cache_ram_mib` passes no flag, for a llama-server too old
+to know it.
+
+Not changed: `n_parallel = 4`. With `kv_unified = true` the four slots share one
+4096-token KV cache, so they do not multiply it.
+
+The other half of #499 was that none of these lines were reachable. mavend
+scraped llama-server's stderr for the listen line and discarded it, and never
+piped stdout at all. Both streams now go to mavend's log with a `llama:` prefix.
+The last 12 startup lines go into the error when the server dies before it
+listens.
+
+## Deployed
+
+The `mavenai:latest` image was rebuilt and `maven-mavend-1` recreated the same
+day. The daemon's own log now carries the child's startup, it reads
+`prompt cache is enabled, size limit: 512 MiB`, and the resident server sat at
+439MB RSS after load and 613MB after one served turn.

docs/evals/2026-08-03-personal-boundary.md

@@ -0,0 +1,46 @@
+# Personal boundary, seed scoring vs possession markers, 2026-08-03
+
+Vikunja #495. `что я говорил про бэкапы?` walked past the personal boundary into
+SearXNG and came back answered from a Habr article. The boundary matched
+possession words only, so a first-person speech verb was not a personal
+question.
+
+## What changed
+
+The boundary now scores the turn's query vector against two frozen seed sets.
+It claims the turn when the personal side is nearer than the world side. Seeds
+and code are in `cmd/mavend/personalboundary.go`. The possession markers stay as
+the offline floor for a handler with no embedder.
+
+A regex speech class was written first and dropped. Russian gives every verb a
+dozen surface forms, and the "как я говорил, ..." preamble list has no end. Each
+form the lexicon missed was one more question reaching the world.
+
+## Measurement
+
+Embedder: multilingual-e5-small int8, the one homesrv runs. Both sides are
+embedded on the query side. Cases are held out, none of them a seed. `make test`
+runs the offline part. The scored part is opt-in through `MAVEN_ONNX_LIB`, like
+`TestONNXRecall`.
+
+    19/19 held-out utterances correct   (TestONNXPersonalBoundary)
+
+    true positive margins   +0.014 to +0.089
+    nearest true negative   -0.005  ("кто такой гагарин")
+
+One case missed during the first pass and is not held out any more: `as i said,
+what is the population of india`, +0.008 to the personal side. It is a world seed
+now.
+
+The gate is the sign of the difference and nothing tighter. The margins are too
+thin for a threshold. The asymmetry favours claiming: a false claim costs one
+honest "не знаю", a false pass sends his life to an upstream engine.
+
+`make eval-recall` unchanged, 18/27 answered at gate 0.55. Recall does not touch
+this path.
+
+## Not verified
+
+The live probe on the deployed box. The daemon was not rebuilt in this session.
+The reply to `что я говорил про бэкапы?` with no matching note is still untested
+against a real SearXNG.

docs/evals/2026-08-03-recall-topic-veto.md

@@ -0,0 +1,65 @@
+# Recall topic veto, what it costs and what it buys, 2026-08-03
+
+Vikunja #496. The task asked for a cross-language fix. Skip the topic veto in
+`memory.RecallAllowed` when the question and the hit are in different scripts.
+An English question would then stop losing a Russian note.
+
+No such case exists. No fixture case puts the question and its wanted note in
+different scripts. The case the task named is not one either.
+
+    en-hard-024
+      query  "what fixed the screen problem"
+      note   "the flicker went away once i swapped the display cable"
+
+Both are English. It is a paraphrase failure, not a language failure. A script
+test would not have changed a single case, and neither would a bilingual stem
+map.
+
+## What the veto is worth today
+
+Measured with the real embedder, multilingual-e5-small int8, gate 0.55, margin
+0.008. The first row is the veto as it ships. The second is `RecallAllowed`
+forced to true.
+
+| | cases passing | answered | false recall | silenced by gate |
+|---|---|---|---|---|
+| veto on | 22/32 | 17/27 | 0/5 | 2 |
+| veto off | 22/32 | 18/27 | 1/5 | 1 |
+
+The pass count does not move. The veto trades one true recall for one false one.
+It costs `en-hard-024` and it buys `ru-silent-029`:
+
+    ru-silent-029
+      query  "во сколько отходит поезд"
+      note   "погулял вдоль реки"        0.835, margin 0.019
+
+The second case counted as silenced by the gate is `ru-home-026` at margin
+0.001, which the margin gate stops. The veto has nothing to do with it.
+
+## Why no lexical rule separates the two
+
+`en-hard-024` and `ru-silent-029` are in the same lexical class. Both questions
+share zero content words with their hit, and neither carries a first-person
+marker. The scores sit on top of each other, 0.826 against 0.835, and so do the
+margins, 0.023 against 0.019. Only one thing separates them. A screen problem
+and a swapped display cable are the same event. A train and a river walk are
+not. The embedder scores that difference at nine thousandths.
+
+So the signal is semantic and the gate is lexical. Any rule cheap enough to sit
+in `RecallAllowed` and strong enough to recover `en-hard-024` also re-admits
+`ru-silent-029`, which puts false recall back to 1/5.
+
+One near-miss rule was tried on paper and rejected: let the veto pass when the
+hit itself is first person. It works on these two, because the English note says
+"i swapped" and the Russian note says only "погулял". It is backwards as a
+principle. A first-person note is exactly the personal note the veto keeps away
+from a world question. The rule would weaken the veto where it was designed to
+bite. It survives here only because Russian drops the pronoun.
+
+## Decision
+
+Accept the loss. `en-hard-024` stays silenced and false recall stays 0/5.
+
+The way out is a reranker, not a longer word list. Recall@3 is 85.2% against
+recall@1 at 70.4%, so the right note is usually in the returned set and ranked
+wrong. That is where the remaining points are, and it is not this task.

docs/qa.md

@@ -498,12 +498,21 @@ rejects `https://api.openai.com`, and forget really deletes
 (`internal/store/memory.go:145` is a real `DELETE`, not a tombstone). Vision is
 19/19, speaker 22/22, media 16/16.
 
-**470 got worse.** Both poisoned facts show `voided` on `/history`, and the
-defect survives. Re-measured at 15:42, after four restarts: `почему небо синее?`
-still answers `какая последняя версия языка Go?` with no `search:` line. What
-comes back is the question he typed, not the value the fact held. So the poison
-is a vector in the memory index, and `revert` does not remove it. There is
-currently no documented way to repair a poisoned box.
+**470 got worse, then closed.** Both poisoned facts showed `voided` on
+`/history` and the defect survived. Re-measured at 15:42, after four restarts:
+`почему небо синее?` still answered `какая последняя версия языка Go?` with no
+`search:` line. What came back was the question he typed, not the value the fact
+held. So the poison was a vector in the memory index, and `revert` did not
+remove it.
+
+Repaired in two parts. 470 stopped the writes: a question is never a fact, and a
+void drops the key's vectors. 493 fixed what the index holds. A fact is indexed
+as the fact and not as the utterance, and a correction drops its superseded
+vector too.
+
+A poisoned box now repairs itself on the next start. `RepairFactVectors`
+re-embeds every fact vector from the fact it names, and deletes the voided and
+superseded ones. It runs once, guarded by a marker, and logs what it did.
 
 ---
 

internal/config/config.go

@@ -1276,6 +1276,12 @@ type PhraserConfig struct {
 	NCtx       int      `json:"n_ctx,omitempty"`
 	Timeout    Duration `json:"timeout,omitempty"`
 
+	// CacheRAMMiB bounds llama-server's prompt cache. Omitted ⇒ 512 MiB, which
+	// is what keeps the resident model near 1 GB of RSS instead of the 7.9 GB
+	// measured on 2026-08-03. Set it to -1 to pass no flag at all and let the
+	// server apply its own 8 GiB default. See phraser.Config.CacheRAMMiB.
+	CacheRAMMiB int `json:"cache_ram_mib,omitempty"`
+
 	// LLMNudges — let the model word nudges again. Off by default: nudges are
 	// worded from hand-written Russian templates now (the model broke the
 	// persona and invented units). Chat, query and reminder phrasing always go

internal/memory/lexical.go

@@ -0,0 +1,134 @@
+package memory
+
+import (
+	"strings"
+	"unicode"
+)
+
+// stopwords — words that carry no topic. A question and a note that share only
+// these share nothing: "почему небо синее" and "сеть какая-то медленная" both
+// contain "какая"-shaped filler and are about different worlds.
+var stopwords = map[string]bool{
+	// interrogatives and demonstratives
+	"что": true, "чего": true, "какой": true, "какая": true, "какое": true,
+	"какие": true, "каких": true, "кто": true, "кого": true, "кому": true,
+	"почему": true, "зачем": true, "где": true, "куда": true, "откуда": true,
+	"когда": true, "сколько": true, "как": true, "то": true, "это": true,
+	"этот": true, "тот": true, "там": true, "тут": true, "такой": true,
+	// pronouns — every sentence he says is about him, so "я" is not a topic
+	"я": true, "меня": true, "мне": true, "мой": true, "моя": true, "мои": true,
+	"ты": true, "тебя": true, "тебе": true, "твой": true, "он": true, "она": true,
+	"они": true, "мы": true, "себя": true, "свой": true,
+	// prepositions, conjunctions, particles, copulas
+	"в": true, "во": true, "на": true, "с": true, "со": true, "у": true,
+	"о": true, "об": true, "про": true, "за": true, "из": true, "по": true,
+	"до": true, "от": true, "для": true, "над": true, "под": true, "при": true,
+	"и": true, "а": true, "но": true, "или": true, "же": true, "ли": true,
+	"не": true, "ни": true, "бы": true, "был": true, "была": true, "было": true,
+	"быть": true, "есть": true, "был-ли": true, "уже": true, "ещё": true,
+	"еще": true, "так": true, "вот": true, "там-же": true,
+	// English filler, for the mixed utterances he does say
+	"the": true, "a": true, "an": true, "is": true, "are": true, "was": true,
+	"were": true, "be": true, "of": true, "in": true, "on": true, "at": true,
+	"to": true, "for": true, "about": true, "and": true, "or": true, "not": true,
+	"what": true, "who": true, "why": true, "when": true, "where": true,
+	"which": true, "how": true, "i": true, "my": true, "me": true, "it": true,
+	"this": true, "that": true,
+}
+
+// firstPerson — the words that make an utterance a question about his own
+// life. Not possession only: "как я восстановил конфиги" owns nothing and is
+// still about him.
+var firstPerson = map[string]bool{
+	"я": true, "меня": true, "мне": true, "мной": true, "мой": true,
+	"моя": true, "моё": true, "мое": true, "мои": true, "моего": true,
+	"моей": true, "моих": true, "моим": true, "себя": true, "свой": true,
+	"своя": true, "свои": true, "своего": true, "мною": true,
+	"i": true, "me": true, "my": true, "mine": true, "myself": true,
+}
+
+// RecallAllowed is the second half of the recall gate (#470). A hit that
+// cleared the score and margin gate may still be about something else
+// entirely: the held-out fixture puts the right note at 0.791-0.890 and the
+// must-be-silent cases at 0.795-0.835, so no threshold sits between them, and
+// a note about his slow network answered "почему небо синее?".
+//
+// The veto applies only to a question that mentions nothing of his. That
+// restriction is what keeps the fix from costing more than it saves: recall
+// exists to find the note whose words he no longer remembers, and demanding a
+// shared word of every recall silenced four true recalls on the fixture to
+// kill one false one. A question about his own life keeps the embedder alone
+// as its judge. A question about the world has to name something the memory
+// actually mentions.
+//
+// The veto's price was re-measured on 2026-08-03 (#496,
+// docs/evals/2026-08-03-recall-topic-veto.md). It costs one true recall and
+// buys one false one, and the fixture pass count is the same either way. The
+// lost case is an English paraphrase, not the cross-language loss it was
+// reported as, and the fixture has no cross-language case at all. Do not add a
+// script test or a bilingual stem map for it — both are no-ops here. The
+// separating signal is semantic and belongs in a reranker, not in this file.
+func RecallAllowed(query, text string) bool {
+	if mentionsHim(query) {
+		return true
+	}
+	return SharesContentWord(query, text)
+}
+
+func mentionsHim(query string) bool {
+	for _, w := range strings.FieldsFunc(strings.ToLower(query), func(r rune) bool {
+		return !unicode.IsLetter(r) && !unicode.IsDigit(r)
+	}) {
+		if firstPerson[w] {
+			return true
+		}
+	}
+	return false
+}
+
+// SharesContentWord reports whether query and text have at least one topic
+// word in common, after dropping the words that carry no topic. Stems are
+// compared, so the note and the question do not have to inflect alike.
+func SharesContentWord(query, text string) bool {
+	q := contentWords(query)
+	if len(q) == 0 {
+		// Nothing to compare — a question made entirely of filler. The score
+		// gate is then the only judge it can have.
+		return true
+	}
+	t := contentWords(text)
+	for _, a := range q {
+		for _, b := range t {
+			if a == b || sameStem(a, b) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func contentWords(s string) []string {
+	var out []string
+	for _, w := range strings.FieldsFunc(strings.ToLower(s), func(r rune) bool {
+		return !unicode.IsLetter(r) && !unicode.IsDigit(r)
+	}) {
+		if !stopwords[w] {
+			out = append(out, w)
+		}
+	}
+	return out
+}
+
+// sameStem is inflection and derivation tolerance: Russian marks case and
+// tense on the ending, and the note and the question rarely use the same form.
+// "воду" and "вода" are the same water, "кормить" and "корм" the same feeding.
+// All but the last rune of the shorter word must match, and never fewer than
+// three, which is what keeps "сеть" clear of "сеанс".
+func sameStem(a, b string) bool {
+	ar, br := []rune(a), []rune(b)
+	n := min(len(ar), len(br)) - 1
+	if n < 3 {
+		return false
+	}
+	return string(ar[:n]) == string(br[:n])
+}

