mavgpud: turn gemma's thinking off in the chat template (V-489)

Owner's call, 02-08-2026. Without it the 12B spends the reply budget on
reasoning tokens and answers empty at low max_tokens. Verified on the box:
"Столица Франции?" now answers "Париж" with no reasoning_content.

.claude/prose-dictionary.yaml

@@ -0,0 +1,160 @@
+# Maven project dictionary for the direct-prose skill.
+#
+# These terms override every word preference in the skill's word-choice tables.
+# Each entry exists because the name drifted in real docs or real answers, not
+# because the word looked improvable.
+#
+# Format and the rule for adding a term: ~/.claude/skills/direct-prose/references/modes.md
+
+terms:
+  resident_model:
+    name: resident model
+    meaning: the one always-warm Qwen3-1.7B llama-server that both routes and phrases
+    avoid:
+      - the model
+      - the LLM
+      - the 1.7B
+      - the phraser model
+    examples:
+      good: The resident model emits GBNF-constrained JSON.
+      bad: The 1.7B emits GBNF-constrained JSON.
+
+  router:
+    name: router
+    meaning: the stage that turns an utterance into a Decision with one of 7 intents
+    avoid:
+      - orchestrator
+      - intent classifier
+      - dispatcher
+
+  classifier:
+    name: classifier
+    meaning: the embedder nearest-neighbour path that runs when the router is off or errors
+    avoid:
+      - the fallback
+      - the floor
+      - the old router
+    examples:
+      good: A router error falls through to the classifier.
+      bad: A router error falls through to the floor.
+
+  cascade:
+    name: cascade
+    meaning: the ordered path stage 0, then router, then classifier
+    avoid:
+      - the pipeline
+      - the chain
+      - the fallback chain
+
+  stage_0:
+    name: stage 0
+    meaning: the deterministic rules that answer before the resident model is called
+    avoid:
+      - the fast path
+      - bypass
+      - deterministic assist
+      - preemption
+    examples:
+      good: Stage 0 routes agenda questions to IntentQuery.
+      bad: The bypass routes agenda questions to IntentQuery.
+
+  query_source:
+    name: query source
+    meaning: one entry in querySources, which either claims a turn or passes
+    avoid:
+      - arm
+      - handler
+      - branch
+      - answerer
+    examples:
+      good: Kiwix is the last query source before the model answers from memory.
+      bad: Kiwix is the last arm before the model answers from memory.
+
+  personal_boundary:
+    name: personal boundary
+    meaning: the query source that stops a question about him from reaching the world
+    avoid:
+      - the boundary
+      - the privacy gate
+      - the personal filter
+
+  clarify:
+    name: clarify
+    meaning: the turn outcome where Maven asks instead of acting
+    avoid:
+      - refusal
+      - rejection
+      - punt
+    examples:
+      good: The gate produced two false clarifies.
+      bad: The gate produced two false refusals.
+
+  fact:
+    name: fact
+    meaning: a keyed, supersedable row in the fact store
+    avoid:
+      - memory entry
+      - datum
+      - record
+
+  note:
+    name: note
+    meaning: free text he captured, indexed for recall
+    avoid:
+      - memo
+      - entry
+
+  memory:
+    name: memory
+    meaning: the embedded index over notes and facts that backs recall
+    avoid:
+      - RAG store
+      - vector db
+      - long-term memory
+
+  nudge:
+    name: nudge
+    meaning: one proactive message the digestion worker proposes and the dispatcher sends
+    avoid:
+      - suggestion
+      - proposal
+      - proactive prompt
+      - reminder
+    examples:
+      good: A fact can close the nudge that asked for it.
+      bad: A fact can close the suggestion that asked for it.
+
+  digestion_worker:
+    name: digestion worker
+    meaning: the background engine that consolidates memory and proposes nudges
+    avoid:
+      - digestion tick
+      - background engine
+      - reflection loop
+
+  reach:
+    name: reach
+    meaning: an outbound channel Maven speaks through, such as telegram, ntfy or voice
+    avoid:
+      - sink
+      - delivery channel
+      - notification backend
+
+  ecosystem:
+    name: ecosystem
+    meaning: Nexus, Praxis and Hexis together
+    avoid:
+      - the services
+      - the integrations
+      - upstream
+
+  act:
+    name: act
+    meaning: the intent that runs a capability through Hexis
+    avoid:
+      - action
+      - command
+      - execution
+    examples:
+      good: An act with no allowlisted fn is gated to a clarify.
+      bad: An action with no allowlisted fn is gated to a clarify.

