memory: record the recall topic veto's real price (V-496)

#496 asked to skip the veto when the question and the hit are in
different scripts, so an English question stops losing a Russian note.
Measured first: the fixture has no cross-language case, and en-hard-024
is an English question against an English note. Both proposed fixes are
no-ops.

What the veto actually does on the fixture, with the real embedder: it
costs en-hard-024 and buys ru-silent-029. Pass count is 22/32 either
way; false recall is 0/5 with it and 1/5 without. The two cases are one
lexical class, so no rule cheap enough for RecallAllowed separates them.

Accepts the loss and pins both sides in a test.

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

cmd/mavend/actions_query.go

@@ -715,7 +715,7 @@ 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)
@@ -740,7 +740,9 @@ 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
@@ -753,6 +755,23 @@ 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/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/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.

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.

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

@@ -60,6 +60,14 @@ var firstPerson = map[string]bool{
 // 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

internal/memory/lexical_test.go

@@ -31,6 +31,22 @@ func TestRecallAllowed(t *testing.T) {
 	}
 }
 
+// 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) {

internal/phraser/llmphraser.go

@@ -1,6 +1,7 @@
 package phraser
 
 import (
+	"bufio"
 	"bytes"
 	"context"
 	"encoding/json"
@@ -8,6 +9,7 @@ import (
 	"io"
 	"log"
 	"net/http"
+	"os"
 	"os/exec"
 	"regexp"
 	"strings"
@@ -83,6 +85,19 @@ 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 +131,9 @@ 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 +242,10 @@ 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 +254,15 @@ 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 +273,104 @@ 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.

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)
 	}