internal/memory/lexical_test.go

@@ -0,0 +1,56 @@
+package memory
+
+import "testing"
+
+func TestRecallAllowed(t *testing.T) {
+	cases := []struct {
+		name        string
+		query, text string
+		want        bool
+	}{
+		// The #470 shape: a world question and a note about his box.
+		{"world question, unrelated note", "почему небо синее", "сеть какая-то медленная", false},
+		{"world question, unrelated fact", "какая столица Франции", "какая последняя версия языка Go", false},
+		{"silent fixture case", "во сколько отходит поезд", "бэкап запускается в три ночи", false},
+
+		// A world question that does name the topic keeps its answer.
+		{"world question, same topic", "какой поезд идёт в Минск", "поезда в Минск ходят утром", true},
+
+		// A question about his own life is judged by the embedder alone,
+		// because recall exists for words he no longer remembers.
+		{"about him, no shared word", "во сколько я обычно засыпаю", "ложусь около одиннадцати", true},
+		{"about him, english", "which colour scheme do i like", "тёмная тема везде", true},
+
+		// Inflection must not break a match.
+		{"inflected", "чем кормить кота", "корм для кота в шкафу", true},
+	}
+	for _, c := range cases {
+		if got := RecallAllowed(c.query, c.text); got != c.want {
+			t.Errorf("%s: RecallAllowed(%q, %q) = %v, want %v", c.name, c.query, c.text, got, c.want)
+		}
+	}
+}
+
+// The known cost of the veto and the thing that pays for it, both measured on
+// the held-out fixture with the real embedder (#496,
+// docs/evals/2026-08-03-recall-topic-veto.md). The two are one lexical class:
+// zero shared content words, no first-person marker, scores 0.826 against 0.835
+// and margins 0.023 against 0.019. Recovering the first re-admits the second,
+// which puts false recall back to 1/5. Anyone loosening the veto has to move
+// the first line without moving the second.
+func TestRecallVetoTradeIsPinned(t *testing.T) {
+	if RecallAllowed("what fixed the screen problem", "the flicker went away once i swapped the display cable") {
+		t.Error("en-hard-024 is expected to stay vetoed — if this passes now, re-measure false recall before celebrating")
+	}
+	if RecallAllowed("во сколько отходит поезд", "погулял вдоль реки") {
+		t.Error("ru-silent-029 must stay vetoed — this is the false recall the veto exists to stop")
+	}
+}
+
+// A question made only of filler has no topic word to match on, and the score
+// gate is then the only judge it can have.
+func TestRecallAllowedFallsBackWhenNothingToCompare(t *testing.T) {
+	if !RecallAllowed("что это", "сеть какая-то медленная") {
+		t.Error("a question with no content word must not be vetoed")
+	}
+}

internal/memory/recalleval/recalleval.go

@@ -378,7 +378,7 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 		if len(hits) > 1 {
 			o.Margin = hits[0].Score - hits[1].Score
 		}
-		o.Recalled = bestRecall(hits, minScore, minMargin)
+		o.Recalled = bestRecall(c.Query, hits, minScore, minMargin)
 	}
 	for i, h := range hits {
 		if h.ID != c.Want {
@@ -424,11 +424,19 @@ func rankNote(inTop3 bool) string {
 // is not importable; recalleval_test.go asserts the two agree in behaviour.
 // The daemon returns the whole hit (a note and a fact are said differently);
 // the harness only scores what came back, so it keeps returning the text.
-func bestRecall(results []memory.Result, minScore, minMargin float64) string {
+// bestRecall mirrors the daemon's gate in cmd/mavend/recall.go, including the
+// topic veto added for #470: a score that clears the gate still has to be
+// about what he asked. Keep the two in step — a fixture that measures a
+// weaker gate than the daemon runs flatters it.
+func bestRecall(query string, results []memory.Result, minScore, minMargin float64) string {
 	if !memory.Confident(results, minScore, minMargin) {
 		return ""
 	}
-	return results[0].Meta["text"]
+	text := results[0].Meta["text"]
+	if !memory.RecallAllowed(query, text) {
+		return ""
+	}
+	return text
 }
 
 func bump(m map[string]TagStat, key string, pass bool) {

internal/memory/recalleval/recalleval_test.go

@@ -140,21 +140,22 @@ func words(s string) []string {
 
 // TestBestRecallMatchesDaemon — the harness duplicates bestRecall from
 // cmd/mavend/recall.go (package main is not importable). This pins the copy to
-// the original's three rules: no hits, below the gate, or no text ⇒ silence.
+// the original's rules: no hits, below the gate, no text, or no shared topic
+// word ⇒ silence.
 func TestBestRecallMatchesDaemon(t *testing.T) {
-	if got := bestRecall(nil, 0.55, 0); got != "" {
+	if got := bestRecall("чай", nil, 0.55, 0); got != "" {
 		t.Errorf("no hits: got %q, want silence", got)
 	}
 	low := []memory.Result{{ID: "a", Score: 0.4, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(low, 0.55, 0); got != "" {
+	if got := bestRecall("чай", low, 0.55, 0); got != "" {
 		t.Errorf("below gate: got %q, want silence", got)
 	}
 	noText := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{}}}
-	if got := bestRecall(noText, 0.55, 0); got != "" {
+	if got := bestRecall("чай", noText, 0.55, 0); got != "" {
 		t.Errorf("no text: got %q, want silence", got)
 	}
 	ok := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(ok, 0.55, 0); got != "чай" {
+	if got := bestRecall("чай", ok, 0.55, 0); got != "чай" {
 		t.Errorf("above gate: got %q, want %q", got, "чай")
 	}
 	// Margin: a close runner-up means the embedder cannot tell the two apart,
@@ -163,17 +164,23 @@ func TestBestRecallMatchesDaemon(t *testing.T) {
 		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
 		{ID: "b", Score: 0.85, Meta: map[string]string{"text": "кофе"}},
 	}
-	if got := bestRecall(close, 0.55, 0.03); got != "" {
+	if got := bestRecall("чай", close, 0.55, 0.03); got != "" {
 		t.Errorf("thin margin: got %q, want silence", got)
 	}
-	if got := bestRecall(close, 0.55, 0); got != "чай" {
+	if got := bestRecall("чай", close, 0.55, 0); got != "чай" {
 		t.Errorf("margin off: got %q, want %q", got, "чай")
 	}
+	// The topic veto (#470): the score is fine and the note is about
+	// something else.
+	offTopic := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{"text": "сеть какая-то медленная"}}}
+	if got := bestRecall("почему небо синее", offTopic, 0.55, 0); got != "" {
+		t.Errorf("off topic: got %q, want silence", got)
+	}
 	clear := []memory.Result{
 		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
 		{ID: "b", Score: 0.70, Meta: map[string]string{"text": "кофе"}},
 	}
-	if got := bestRecall(clear, 0.55, 0.03); got != "чай" {
+	if got := bestRecall("чай", clear, 0.55, 0.03); got != "чай" {
 		t.Errorf("wide margin: got %q, want %q", got, "чай")
 	}
 }

internal/phraser/eval/talk.go

@@ -26,20 +26,22 @@ import (
 	"time"
 
 	"github.com/kami/maven/internal/dialogue"
+	"github.com/kami/maven/internal/router"
 )
 
 //go:embed talk_v1.json
 var talkFixtureJSON []byte
 
-// The three phrasing paths under test. Values match the fixture's "path" field.
+// The phrasing paths under test. Values match the fixture's "path" field.
 const (
 	PathChat      = "chat"      // PhraseChat
 	PathQuery     = "query"     // PhraseQuery with notes
 	PathKnowledge = "knowledge" // PhraseQuery with no notes
+	PathReply     = "reply"     // PhraseReply, the reactive confirmation
 )
 
 // TalkPaths — report order.
-var TalkPaths = []string{PathChat, PathQuery, PathKnowledge}
+var TalkPaths = []string{PathChat, PathQuery, PathKnowledge, PathReply}
 