.claude/settings.json

@@ -0,0 +1,20 @@
+{
+  "hooks": {
+    "SessionStart": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "f=.claude/prose-dictionary.yaml; [ -f \"$f\" ] && jq -Rs '{hookSpecificOutput:{hookEventName:\"SessionStart\",additionalContext:(\"Project prose dictionary. These terms override every word preference in the direct-prose output style. Use the name, never the avoid list.\\n\\n\"+.)}}' \"$f\" 2>/dev/null || true",
+            "statusMessage": "Loading prose dictionary"
+          },
+          {
+            "type": "command",
+            "command": "f=HANDOFF.md; [ -f \"$f\" ] && jq -Rs '{hookSpecificOutput:{hookEventName:\"SessionStart\",additionalContext:(\"An unconsumed HANDOFF.md is present. Run the pickup skill before anything else: read it, read the Vikunja task it names, restate the assumption set in at most five bullets, and wait for the user to confirm before writing code. It is a claim from the previous session, not truth. Delete it once consumed.\\n\\n\"+.)}}' \"$f\" 2>/dev/null || true",
+            "statusMessage": "Loading handoff"
+          }
+        ]
+      }
+    ]
+  }
+}

.claude/skills/pickup/SKILL.md

@@ -0,0 +1,78 @@
+---
+name: pickup
+description: Start a work session on a Maven task. Runs task start, reads the brief and the disposable handoff, restates the assumption set, and waits for correction before touching code. Use at the start of any session that continues earlier work, when the user says "pickup", "continue", "resume", or names a Vikunja task id.
+---
+
+# Pickup
+
+The point of this skill is the pause in step 5. Every wasted session in this repo
+started with an agent that inferred the goal instead of stating it back.
+
+## 0. Get on the branch
+
+```sh
+task start <vikunja-id>
+```
+
+`~/.local/bin/task` owns the branch, the identity and the PR. It cuts
+`task/<id>-<slug>` off `origin/master` and sets the commit author to the `claude`
+gitea user. It writes `TASK.md` from the Vikunja task, and pulls any waiting
+review comments into `.task/review-comments.md`. Do not hand-roll any of that.
+
+`TASK.md` is the brief and it is immutable. If it says a PR already exists, this
+is a review-fix session and not new work. Read the comments first.
+
+## 1. Read the handoff
+
+`HANDOFF.md` at the repo root, if it exists. It is gitignored, it belongs to one
+session, and it holds only what is needed to resume. Treat it as a claim from the
+previous agent, not as truth. It can be stale or wrong.
+
+If there is no handoff, that is normal. It means the last session closed clean.
+
+## 2. Read the durable state
+
+In this order, and stop as soon as you have enough:
+
+- The Vikunja task, by id. Project Maven is ID 2, MCP at `http://localhost:9100/mcp`.
+  The task description and its comments hold the goal, the constraints, and the
+  assumption ledger. This outranks the handoff on every conflict.
+- `CLAUDE.md`, the section that covers the area you are about to touch.
+- The one file under `docs/` that owns the area. Check its `Last verified` line.
+  If the sha is behind the code you are reading, say so in step 4 and trust the code.
+
+Do not read the dated files under `docs/evals/`. They are measurements from one day,
+never updated. Read one only when you need the number it recorded.
+
+If no task id is known, ask for one before doing anything else. Work without a task
+is work nobody can resume.
+
+## 3. Look at the ground
+
+`git status`, `git log --oneline -5`, and the diff on the current branch. What the
+repo says beats what any document says.
+
+## 4. Restate, then stop
+
+Write at most five bullets and stop. Do not write code, do not open files to "check
+one thing first", do not start with a small safe change.
+
+```
+Task:        V-359, one line.
+Done:        what is already on the branch.
+Next:        the one thing this session does.
+Constraints: what would make this wrong.
+Assuming:    the beliefs that, if false, waste the session.
+```
+
+Then ask: is this right? Wait for the answer.
+
+A corrected assumption goes into the Vikunja task as a comment, not into the handoff.
+The handoff dies tonight. The task does not.
+
+## 5. Then begin
+
+- Delete `HANDOFF.md`. It has been consumed and must not outlive this step.
+- On master, cut the branch: `scripts/task-branch.sh <id> <slug>`.
+- One task per session. When context passes roughly half, run `/wrap` rather than
+  pushing on. A compacted session is a session that forgot why it made a choice.