 // TalkCheckNames — the checks that apply to a free-form reply, in report order.
 // Deliberately a subset of CheckNames: length, mood and "no questions" are nudge
@@ -58,17 +60,30 @@ var TalkCheckNames = []string{
 // WantAny is the on-topic contract: at least one lowercased fragment must appear
 // in the reply. Fragments are stems ("пароль" → "парол") so declension does not
 // defeat them.
+//
+// Intent, Key and Value carry the reply path's decision: that path is phrased
+// from what the router already resolved, not from the raw utterance. Utterance
+// stays filled anyway, because it is what a human reads in the report.
 type TalkCase struct {
 	ID        string   `json:"id"`
 	Path      string   `json:"path"`
 	Utterance string   `json:"utterance"`
 	History   []string `json:"history,omitempty"`
 	Notes     []string `json:"notes,omitempty"`
+	Intent    string   `json:"intent,omitempty"`
+	Key       string   `json:"key,omitempty"`
+	Value     string   `json:"value,omitempty"`
 	WantAny   []string `json:"want_any"`
 	Tags      []string `json:"tags,omitempty"`
 	Note      string   `json:"note,omitempty"`
 }
 
+// TalkSchemaVersion — the version this loader understands. Separate from the
+// nudge fixture's SchemaVersion: the two fixtures have different shapes and
+// change on different days, and one shared constant would force a bump on the
+// fixture that did not move.
+const TalkSchemaVersion = 1
+
 // TalkFixture — the versioned envelope, same gating as Fixture.
 type TalkFixture struct {
 	SchemaVersion int        `json:"schema_version"`
@@ -83,8 +98,8 @@ func LoadTalk() (TalkFixture, error) {
 	if err := json.Unmarshal(talkFixtureJSON, &f); err != nil {
 		return TalkFixture{}, fmt.Errorf("parse talk fixture: %w", err)
 	}
-	if f.SchemaVersion != SchemaVersion {
-		return TalkFixture{}, fmt.Errorf("talk fixture schema_version %d, want %d", f.SchemaVersion, SchemaVersion)
+	if f.SchemaVersion != TalkSchemaVersion {
+		return TalkFixture{}, fmt.Errorf("talk fixture schema_version %d, want %d", f.SchemaVersion, TalkSchemaVersion)
 	}
 	if len(f.Cases) == 0 {
 		return TalkFixture{}, fmt.Errorf("talk fixture has no cases")
@@ -92,13 +107,27 @@ func LoadTalk() (TalkFixture, error) {
 	return f, nil
 }
 
-// Talker — the two methods a conversational path must have to be scorable.
-// *phraser.LLMPhraser satisfies it; same trick as Nudger.
+// Talker — the methods a conversational path must have to be scorable.
+// *phraser.LLMPhraser satisfies the first two; *phraser.Replier satisfies the
+// third, so a run that scores all four paths passes a Pair.
 type Talker interface {
 	PhraseChat(ctx context.Context, utterance string, history []dialogue.Turn) (string, error)
 	PhraseQuery(ctx context.Context, utterance string, notes []string) (string, error)
 }
 
+// Confirmer — the reply path. *phraser.Replier satisfies it.
+type Confirmer interface {
+	PhraseReply(ctx context.Context, d router.Decision) (string, error)
+}
+
+// Pair joins the two objects the daemon wires separately — the phraser and the
+// replier — so one ScoreTalk call covers every path Maven speaks through. A bare
+// Talker still works; its reply cases score as errors, which is honest.
+type Pair struct {
+	Talker
+	Confirmer
+}
+
 // TalkOutcome — one scored case.
 type TalkOutcome struct {
 	Case    TalkCase
@@ -194,10 +223,30 @@ func (c TalkCase) run(ctx context.Context, t Talker) (string, error) {
 		return t.PhraseQuery(ctx, c.Utterance, c.Notes)
 	case PathKnowledge:
 		return t.PhraseQuery(ctx, c.Utterance, nil)
+	case PathReply:
+		conf, ok := t.(Confirmer)
+		if !ok {
+			return "", fmt.Errorf("target cannot phrase replies — pass a Pair")
+		}
+		return conf.PhraseReply(ctx, c.decision())
 	}
 	return "", fmt.Errorf("unknown path %q", c.Path)
 }
 
+// decision rebuilds what the router would have handed the replier. Text is the
+// utterance for a note or a reminder, which is what the router puts there.
+func (c TalkCase) decision() router.Decision {
+	return router.Decision{
+		Intent: router.Intent(c.Intent),
+		Slots: router.Slots{
+			Key:    c.Key,
+			Value:  c.Value,
+			Text:   c.Utterance,
+			HasKey: c.Key != "",
+		},
+	}
+}
+
 func (c TalkCase) turns() []dialogue.Turn {
 	turns := make([]dialogue.Turn, 0, len(c.History))
 	for _, h := range c.History {

internal/phraser/eval/talk_test.go

@@ -11,6 +11,7 @@ import (
 	"github.com/kami/maven/internal/llm"
 	"github.com/kami/maven/internal/persona"
 	"github.com/kami/maven/internal/phraser"
+	"github.com/kami/maven/internal/router"
 )
 
 // perPathMinimum — the resolution floor. A per-path score built on a handful of
@@ -36,6 +37,10 @@ func TestTalkFixture(t *testing.T) {
 
 		switch c.Path {
 		case PathChat, PathQuery, PathKnowledge:
+		case PathReply:
+			if c.Intent == "" {
+				t.Errorf("%s: reply case has no intent — the replier is phrased from the decision", c.ID)
+			}
 		default:
 			t.Errorf("%s: unknown path %q", c.ID, c.Path)
 		}
@@ -69,10 +74,15 @@ type fakeTalker struct{ reply string }
 func (f fakeTalker) PhraseChat(context.Context, string, []dialogue.Turn) (string, error) {
 	return f.reply, nil
 }
+
 func (f fakeTalker) PhraseQuery(context.Context, string, []string) (string, error) {
 	return f.reply, nil
 }
 
+func (f fakeTalker) PhraseReply(context.Context, router.Decision) (string, error) {
+	return f.reply, nil
+}
+
 // TestScoreTalkCounts — a reply that fails on purpose must be counted on every
 // path, so a real run cannot report a hidden zero.
 func TestScoreTalkCounts(t *testing.T) {
@@ -104,7 +114,7 @@ func TestScoreTalkCounts(t *testing.T) {
 	}
 }
 
-// TestLLMTalkBaseline — the resident model on the three conversational paths.
+// TestLLMTalkBaseline — the resident model on all four phrasing paths.
 // Opt-in exactly like TestLLMPhrasingBaseline: CI has no model and a run costs
 // minutes on the CPU target.
 //
@@ -132,32 +142,32 @@ func TestLLMTalkBaseline(t *testing.T) {
 	p := phraser.NewLLMPhraserAt(base, cfg)
 	defer p.Close()
 
-	// Unreachable server is fatal here, not a logged warning, and that differs
-	// from the nudge test on purpose. PhraseNudge returns its errors, so a dead
-	// server there shows up honestly in the Errors column. PhraseChat and
-	// PhraseQuery do NOT: they swallow every failure and return a canned string
-	// ("поговорили.", "не знаю.", "вот что я нашла: …"). So on these three paths
-	// a dead server produces a full report with 0 errors and a terrible score —
-	// a number that looks like bad phrasing and is really no phrasing at all.
-	// Refusing to score without a confirmed model is the only guard available
-	// until the phraser reports its failures (Vikunja #397).
+	// The model id names the run in the report. Since Vikunja #397 every path
+	// returns its errors, so a server that dies mid-run shows up in the Errors
+	// column instead of scoring as bad phrasing — the before-and-after probe that
+	// used to stand in for that is gone.
 	model, err := llm.ModelID(ctx, base)
 	if err != nil {
-		t.Fatalf("no model at %s: %v — refusing to score, these paths hide their errors "+
-			"and would report a plausible-looking result off a dead server", base, err)
+		t.Fatalf("no model at %s: %v", base, err)
 	}
 	t.Logf("scoring model %s at %s", model, base)
 
-	rep, err := ScoreTalk(ctx, "llm ("+model+", built-in persona)", p, f)
+	// The reply path is a separate object in the daemon too: the phraser owns its
+	// own llama-server, the replier is handed an llm.Client. Pair scores both.
+	block := func() string { return persona.Facts{}.Block(time.Now()) }
+	target := Pair{Talker: p, Confirmer: phraser.NewReplier(llm.New(base, cfg.Timeout), block)}
+
+	rep, err := ScoreTalk(ctx, "llm ("+model+", built-in persona)", target, f)
 	if err != nil {
 		t.Fatalf("ScoreTalk: %v", err)
 	}
 	t.Log("\n" + rep.String() + "\nreplies:\n" + rep.Replies() + "\nfailures:\n" + rep.Failures())
 
-	// And again afterwards: the run takes minutes, and a server that died or got
-	// OOM-killed halfway through would leave the first cases scored and the rest
-	// silently canned. Checking only at the start would not catch that.
-	if _, err := llm.ModelID(ctx, base); err != nil {
-		t.Fatalf("model at %s went away during the run: %v — the score above is not trustworthy", base, err)
+	// A run where nothing was phrased is not a low score, it is no measurement.
+	if rep.Errors == rep.Total {
+		t.Fatalf("every case errored — nothing was measured, the score above is not a phrasing result")
+	}
+	if rep.Errors > 0 {
+		t.Logf("%d/%d cases errored — those are model failures, not phrasing failures", rep.Errors, rep.Total)
 	}
 }

internal/phraser/eval/talk_v1.json

@@ -222,6 +222,90 @@
       "utterance": "почему гром слышно позже молнии?",
       "want_any": ["звук", "све", "быстр", "гром", "молни"],
       "tags": ["general"]
+    },
+    {
+      "id": "reply-fact-coffee",
+      "path": "reply",
+      "intent": "fact",
+      "key": "кофе",
+      "value": "закончился",
+      "utterance": "кофе закончился",
+      "want_any": ["коф"],
+      "tags": ["fact"],
+      "note": "The plainest confirmation there is, and the sentence he hears most often."
+    },
+    {
+      "id": "reply-fact-weight",
+      "path": "reply",
+      "intent": "fact",
+      "key": "вес",
+      "value": "82",
+      "utterance": "мой вес 82",
+      "want_any": ["вес", "82"],
+      "tags": ["fact", "number"],
+      "note": "A number must survive into the confirmation; a paraphrase that drops it is useless."
+    },
+    {
+      "id": "reply-fact-pill",
+      "path": "reply",
+      "intent": "fact",
+      "key": "таблетки",
+      "value": "выпил",
+      "utterance": "таблетки выпил",
+      "want_any": ["таблетк"],
+      "tags": ["fact", "feminine"],
+      "note": "He says 'выпил', masculine and about himself. She must not copy the form onto herself."
+    },
+    {
+      "id": "reply-note-router",
+      "path": "reply",
+      "intent": "note",
+      "utterance": "роутер перезагружается сам по ночам",
+      "want_any": ["роутер"],
+      "tags": ["note"]
+    },
+    {
+      "id": "reply-note-long",
+      "path": "reply",
+      "intent": "note",
+      "utterance": "если диск снова отвалится, посмотреть кабель, а не контроллер, в прошлый раз был кабель",
+      "want_any": ["диск", "кабел"],
+      "tags": ["note", "length"],
+      "note": "A long note baits a long confirmation. One sentence is the contract."
+    },
+    {
+      "id": "reply-reminder-evening",
+      "path": "reply",
+      "intent": "reminder",
+      "utterance": "напомни вечером полить цветы",
+      "want_any": ["цвет", "полит", "вечер"],
+      "tags": ["reminder"]
+    },
+    {
+      "id": "reply-reminder-tomorrow",
+      "path": "reply",
+      "intent": "reminder",
+      "utterance": "напомни завтра позвонить в поликлинику",
+      "want_any": ["поликлиник", "позвон", "звон"],
+      "tags": ["reminder"]
+    },
+    {
+      "id": "reply-formality-bait",
+      "path": "reply",
+      "intent": "note",
+      "utterance": "запишите пожалуйста что счётчики я сдал",
+      "want_any": ["счётчик", "счетчик"],
+      "tags": ["note", "persona-bait", "address"],
+      "note": "Polite plural in the input. The confirmation must still be на ты."
+    },
+    {
+      "id": "reply-question-bait",
+      "path": "reply",
+      "intent": "note",
+      "utterance": "надо купить фильтр для воды, не помню какой",
+      "want_any": ["фильтр"],
+      "tags": ["note", "no-question"],
+      "note": "An unresolved note invites her to ask which filter. A confirmation does not ask."
     }
   ]
 }

internal/phraser/failure_test.go

@@ -0,0 +1,72 @@
+package phraser
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// A dead server must be distinguishable from bad phrasing. Both PhraseChat and
+// PhraseQuery keep the turn alive with canned text — ChatFallback, "не знаю.",
+// "вот что я нашла: …" — and every one of those is also a legitimate reply, so
+// the text alone cannot say which happened. The error is the only signal, and
+// before Vikunja #397 it was dropped: the talk scorer reported a full run with
+// zero errors off a server that answered nothing.
+func TestPhrasingReportsTheFailureWithTheFallback(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		http.Error(w, "model not loaded", http.StatusServiceUnavailable)
+	}))
+	t.Cleanup(srv.Close)
+	p := NewLLMPhraserAt(srv.URL, Config{})
+
+	cases := []struct {
+		name string
+		call func() (string, error)
+		want string
+	}{
+		{"chat", func() (string, error) {
+			return p.PhraseChat(context.Background(), "как дела", nil)
+		}, ChatFallback},
+		{"knowledge", func() (string, error) {
+			return p.PhraseQuery(context.Background(), "кто написал войну и мир", nil)
+		}, "не знаю."},
+		{"evidence", func() (string, error) {
+			return p.PhraseQuery(context.Background(), "сколько воды я выпил", []string{"два литра"})
+		}, "вот что я нашла: два литра"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got, err := c.call()
+			if err == nil {
+				t.Fatalf("no error from a dead server; the scorer would count this as bad phrasing")
+			}
+			if got != c.want {
+				t.Errorf("fallback text = %q, want %q — the daemon still has to say something", got, c.want)
+			}
+		})
+	}
+}
+
+// An empty answer is a failure too: the server is up and produced no tokens,
+// which is not an answer and must not score as one.
+func TestEmptyKnowledgeAnswerIsAnError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.Write([]byte(`{"choices":[{"message":{"content":""}}]}`))
+	}))
+	t.Cleanup(srv.Close)
+	p := NewLLMPhraserAt(srv.URL, Config{})
+
+	got, err := p.PhraseQuery(context.Background(), "кто написал войну и мир", nil)
+	if err == nil {
+		t.Fatal("an empty response scored as an answer")
+	}
+	if got != "не знаю." {
+		t.Errorf("fallback text = %q, want \"не знаю.\"", got)
+	}
+	if !strings.Contains(err.Error(), "empty") {
+		t.Errorf("error = %v; want it to name the empty response", err)
+	}
+}

internal/phraser/llmphraser.go

@@ -1,13 +1,16 @@
 package phraser
 
 import (
+	"bufio"
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"log"
 	"net/http"
+	"os"
 	"os/exec"
 	"regexp"
 	"strings"
@@ -24,6 +27,11 @@ import (
 
 var listenRE = regexp.MustCompile(`listening on (https?://\S+)`)
 
+// errEmptyResponse — the server answered and said nothing. Separate from a
+// transport failure: the model is up and produced no tokens, which is still not
+// an answer and must not score as one.
+var errEmptyResponse = errors.New("phraser: empty response from the model")
+
 type LLMPhraser struct {
 	cfg    Config
 	client *http.Client
@@ -83,6 +91,19 @@ type Config struct {
 	NCtx       int
 	Timeout    time.Duration
 
+	// CacheRAMMiB bounds llama-server's prompt cache, which is what actually ate
+	// this box. Measured on homesrv 2026-08-03: the server's own default limit is
+	// 8192 MiB, it stores the full KV state of every idle slot it evicts (112 kiB
+	// per token, so 166 MiB for one 1521-token prompt), and RSS climbed by that
+	// much per distinct prompt until it hit 7.9 GB and half a gigabyte went to
+	// swap. Weights are only 1.1 GB and mmapped, and -ngl 99 costs almost no RSS
+	// because RADV keeps device memory outside the process.
+	//
+	// 0 ⇒ the flag is not passed and the server's own 8 GiB default applies. That
+	// is the escape hatch for a llama-server too old to know --cache-ram, not a
+	// recommendation. See docs/evals/2026-08-03-llama-prompt-cache.md.
+	CacheRAMMiB int
+
 	// ContextBlock renders the shared context block (who he is, how to
 	// address him, the time) fresh for each turn. See internal/persona.
 	// nil ⇒ no block, the prompts stand alone.
@@ -116,7 +137,9 @@ func DefaultConfig(modelPath string) Config {
 		Listen:     "127.0.0.1:0",
 		NGpuLayers: -1,
 		NCtx:       2048,
-		Timeout:    30 * time.Second,
+		// 512 MiB caps total RSS near 1 GB and still holds several recent prompts.
+		CacheRAMMiB: 512,
+		Timeout:     30 * time.Second,
 	}
 }
 
@@ -225,8 +248,10 @@ func spawnLlamaServer(ctx context.Context, cfg Config) (backend, error) {
 	return p, nil
 }
 
-func startLlamaProc(ctx context.Context, cfg Config) (*llamaProc, error) {
-	p := &llamaProc{}
+// llamaArgs is the command line for one resident server. It is a function and
+// not an inline literal because kill-maven.sh's orphan sweep matches against
+// this exact line, and a test pins the two together.
+func llamaArgs(cfg Config) []string {
 	args := []string{
 		"-m", cfg.ModelPath,
 		"--host", "127.0.0.1",
@@ -235,7 +260,15 @@ func startLlamaProc(ctx context.Context, cfg Config) (*llamaProc, error) {
 		"-ngl", fmt.Sprintf("%d", cfg.NGpuLayers),
 		"--no-webui",
 	}
-	cmd := exec.CommandContext(ctx, cfg.BinPath, args...)
+	if cfg.CacheRAMMiB > 0 {
+		args = append(args, "--cache-ram", fmt.Sprintf("%d", cfg.CacheRAMMiB))
+	}
+	return args
+}
+
+func startLlamaProc(ctx context.Context, cfg Config) (*llamaProc, error) {
+	p := &llamaProc{}
+	cmd := exec.CommandContext(ctx, cfg.BinPath, llamaArgs(cfg)...)
 	// Pdeathsig: the kernel SIGKILLs llama-server the moment mavend dies — by
 	// ANY means, including SIGKILL/OOM/panic where our Close() never runs. Without
 	// it a hard-killed mavend orphans its llama-server (reparented to init, keeps
@@ -246,63 +279,104 @@ func startLlamaProc(ctx context.Context, cfg Config) (*llamaProc, error) {
 	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true, Pdeathsig: syscall.SIGKILL}
 	p.cmd = cmd
 
-	stderr, err := cmd.StderrPipe()
+	// One pipe for both streams. llama.cpp writes its buffer sizes, KV-cache
+	// layout and offload lines to stderr and its request log to stdout, and
+	// stdout used to go nowhere at all — so nothing about the model's memory was
+	// diagnosable from a running box. Both ends land in mavend's log now.
+	pr, pw, err := os.Pipe()
 	if err != nil {
-		return nil, fmt.Errorf("llm: stderr pipe: %w", err)
+		return nil, fmt.Errorf("llm: output pipe: %w", err)
 	}
+	cmd.Stdout = pw
+	cmd.Stderr = pw
 
 	if err := cmd.Start(); err != nil {
-		stderr.Close()
+		pr.Close()
+		pw.Close()
 		return nil, fmt.Errorf("llm: start: %w", err)
 	}
+	// The child holds the only other reference to the write end. Dropping ours
+	// is what makes the reader see EOF when the child dies.
+	pw.Close()
 
 	portCh := make(chan string, 1)
 	errCh := make(chan error, 1)
+	tail := &lineTail{}
 	p.wg.Add(1)
 	go func() {
 		defer p.wg.Done()
-		buf := make([]byte, 4096)
-		var leftover []byte
-		for {
-			n, err := stderr.Read(buf)
-			if n > 0 {
-				data := append(leftover, buf[:n]...)
-				lines := bytes.Split(data, []byte("\n"))
-				for _, line := range lines[:len(lines)-1] {
-					if m := listenRE.FindSubmatch(line); len(m) > 1 {
-						addr := string(m[1])
-						portCh <- addr
-						close(portCh)
-					}
+		defer pr.Close()
+		sc := bufio.NewScanner(pr)
+		// llama.cpp prints one prompt per line and a prompt can be long.
+		sc.Buffer(make([]byte, 0, 64*1024), 1024*1024)
+		listening := false
+		for sc.Scan() {
+			line := sc.Bytes()
+			log.Printf("llama: %s", line)
+			if !listening {
+				tail.add(string(line))
+				if m := listenRE.FindSubmatch(line); len(m) > 1 {
+					listening = true
+					portCh <- string(m[1])
+					close(portCh)
 				}
-				leftover = lines[len(lines)-1]
-			}
-			if err != nil {
-				errCh <- err
-				return
 			}
 		}
+		err := sc.Err()
+		if err == nil {
+			err = io.EOF
+		}
+		errCh <- err
 	}()
 
+	fail := func(err error) (*llamaProc, error) {
+		_ = cmd.Process.Kill()
+		_ = cmd.Wait()
+		return nil, err
+	}
 	select {
 	case addr := <-portCh:
 		p.base = addr
 		return p, nil
 	case err := <-errCh:
-		_ = cmd.Process.Kill()
-		_ = cmd.Wait()
-		return nil, fmt.Errorf("llm: server output: %w", err)
+		// The tail is the whole diagnosis when the server dies during load: bare
+		// "EOF" never said which layer or which allocation it choked on.
+		return fail(fmt.Errorf("llm: server output: %w; last output: %s", err, tail.String()))
 	case <-ctx.Done():
-		_ = cmd.Process.Kill()
-		_ = cmd.Wait()
-		return nil, ctx.Err()
+		return fail(ctx.Err())
 	case <-time.After(60 * time.Second):
-		_ = cmd.Process.Kill()
-		_ = cmd.Wait()
-		return nil, fmt.Errorf("llm: server did not start within 60s")
+		return fail(fmt.Errorf("llm: server did not start within 60s; last output: %s", tail.String()))
 	}
 }
 
+// lineTail keeps the last few startup lines so a server that dies before it
+// listens can say why in the error, not just "EOF". Written by the reader
+// goroutine and read by whoever gives up on startup, so it takes a lock.
+type lineTail struct {
+	mu    sync.Mutex
+	lines []string
+}
+
+const lineTailMax = 12
+
+func (t *lineTail) add(line string) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.lines = append(t.lines, line)
+	if len(t.lines) > lineTailMax {
+		t.lines = t.lines[len(t.lines)-lineTailMax:]
+	}
+}
+
+func (t *lineTail) String() string {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if len(t.lines) == 0 {
+		return "(no output)"
+	}
+	return strings.Join(t.lines, " | ")
+}
+
 // BaseURL is the llama-server this phraser talks to right now. It changes when
 // the model is swapped, so callers that cache it must register an observer
 // (OnSwap) rather than keeping the string forever.
@@ -360,8 +434,11 @@ func (p *LLMPhraser) PhraseNudge(ctx context.Context, c loop.Candidate) (deliver
 }
 
 // PhraseQuery prompts the LLM with the user's utterance and matching notes to
-// compose a natural answer. Falls back to "вот что я нашла: <notes>" on any
-// LLM error — better to give the raw data than silence.
+// compose a natural answer. On any LLM error it returns the fallback text —
+// "вот что я нашла: <notes>", or "не знаю." with no notes — and the error
+// together. The daemon uses the text and keeps the turn alive; a caller that is
+// measuring counts the failure. Until Vikunja #397 the error was dropped, so a
+// dead server scored as bad phrasing.
 func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []string) (string, error) {
 	// Blank sources are no sources. A caller that hands over one empty string —
 	// a page that fetched to nothing, a snippet trimmed away — used to take the
@@ -371,13 +448,15 @@ func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []
 	if len(notes) == 0 {
 		sys, prompt := p.knowledgePrompt(utterance)
 		resp, err := p.chatWithSystem(ctx, sys, prompt, 768)
-		if err != nil || resp == "" {
-			return "не знаю.", nil
+		if err != nil {
+			return "не знаю.", fmt.Errorf("phrase query (knowledge): %w", err)
+		}
+		if resp == "" {
+			return "не знаю.", errEmptyResponse
 		}
 		text, _, perr := parseResponseMood(resp)
 		if perr != nil {
-			log.Printf("phraser: PhraseQuery: %v", perr)
-			return "не знаю.", nil
+			return "не знаю.", fmt.Errorf("phrase query (knowledge): %w", perr)
 		}
 		if text != "" {
 			return text, nil
@@ -389,13 +468,12 @@ func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []
 	text, _, perr := parseResponseMood(resp)
 	if err != nil || perr != nil {
 		// Read the notes out rather than ship a broken fragment.
-		if perr != nil {
-			log.Printf("phraser: PhraseQuery: %v", perr)
+		cause := err
+		if cause == nil {
+			cause = perr
 		}
-		if len(notes) == 1 {
-			return "вот что я нашла: " + notes[0], nil
-		}
-		return "вот что я нашла: " + strings.Join(notes, "; "), nil
+		return "вот что я нашла: " + strings.Join(notes, "; "),
+			fmt.Errorf("phrase query (evidence): %w", cause)
 	}
 	if text != "" {
 		return text, nil
@@ -404,8 +482,9 @@ func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []
 }
 
 // PhraseChat uses the LLM to respond conversationally, building a multi-turn
-// message array from dialogue history + the current user utterance. Falls back
-// to a simple greeting on any LLM error — better to say something than nothing.
+// message array from dialogue history + the current user utterance. On any LLM
+// error it returns both ChatFallback and the error, on the same rule as
+// PhraseQuery: the fallback keeps the turn alive, the error stays visible.
 func (p *LLMPhraser) PhraseChat(ctx context.Context, utterance string, history []dialogue.Turn) (string, error) {
 	sys := chatSystemPrompt(p.cfg.ContextBlock)
 	msgs := []chatMsg{
@@ -422,13 +501,11 @@ func (p *LLMPhraser) PhraseChat(ctx context.Context, utterance string, history [
 
 	resp, err := p.chatWithMessages(ctx, msgs, 768)
 	if err != nil {
-		log.Printf("phraser: PhraseChat: %v", err)
-		return "поговорили.", nil
+		return ChatFallback, fmt.Errorf("phrase chat: %w", err)
 	}
 	text, _, perr := parseResponseMood(resp)
 	if perr != nil {
-		log.Printf("phraser: PhraseChat: %v", perr)
-		return "поговорили.", nil
+		return ChatFallback, fmt.Errorf("phrase chat: %w", perr)
 	}
 	if text != "" {
 		return text, nil

internal/phraser/phraser.go

@@ -66,11 +66,17 @@ type Stub struct{}
 // NewStub builds the floor phraser. no config — the Stub is stateless.
 func NewStub() *Stub { return &Stub{} }
 
+// ChatFallback — what she says on the chat path when the model gave her
+// nothing to say. It replaced "поговорили.", which reads as a summary of a
+// conversation that did not happen. Said out loud this one is an admission,
+// which is what it is.
+const ChatFallback = "даже не знаю, что сказать."
+
 // PhraseChat returns a stub reply — the LLMPhraser replaces this with a
 // prompted response from the model. The history parameter is accepted but
 // ignored at the stub level (the production impl uses it for multi-turn).
 func (s *Stub) PhraseChat(_ context.Context, _ string, _ []dialogue.Turn) (string, error) {
-	return "поговорили.", nil
+	return ChatFallback, nil
 }
 
 // PhraseQuery returns a deterministic summary of the best matching notes.

internal/phraser/replier.go

@@ -0,0 +1,112 @@
+// phraser/replier.go — reactive reply phrasing, the confirmation he hears
+// after every fact, note and reminder.
+//
+// It lived in cmd/mavend as package main until Vikunja #396, which meant the
+// most frequently heard sentence Maven says was the one path the phrasing eval
+// could not import, let alone score. Nothing here talks to the daemon: the
+// caller supplies the completer and the context block, and cmd/mavend keeps the
+// stub fallback so a model error still answers.
+package phraser
+
+import (
+	"context"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/persona"
+	"github.com/kami/maven/internal/router"
+)
+
+// Completer is the model seam for the replier, a subset of router.Completer.
+// *llm.Client satisfies it.
+type Completer interface {
+	Complete(ctx context.Context, r llm.Req) (string, error)
+}
+
+// replyTimeout bounds one reply. Generous because the resident model on the CPU
+// floor is slow and the caller has a deterministic fallback anyway.
+const replyTimeout = 60 * time.Second
+
+// ReplySystemPrompt — the reactive confirmation contract: one short Russian
+// sentence, feminine self-reference, informal address, no question.
+const ReplySystemPrompt = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Владелец — мужчина, говоришь с ним на "ты", в единственном числе; никогда не "вы"/"ваш" и не "он"/"его". Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), по-русски, спокойно и без официальных формулировок. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
+Пример: {"response": "Записала, что ты выпил стакан воды.", "mood": "neutral"}
+Никогда не пиши "..." в поле response.`
+
+// Replier phrases reactive confirmations with the resident model. It has no
+// fallback of its own: an error is returned, and the daemon answers from the
+// deterministic stub. That is also what makes it scorable — a dead server shows
+// up as an error rather than as bad phrasing.
+type Replier struct {
+	c Completer
+
+	// block renders the shared context block per turn (who he is, the time).
+	// nil ⇒ the prompt stands alone.
+	block func() string
+}
+
+// NewReplier builds a replier over c. block may be nil.
+func NewReplier(c Completer, block func() string) *Replier {
+	return &Replier{c: c, block: block}
+}
+
+// PhraseReply returns the confirmation for one decision. An empty string with a
+// nil error means the model produced nothing usable, which the caller must
+// treat exactly like an error.
+func (r *Replier) PhraseReply(ctx context.Context, d router.Decision) (string, error) {
+	ctx, cancel := context.WithTimeout(ctx, replyTimeout)
+	defer cancel()
+	out, err := r.c.Complete(ctx, llm.Req{
+		System:        persona.Prepend(r.block, ReplySystemPrompt),
+		User:          replyContext(d),
+		Grammar:       ResponseGrammar,
+		MaxTokens:     512,
+		RepeatPenalty: 1.3,
+	})
+	if err != nil {
+		return "", err
+	}
+	out = stripThink(out)
+	if response, _, perr := parseResponseMood(out); perr != nil {
+		return "", perr
+	} else if response != "" {
+		return response, nil
+	}
+	// fallback: the model answered in bare prose, which is fine here.
+	return firstSentence(out), nil
+}
+
+// firstSentence trims the model's output to a single clean confirmation: first
+// line, first sentence, whitespace-normalized — the last-line defense against a
+// small model that rambles past the first period despite the prompt + stop.
+func firstSentence(s string) string {
+	s = strings.TrimSpace(s)
+	if i := strings.IndexByte(s, '\n'); i >= 0 {
+		s = s[:i]
+	}
+	// keep up to and including the first sentence-ending punctuation.
+	if i := strings.IndexAny(s, ".!?"); i >= 0 {
+		s = s[:i+1]
+	}
+	return strings.TrimSpace(s)
+}
+
+// replyContext renders the decision into a compact RU description for the model.
+func replyContext(d router.Decision) string {
+	switch d.Intent {
+	case router.IntentFact:
+		return "записала факт: " + d.Slots.Key + " " + d.Slots.Value
+	case router.IntentNote:
+		return "сохранила заметку: " + d.Slots.Text
+	case router.IntentReminder:
+		return "поставила напоминание: " + d.Slots.Text
+	default:
+		return string(d.Intent) + ": " + d.Slots.Text
+	}
+}
+
+// StripThink removes the <think> block a Thinking-variant model emits before its
+// answer. Exported for the daemon's own model callers, which parse output that
+// never passes through a phraser method.
+func StripThink(s string) string { return stripThink(s) }

internal/phraser/replier_test.go

@@ -0,0 +1,90 @@
+package phraser
+
+import (
+	"context"
+	"testing"
+
+	"github.com/kami/maven/internal/llm"
+	"github.com/kami/maven/internal/router"
+)
+
+type mockCompleter struct {
+	out string
+	err error
+}
+
+func (m mockCompleter) Complete(_ context.Context, _ llm.Req) (string, error) { return m.out, m.err }
+
+func TestReplierReturnsLLMReply(t *testing.T) {
+	r := NewReplier(mockCompleter{out: `{"response":"записала, кофе закончился","mood":"neutral"}`}, nil)
+	got, err := r.PhraseReply(context.Background(), noteDecision())
+	if err != nil || got != "записала, кофе закончился" {
+		t.Errorf("got %q, %v, want %q, nil", got, err, "записала, кофе закончился")
+	}
+}
+
+func TestReplierFallsBackToPlainText(t *testing.T) {
+	r := NewReplier(mockCompleter{out: "записала, кофе закончился"}, nil)
+	got, err := r.PhraseReply(context.Background(), noteDecision())
+	if err != nil || got != "записала, кофе закончился" {
+		t.Errorf("got %q, %v, want %q, nil", got, err, "записала, кофе закончился")
+	}
+}
+
+func TestReplierReportsTheModelError(t *testing.T) {
+	r := NewReplier(mockCompleter{err: errTestLLMDown}, nil)
+	got, err := r.PhraseReply(context.Background(), noteDecision())
+	if err == nil {
+		t.Errorf("got %q, nil error — a dead model must be reported, not phrased around", got)
+	}
+}
+
+// A fragment the grammar left half-open is a failed generation. It must come
+// back as an error so the daemon reaches its stub, not as a reply.
+func TestReplierRejectsBrokenJSON(t *testing.T) {
+	r := NewReplier(mockCompleter{out: `{"response":"запис`}, nil)
+	got, err := r.PhraseReply(context.Background(), noteDecision())
+	if err == nil || got != "" {
+		t.Errorf("got %q, %v, want empty and an error", got, err)
+	}
+}
+
+func TestReplierEmptyOutputIsEmpty(t *testing.T) {
+	r := NewReplier(mockCompleter{out: ""}, nil)
+	got, err := r.PhraseReply(context.Background(), noteDecision())
+	if err != nil || got != "" {
+		t.Errorf("got %q, %v, want empty and no error", got, err)
+	}
+}
+
+// grammarRecorder captures the request so the grammar can be asserted on.
+type grammarRecorder struct{ req llm.Req }
+
+func (g *grammarRecorder) Complete(_ context.Context, r llm.Req) (string, error) {
+	g.req = r
+	return `{"response":"записала","mood":"neutral"}`, nil
+}
+
+func TestReplierCarriesTheResponseGrammar(t *testing.T) {
+	rec := &grammarRecorder{}
+	r := NewReplier(rec, nil)
+	if _, err := r.PhraseReply(context.Background(), noteDecision()); err != nil {
+		t.Fatalf("PhraseReply: %v", err)
+	}
+	if rec.req.Grammar != ResponseGrammar {
+		t.Errorf("grammar = %q, want ResponseGrammar", rec.req.Grammar)
+	}
+	if rec.req.System != ReplySystemPrompt {
+		t.Errorf("system prompt = %q, want ReplySystemPrompt", rec.req.System)
+	}
+}
+
+func noteDecision() router.Decision {
+	return router.Decision{Intent: router.IntentNote, Slots: router.Slots{Text: "кофе закончился"}}
+}
+
+var errTestLLMDown = errTest("llm down")
+
+type errTest string
+
+func (e errTest) Error() string { return string(e) }

internal/phraser/spawn_test.go

@@ -1,9 +1,11 @@
 package phraser
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
+	"log"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -59,6 +61,19 @@ func TestExtractPort(t *testing.T) {
 	}
 }
 
+// The prompt cache is what ate 6.8GB of the deployed server's RSS, so the cap
+// has to reach the command line, and the opt-out has to leave it off.
+func TestLlamaArgsCapsPromptCache(t *testing.T) {
+	cfg := DefaultConfig("/m.gguf")
+	if got := strings.Join(llamaArgs(cfg), " "); !strings.Contains(got, "--cache-ram 512") {
+		t.Errorf("default args = %q, want --cache-ram 512", got)
+	}
+	cfg.CacheRAMMiB = 0
+	if got := strings.Join(llamaArgs(cfg), " "); strings.Contains(got, "--cache-ram") {
+		t.Errorf("args with the cap off = %q, want no --cache-ram flag", got)
+	}
+}
+
 func TestStartLlamaProcScrapesPortAndReaps(t *testing.T) {
 	bin := fakeLlama(t, listensThenSleeps)
 	ctx, cancel := context.WithCancel(context.Background())
@@ -86,6 +101,48 @@ func TestStartLlamaProcScrapesPortAndReaps(t *testing.T) {
 	}
 }
 
+// captureLog redirects the standard logger for the duration of a test and
+// returns what was written to it.
+func captureLog(t *testing.T) *bytes.Buffer {
+	t.Helper()
+	var buf bytes.Buffer
+	old := log.Writer()
+	flags := log.Flags()
+	log.SetOutput(&buf)
+	log.SetFlags(0)
+	t.Cleanup(func() { log.SetOutput(old); log.SetFlags(flags) })
+	return &buf
+}
+
+// The child's buffer-size, KV-cache and offload lines are the only way to
+// account for its memory on a running box, and they used to be dropped: stderr
+// was scraped for the listen line and thrown away, stdout was never piped.
+func TestStartLlamaProcForwardsChildOutput(t *testing.T) {
+	buf := captureLog(t)
+	bin := fakeLlama(t, `echo "load_tensors: Vulkan0 model buffer size = 1053.34 MiB" >&2
+echo "llama_context: KV self size = 448.00 MiB"
+`+listensThenSleeps)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	p, err := startLlamaProc(ctx, testCfg(bin))
+	if err != nil {
+		t.Fatalf("startLlamaProc: %v", err)
+	}
+	p.cancel = cancel
+	defer p.Close()
+
+	got := buf.String()
+	for _, want := range []string{
+		"llama: load_tensors: Vulkan0 model buffer size = 1053.34 MiB", // stderr
+		"llama: llama_context: KV self size = 448.00 MiB",              // stdout, previously discarded
+	} {
+		if !strings.Contains(got, want) {
+			t.Errorf("log missing %q\nlog was:\n%s", want, got)
+		}
+	}
+}
+
 func TestStartLlamaProcFailureArms(t *testing.T) {
 	t.Run("binary missing", func(t *testing.T) {
 		cfg := testCfg(filepath.Join(t.TempDir(), "does-not-exist"))
@@ -96,13 +153,18 @@ func TestStartLlamaProcFailureArms(t *testing.T) {
 	})
 
 	t.Run("server exits without listening", func(t *testing.T) {
-		// stderr closes, so the reader goroutine reports EOF on errCh.
+		// stderr closes, so the reader goroutine reports EOF on errCh. The error
+		// must carry the child's last words: bare "EOF" named no cause.
+		captureLog(t)
 		bin := fakeLlama(t, `echo "ggml_vulkan: no device" >&2
 exit 1`)
 		_, err := startLlamaProc(context.Background(), testCfg(bin))
 		if err == nil || !strings.Contains(err.Error(), "llm: server output") {
 			t.Fatalf("err = %v, want the server-output arm", err)
 		}
+		if !strings.Contains(err.Error(), "ggml_vulkan: no device") {
+			t.Errorf("err = %v, want the child's last output in it", err)
+		}
 	})
 
 	t.Run("context cancelled during startup", func(t *testing.T) {
@@ -241,15 +303,7 @@ func TestKillMavenScriptMatchesRealCommandLine(t *testing.T) {
 	// startLlamaProc that breaks the sweep fails here instead of on the box.
 	cfg := DefaultConfig("/opt/maven/models/llm/Qwen3-1.7B-UD-Q4_K_XL.gguf")
 	cfg.NCtx, cfg.NGpuLayers = 4096, 99
-	cmdline := strings.Join([]string{
-		cfg.BinPath,
-		"-m", cfg.ModelPath,
-		"--host", "127.0.0.1",
-		"--port", extractPort(cfg.Listen),
-		"-c", fmt.Sprintf("%d", cfg.NCtx),
-		"-ngl", fmt.Sprintf("%d", cfg.NGpuLayers),
-		"--no-webui",
-	}, " ")
+	cmdline := cfg.BinPath + " " + strings.Join(llamaArgs(cfg), " ")
 	if !pat.MatchString(cmdline) {
 		t.Fatalf("kill-maven.sh pattern %q does not match %q — orphans would leak", m[1], cmdline)
 	}

internal/phraser/swap_test.go

@@ -215,10 +215,12 @@ func TestSwap_RollbackFailureLeavesNoBackendAndDegrades(t *testing.T) {
 	if _, _, aerr := p.acquire(); !errors.Is(aerr, ErrNoBackend) {
 		t.Errorf("acquire error = %v; want ErrNoBackend", aerr)
 	}
-	// Phrasing degrades to its fallback instead of failing the turn.
+	// Phrasing degrades to its fallback instead of failing the turn, and since
+	// Vikunja #397 it reports the error next to that fallback so a measuring
+	// caller can tell "no model" from "bad phrasing".
 	got, err := p.PhraseChat(context.Background(), "привет", nil)
-	if err != nil {
-		t.Fatalf("PhraseChat after a total failure returned an error: %v", err)
+	if !errors.Is(err, ErrNoBackend) {
+		t.Errorf("PhraseChat error = %v; want ErrNoBackend alongside the fallback", err)
 	}
 	if got == "" {
 		t.Error("PhraseChat returned empty; the fallback must still say something")

internal/router/question.go

@@ -0,0 +1,64 @@
+package router
+
+import "strings"
+
+// interrogatives — the question words that mark an utterance as asking rather
+// than telling. Tokenized, never substring: "что" inside "чтобы" and "как"
+// inside "какао" are not questions.
+var interrogatives = []string{
+	"что", "чего", "какой", "какая", "какое", "какие", "каких",
+	"кто", "кого", "кому", "чей", "почему", "зачем", "отчего",
+	"где", "куда", "откуда", "когда", "сколько", "как",
+	"what", "who", "whom", "why", "when", "where", "which", "how",
+}
+
+// narrativeRequests — "tell me about X" asks for knowledge Maven does not
+// hold about him. It carries no question mark and no interrogative, which is
+// how "расскажи про битву при Ватерлоо" reached the fact store (#470).
+var narrativeRequests = []string{
+	"расскажи", "объясни", "опиши", "перечисли",
+	"tell", "explain", "describe",
+}
+
+// captureVerbs — an explicit instruction to record something. These win over
+// every test below, because "запиши что я пил воду" contains an interrogative
+// and is still a capture: the word he said is "запиши".
+var captureVerbs = []string{
+	"запиши", "запомни", "отметь", "заметь", "добавь", "сохрани",
+	"note", "remember", "log", "save",
+}
+
+// IsQuestionShaped reports whether text asks for something rather than
+// records it. It is a deterministic offline test over tokens, so it costs
+// nothing and never depends on the model that produced the routing decision.
+//
+// It exists because a mis-routed question used to be persisted as a fact
+// about the owner, with the model's invented answer as the value (#470). The
+// predicate is deliberately blunt: refusing to store a question is cheap and
+// reversible, storing an invented fact about him is neither.
+func IsQuestionShaped(text string) bool {
+	t := strings.TrimSpace(text)
+	if t == "" {
+		return false
+	}
+	toks := planTokens(t)
+	for _, v := range captureVerbs {
+		if hasTok(toks, v) {
+			return false
+		}
+	}
+	if strings.HasSuffix(t, "?") {
+		return true
+	}
+	for _, w := range interrogatives {
+		if hasTok(toks, w) {
+			return true
+		}
+	}
+	for _, w := range narrativeRequests {
+		if hasTok(toks, w) {
+			return true
+		}
+	}
+	return false
+}

internal/router/question_test.go

@@ -0,0 +1,48 @@
+package router
+
+import "testing"
+
+func TestIsQuestionShaped(t *testing.T) {
+	// The seven utterances #470 recorded, plus the captures that must keep
+	// working. A capture misread as a question loses a fact; a question
+	// misread as a capture poisons recall, so the captures are the ones worth
+	// pinning here.
+	cases := []struct {
+		text string
+		want bool
+	}{
+		{"какая последняя версия языка Go?", true},
+		{"что дальше?", true},
+		{"расскажи про битву при Ватерлоо", true},
+		{"почему небо синее?", true},
+		{"какая столица Австралии?", true},
+		{"кто такой Никола Тесла?", true},
+		{"сколько стоит доллар", true},
+		{"who is the premier of Japan", true},
+		{"объясни линии Фраунгофера", true},
+
+		{"запиши что я пил воду", false},
+		{"запомни какая у меня машина", false},
+		{"отметь что я поужинал", false},
+		{"поужинал", false},
+		{"я выпил кофе", false},
+		{"вода", false},
+		{"привет", false},
+		{"", false},
+	}
+	for _, c := range cases {
+		if got := IsQuestionShaped(c.text); got != c.want {
+			t.Errorf("IsQuestionShaped(%q) = %v, want %v", c.text, got, c.want)
+		}
+	}
+}
+
+// Substring matching is what made the day-plan predicates wrong before, and
+// this predicate gates a write, so it gets the same guard.
+func TestIsQuestionShapedIsTokenized(t *testing.T) {
+	for _, text := range []string{"чтобы не забыть, я полил кактус", "какао выпил"} {
+		if IsQuestionShaped(text) {
+			t.Errorf("IsQuestionShaped(%q) = true; a question word inside a longer word is not a question", text)
+		}
+	}
+}

internal/store/facts.go

@@ -6,6 +6,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log"
 	"strings"
 	"time"
 
@@ -46,6 +47,39 @@ func (s *Store) WriteFact(ctx context.Context, ts time.Time, kind FactKind, key,
 	return id, nil
 }
 
+// FactRecallText is the text a fact is indexed under and read back as (#493).
+//
+// It used to be the utterance that wrote the fact, so recall of ANY
+// voice-tapped fact answered with the sentence he said instead of the value
+// stored: `go_version = 1.20` was indexed as "какая последняя версия языка
+// Go?", and that question is what came back. The poisoned rows made the defect
+// visible; the shape was wrong for legitimate facts too.
+//
+// The key is spoken with its underscores dropped, because a key is written for
+// the store and this string is read out loud.
+func FactRecallText(key, value string) string {
+	spoken := strings.TrimSpace(strings.ReplaceAll(key, "_", " "))
+	v := strings.TrimSpace(DecodeFactValue(value))
+	switch {
+	case v == "":
+		return spoken
+	case spoken == "":
+		return v
+	}
+	return spoken + " — " + v
+}
+
+// DecodeFactValue unwraps a stored value for reading. The column holds raw json
+// when the writer serialized one (SetValue, CorrectValue) and a plain string
+// when it did not (a voice tap), so a reader that wants the text handles both.
+func DecodeFactValue(value string) string {
+	var s string
+	if err := json.Unmarshal([]byte(value), &s); err == nil {
+		return s
+	}
+	return value
+}
+
 // LatestFact returns the latest non-voided fact for key, or ErrNoFact.
 // "Non-voided" = no later row has voids_id pointing at it. We resolve this by
 // taking the newest row whose id is not referenced by any voids_id.
@@ -290,6 +324,20 @@ func (s *Store) CorrectValue(ctx context.Context, key, source string, value any,
 	if err != nil {
 		return 0, fmt.Errorf("last insert id: %w", err)
 	}
+	// The same repair a void needs, for the same reason (#493). A correction
+	// supersedes the value, and the vector still holds the old one, so recall
+	// kept answering with the value he had just corrected. Dropping it costs
+	// the key its recall vector until the fact is tapped again: this layer has
+	// no embedder, and a missing vector loses a question while a stale one
+	// answers it wrongly.
+	//
+	// Best-effort: the corrected row is committed, and a correction that lands
+	// beats one that fails on cleanup.
+	if n, derr := s.VectorMemory().DeletePrefix(ctx, "fact:"+key+":"); derr != nil {
+		log.Printf("store: correct %q: memory vectors survive: %v", key, derr)
+	} else if n > 0 {
+		log.Printf("store: correct %q: dropped %d superseded memory vector(s)", key, n)
+	}
 	return newID, nil
 }
 
@@ -332,6 +380,20 @@ func (s *Store) VoidLatestFact(ctx context.Context, key, source string, ts time.
 	if err != nil {
 		return 0, 0, fmt.Errorf("void: last insert id: %w", err)
 	}
+	// The other half of the repair (#470). A fact reaches recall through a
+	// vector keyed `fact:<key>:<unix>`, holding the utterance that wrote it.
+	// Voiding the row alone left that vector answering questions, so revert
+	// reported success on a box that stayed broken. Deleting every vector for
+	// the key covers the earlier rows too: their values are superseded, and a
+	// superseded value has no business claiming a turn.
+	//
+	// Best-effort by design: the audit trail is already committed, and a fact
+	// that is voided but still recallable is better than a void that failed.
+	if n, derr := s.VectorMemory().DeletePrefix(ctx, "fact:"+key+":"); derr != nil {
+		log.Printf("store: void %q: memory vectors survive: %v", key, derr)
+	} else if n > 0 {
+		log.Printf("store: void %q: dropped %d memory vector(s)", key, n)
+	}
 	return oldID, newID, nil
 }
 

internal/store/factvectors.go

@@ -0,0 +1,177 @@
+package store
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// metaKeyFactVectorShape names the shape the stored fact vectors were written
+// in. It exists so the repair below runs once per box instead of on every
+// start: the rows it fixes were written by a code path that no longer exists,
+// and once fixed nothing writes that shape again.
+const metaKeyFactVectorShape = "fact_vector_shape"
+
+// factVectorShapeFact is the shape FactRecallText produces. Anything else in
+// the marker (including nothing, which is every box written before #493) means
+// the fact vectors still hold utterances.
+const factVectorShapeFact = "fact-text (#493)"
+
+// FactVectorRepair is what one repair run did, for logging.
+type FactVectorRepair struct {
+	Skipped   bool // marker already matched — nothing to do
+	Rewritten int  // rows re-embedded from the fact they name
+	Dropped   int  // rows deleted: voided, superseded, or naming no fact at all
+	Kept      int  // rows already holding the right text
+	Took      time.Duration
+}
+
+// RepairFactVectors brings the fact rows of memory_vectors in line with the
+// facts they name, and is the operator recovery a poisoned box had no path to
+// (#470 point 4, #493).
+//
+// Three defects put wrong text in that index, and all three are write-path
+// fixes that do nothing for rows already stored:
+//
+//   - the indexed text was the utterance, so every fact row reads back a
+//     sentence rather than a value;
+//   - a void left its vector behind, so retracted junk kept answering;
+//   - a correction left its vector behind, so the superseded value did.
+//
+// So each fact row is resolved against the fact store and one of three things
+// happens. It is dropped when the key has no fact, when the newest row for the
+// key is a void marker, or when a newer vector for the same key exists — a
+// superseded value has no business claiming a turn. It is re-embedded when its
+// text is not what FactRecallText says the fact is. Otherwise it is left alone.
+//
+// Idempotent, and safe to interrupt: every step compares before writing and the
+// marker is written last, so a run that dies partway is simply redone.
+func (s *Store) RepairFactVectors(ctx context.Context, embed EmbedFunc) (FactVectorRepair, error) {
+	start := time.Now()
+	var res FactVectorRepair
+
+	shape, err := s.Meta(ctx, metaKeyFactVectorShape)
+	if err != nil {
+		return res, err
+	}
+	if shape == factVectorShapeFact {
+		res.Skipped = true
+		res.Took = time.Since(start)
+		return res, nil
+	}
+
+	rows, err := s.db.QueryContext(ctx, `SELECT id, meta FROM memory_vectors`)
+	if err != nil {
+		return res, fmt.Errorf("repair fact vectors: read: %w", err)
+	}
+	type factVec struct {
+		id, key string
+		meta    map[string]string
+		ts      int64
+	}
+	var vecs []factVec
+	newest := map[string]int64{} // key → newest ts seen for it
+	for rows.Next() {
+		var id, metaJSON string
+		if err := rows.Scan(&id, &metaJSON); err != nil {
+			rows.Close()
+			return res, fmt.Errorf("repair fact vectors: row: %w", err)
+		}
+		meta := map[string]string{}
+		if err := json.Unmarshal([]byte(metaJSON), &meta); err != nil {
+			rows.Close()
+			return res, fmt.Errorf("repair fact vectors: meta for %q: %w", id, err)
+		}
+		if meta["type"] != "fact" {
+			continue
+		}
+		key, ts, ok := splitFactVectorID(id)
+		if !ok {
+			continue
+		}
+		vecs = append(vecs, factVec{id: id, key: key, meta: meta, ts: ts})
+		if ts > newest[key] {
+			newest[key] = ts
+		}
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return res, fmt.Errorf("repair fact vectors: rows: %w", err)
+	}
+
+	for _, v := range vecs {
+		drop := v.ts < newest[v.key]
+		var want string
+		if !drop {
+			f, ferr := s.LatestFact(ctx, v.key)
+			switch {
+			case errors.Is(ferr, ErrNoFact):
+				drop = true
+			case ferr != nil:
+				return res, fmt.Errorf("repair fact vectors: fact %q: %w", v.key, ferr)
+			case DecodeFactValue(f.Value) == "voided":
+				drop = true
+			default:
+				want = FactRecallText(v.key, f.Value)
+			}
+		}
+		if drop {
+			if err := s.VectorMemory().Delete(ctx, v.id); err != nil {
+				return res, err
+			}
+			res.Dropped++
+			continue
+		}
+		if v.meta["text"] == want {
+			res.Kept++
+			continue
+		}
+		vec, err := embed(ctx, want)
+		if err != nil {
+			return res, fmt.Errorf("repair fact vectors: embed %q: %w", v.id, err)
+		}
+		// The whole meta blob is rewritten in Go rather than patched in SQL,
+		// because json_set needs the JSON1 extension and this store is opened
+		// through sqlcipher.
+		v.meta["text"] = want
+		metaJSON, err := json.Marshal(v.meta)
+		if err != nil {
+			return res, fmt.Errorf("repair fact vectors: meta %q: %w", v.id, err)
+		}
+		if _, err := s.db.ExecContext(ctx,
+			`UPDATE memory_vectors SET vec = ?, meta = ? WHERE id = ?`,
+			encodeVec(vec), string(metaJSON), v.id); err != nil {
+			return res, fmt.Errorf("repair fact vectors: write %q: %w", v.id, err)
+		}
+		res.Rewritten++
+	}
+
+	if err := s.SetMeta(ctx, metaKeyFactVectorShape, factVectorShapeFact); err != nil {
+		return res, err
+	}
+	res.Took = time.Since(start)
+	return res, nil
+}
+
+// splitFactVectorID reads the key and write time back out of a fact vector's
+// id, which the write path builds as `fact:<key>:<unix>`. A key may hold a
+// colon, the timestamp may not, so the split is from the right.
+func splitFactVectorID(id string) (key string, ts int64, ok bool) {
+	rest, found := strings.CutPrefix(id, "fact:")
+	if !found {
+		return "", 0, false
+	}
+	cut := strings.LastIndex(rest, ":")
+	if cut <= 0 {
+		return "", 0, false
+	}
+	ts, err := strconv.ParseInt(rest[cut+1:], 10, 64)
+	if err != nil {
+		return "", 0, false
+	}
+	return rest[:cut], ts, true
+}

internal/store/factvectors_test.go

@@ -0,0 +1,151 @@
+package store
+
+import (
+	"context"
+	"database/sql"
+	"testing"
+	"time"
+)
+
+// The write-path half of #493: recall of a fact must read back the fact, not
+// the sentence he happened to say.
+func TestFactRecallText(t *testing.T) {
+	for _, tc := range []struct {
+		name, key, value, want string
+	}{
+		{"json value", "go_version", `"1.20"`, "go version — 1.20"},
+		{"plain value", "water", "выпил", "water — выпил"},
+		{"no value", "shower", "", "shower"},
+		{"underscores are spoken as spaces", "espresso_machine", `"чистая"`, "espresso machine — чистая"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := FactRecallText(tc.key, tc.value); got != tc.want {
+				t.Fatalf("FactRecallText(%q, %q) = %q; want %q", tc.key, tc.value, got, tc.want)
+			}
+		})
+	}
+}
+
+// A correction left the superseded value in the index, so recall answered with
+// the value he had just corrected (#493).
+func TestCorrectValueDropsMemoryVectors(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	now := time.Now()
+	mem := s.VectorMemory()
+
+	if _, err := s.WriteFact(ctx, now, KindSelf, "go_version", `"1.20"`, "tap:voice", 1.0, sql.NullInt64{}); err != nil {
+		t.Fatalf("WriteFact: %v", err)
+	}
+	if err := mem.Insert(ctx, "fact:go_version:1", []float32{1, 0, 0}, map[string]string{
+		"type": "fact", "text": "go version — 1.20",
+	}); err != nil {
+		t.Fatalf("Insert: %v", err)
+	}
+	if _, err := s.CorrectValue(ctx, "go_version", "feedback", "1.25", now.Add(time.Minute)); err != nil {
+		t.Fatalf("CorrectValue: %v", err)
+	}
+	got, err := mem.ByPrefix(ctx, "fact:")
+	if err != nil {
+		t.Fatalf("ByPrefix: %v", err)
+	}
+	if len(got) != 0 {
+		t.Fatalf("after the correction the index still holds %+v; the superseded value must not answer", got)
+	}
+}
+
+// The recovery path a poisoned box had none of (#470 point 4, #493): rows
+// written before the fix hold utterances, voided junk and superseded values,
+// and no write-path change reaches any of them.
+func TestRepairFactVectors(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	now := time.Now()
+	mem := s.VectorMemory()
+	embed := func(ctx context.Context, text string) ([]float32, error) {
+		return []float32{float32(len(text)), 1, 0}, nil
+	}
+
+	// A live fact indexed under the question that wrote it — the defect.
+	if _, err := s.WriteFact(ctx, now, KindSelf, "water", `"выпил"`, "tap:voice", 1.0, sql.NullInt64{}); err != nil {
+		t.Fatalf("WriteFact water: %v", err)
+	}
+	if err := mem.Insert(ctx, "fact:water:100", []float32{9, 9, 9}, map[string]string{
+		"type": "fact", "source": "voice", "text": "запиши что я пил воду",
+	}); err != nil {
+		t.Fatalf("Insert water: %v", err)
+	}
+	// A voided fact whose vector survived the void.
+	if _, err := s.WriteFact(ctx, now, KindSelf, "go_version", `"1.20"`, "tap:voice", 1.0, sql.NullInt64{}); err != nil {
+		t.Fatalf("WriteFact go_version: %v", err)
+	}
+	if _, _, err := s.VoidLatestFact(ctx, "go_version", "feedback", now.Add(time.Minute)); err != nil {
+		t.Fatalf("VoidLatestFact: %v", err)
+	}
+	if err := mem.Insert(ctx, "fact:go_version:100", []float32{9, 9, 9}, map[string]string{
+		"type": "fact", "text": "какая последняя версия языка Go?",
+	}); err != nil {
+		t.Fatalf("Insert go_version: %v", err)
+	}
+	// A key with two vectors: only the newest may answer.
+	if _, err := s.WriteFact(ctx, now, KindSelf, "mood", `"устал"`, "tap:voice", 1.0, sql.NullInt64{}); err != nil {
+		t.Fatalf("WriteFact mood: %v", err)
+	}
+	for _, ts := range []string{"100", "200"} {
+		if err := mem.Insert(ctx, "fact:mood:"+ts, []float32{9, 9, 9}, map[string]string{
+			"type": "fact", "text": "мне грустно",
+		}); err != nil {
+			t.Fatalf("Insert mood %s: %v", ts, err)
+		}
+	}
+	// A note must be left entirely alone.
+	if err := mem.Insert(ctx, "note:7", []float32{5, 5, 5}, map[string]string{
+		"type": "note", "text": "сеть тормозит по вечерам",
+	}); err != nil {
+		t.Fatalf("Insert note: %v", err)
+	}
+
+	res, err := s.RepairFactVectors(ctx, embed)
+	if err != nil {
+		t.Fatalf("RepairFactVectors: %v", err)
+	}
+	if res.Rewritten != 2 || res.Dropped != 2 {
+		t.Fatalf("repair reported %+v; want 2 rewritten (water, newest mood) and 2 dropped (voided go_version, superseded mood)", res)
+	}
+
+	got, err := mem.ByPrefix(ctx, "fact:")
+	if err != nil {
+		t.Fatalf("ByPrefix: %v", err)
+	}
+	texts := map[string]string{}
+	for _, r := range got {
+		texts[r.ID] = r.Meta["text"]
+	}
+	if len(texts) != 2 {
+		t.Fatalf("the index holds %+v; want only fact:water:100 and fact:mood:200", texts)
+	}
+	if texts["fact:water:100"] != "water — выпил" {
+		t.Fatalf("water reads back %q; want the fact, not the utterance", texts["fact:water:100"])
+	}
+	if texts["fact:mood:200"] != "mood — устал" {
+		t.Fatalf("mood reads back %q", texts["fact:mood:200"])
+	}
+	// Provenance the row already carried must survive the rewrite.
+	for _, r := range got {
+		if r.ID == "fact:water:100" && r.Meta["source"] != "voice" {
+			t.Fatalf("water lost its source meta: %+v", r.Meta)
+		}
+	}
+	if notes, err := mem.ByPrefix(ctx, "note:"); err != nil || len(notes) != 1 {
+		t.Fatalf("the note row was touched: %+v (err %v)", notes, err)
+	}
+
+	// Marker written, so a second run is free and changes nothing.
+	again, err := s.RepairFactVectors(ctx, embed)
+	if err != nil {
+		t.Fatalf("second RepairFactVectors: %v", err)
+	}
+	if !again.Skipped {
+		t.Fatalf("second run did work: %+v; the marker must make it a no-op", again)
+	}
+}

internal/store/memory.go

@@ -148,6 +148,27 @@ func (m *MemoryStore) Delete(ctx context.Context, id string) error {
 	return nil
 }
 
+// DeletePrefix removes every vector whose id starts with prefix and returns
+// how many went. Same escaping as ByPrefix, so a key containing % or _ cannot
+// widen the delete.
+//
+// It exists for the repair half of a revert (#470). Voiding a fact row left
+// its vector in the index, so recall kept serving the voided fact's utterance
+// and the documented repair did not repair.
+func (m *MemoryStore) DeletePrefix(ctx context.Context, prefix string) (int64, error) {
+	pattern := escapeLike(prefix) + "%"
+	res, err := m.db.ExecContext(ctx,
+		`DELETE FROM memory_vectors WHERE id LIKE ? ESCAPE '\'`, pattern)
+	if err != nil {
+		return 0, fmt.Errorf("memory: delete prefix %q: %w", prefix, err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return 0, fmt.Errorf("memory: delete prefix %q: rows affected: %w", prefix, err)
+	}
+	return n, nil
+}
+
 // escapeLike neutralises the LIKE wildcards in a literal prefix.
 func escapeLike(s string) string {
 	r := strings.NewReplacer(`\`, `\\`, `%`, `\%`, `_`, `\_`)

internal/store/voidvectors_test.go

@@ -0,0 +1,64 @@
+package store
+
+import (
+	"context"
+	"database/sql"
+	"testing"
+	"time"
+)
+
+// Stage 3 of #470: reverting a fact reported success and left the vector that
+// was answering questions, so the documented repair did not repair.
+func TestVoidLatestFactDropsMemoryVectors(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	now := time.Now()
+	mem := s.VectorMemory()
+
+	if _, err := s.WriteFact(ctx, now, KindSelf, "go_version", `"1.20"`, "tap:voice", 1.0, sql.NullInt64{}); err != nil {
+		t.Fatalf("WriteFact: %v", err)
+	}
+	// The id shape actionFact writes: fact:<key>:<unix>.
+	if err := mem.Insert(ctx, "fact:go_version:1", []float32{1, 0, 0}, map[string]string{
+		"type": "fact", "text": "какая последняя версия языка Go?",
+	}); err != nil {
+		t.Fatalf("Insert: %v", err)
+	}
+	// A vector for another key must survive the void.
+	if err := mem.Insert(ctx, "fact:water:1", []float32{0, 1, 0}, map[string]string{
+		"type": "fact", "text": "запиши что я пил воду",
+	}); err != nil {
+		t.Fatalf("Insert: %v", err)
+	}
+
+	if _, _, err := s.VoidLatestFact(ctx, "go_version", "feedback", now.Add(time.Minute)); err != nil {
+		t.Fatalf("VoidLatestFact: %v", err)
+	}
+
+	got, err := mem.ByPrefix(ctx, "fact:")
+	if err != nil {
+		t.Fatalf("ByPrefix: %v", err)
+	}
+	if len(got) != 1 || got[0].ID != "fact:water:1" {
+		t.Fatalf("after the void the index holds %+v; want only fact:water:1", got)
+	}
+}
+
+func TestDeletePrefixDoesNotWidenOnWildcards(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	mem := s.VectorMemory()
+
+	for _, id := range []string{"fact:a_b:1", "fact:axb:1"} {
+		if err := mem.Insert(ctx, id, []float32{1, 0}, map[string]string{"type": "fact"}); err != nil {
+			t.Fatalf("Insert %q: %v", id, err)
+		}
+	}
+	n, err := mem.DeletePrefix(ctx, "fact:a_b:")
+	if err != nil {
+		t.Fatalf("DeletePrefix: %v", err)
+	}
+	if n != 1 {
+		t.Fatalf("deleted %d rows; the _ in the key must not match x", n)
+	}
+}