.claude/skills/wrap/SKILL.md

@@ -0,0 +1,94 @@
+---
+name: wrap
+description: Close a Maven work session cleanly. Runs the tests, updates the durable docs, commits in reviewable slices with the Vikunja ref, pushes so the PR opens, records state in Vikunja, and leaves a disposable handoff only if work remains. Use when the user says "wrap", "wrap up", "done for now", or when context passes roughly half.
+---
+
+# Wrap
+
+Run every step. A partial wrap is worse than none, because the next session trusts
+the parts that did run.
+
+## 1. Prove it works
+
+`make test`. If something fails, fix it or say plainly in the handoff and in Vikunja
+that it fails, with the output. Never wrap on an untested claim.
+
+## 2. Update the durable docs
+
+Ask what a future agent would have to learn the hard way, and write that down.
+
+- `CLAUDE.md` when a fact an agent needs before touching code has changed: routing
+  behaviour, a measured number, a flag default, a constraint. A commit that changed
+  routing or phrasing without touching the matching CLAUDE.md section is a bug.
+  Correct stale text in place. Do not append a new paragraph next to the wrong one.
+- `AGENTS.md` when the recipe to build, run or preview changed.
+- The one file under `docs/` that owns the area, plus its `Last verified: <date> @ <sha>`
+  line. Only a doc directly under `docs/` carries that line.
+- A new dated file under `docs/evals/` when you measured something. Never edit an
+  existing dated file. A newer measurement is a new file, and the living doc points
+  at it.
+
+Nothing that must survive tonight goes anywhere else. Not into the handoff, not into
+a commit message, not into a comment in the code.
+
+## 3. Commit in slices
+
+Under 300 changed lines per commit in non-markdown files, enforced by `.githooks/pre-commit`.
+Markdown is exempt and may land as one batch.
+
+Each commit is one idea, subject in the repo's voice, lowercase area prefix, and it
+ends with the Vikunja ref:
+
+```
+router: narrow the single-token rule (V-359)
+```
+
+If a change genuinely cannot split under 300 lines, say why in the commit body before
+reaching for `--no-verify`.
+
+## 4. Land it
+
+```sh
+task pr
+```
+
+It refuses a dirty tree, pushes, opens or refreshes the PR against the repo default
+branch, labels the Vikunja task in-review, comments the PR url on it, and pushes an
+ntfy. Do not push by hand and do not call `tea` yourself.
+
+## 5. Record what `task pr` cannot know
+
+Comment on the Vikunja task: what you measured, what is still open. List every
+assumption that turned out to be wrong. If the session found new work, create a task
+for it now rather than describing it in prose.
+
+This step is what makes the handoff disposable.
+
+## 6. Leave the handoff, or leave none
+
+If the task is finished, delete `HANDOFF.md` and stop. An empty root is the correct
+end state.
+
+If work remains, write `HANDOFF.md` with nothing but what the next agent needs to
+resume, and no history:
+
+```markdown
+# Handoff — <date>
+
+Task: V-359 <one line>
+Branch: task/359-<slug>, cut from master
+
+## Where I stopped
+<two sentences, mid-thought detail that is nowhere else>
+
+## Next action
+<the single concrete next step>
+
+## Do not
+<the trap I nearly fell into, or the approach already ruled out>
+```
+
+Nothing else goes in it. No summary of what landed, that is in git and Vikunja. No
+design rationale, that is in `docs/`. No fact an agent needs on any task, that is in
+`CLAUDE.md`. If a line in the handoff would still matter next week, it is in the wrong
+file.

.githooks/commit-msg

@@ -0,0 +1,30 @@
+#!/bin/sh
+# Every commit names the Vikunja task it belongs to.
+#
+#   router: narrow the single-token rule (V-359)
+#
+# V- and not #, because Gitea autolinks #359 to a Gitea issue, which is a
+# different tracker and a wrong link.
+#
+# Exempt: merges, reverts, fixup/squash, and the initial commit.
+
+msg_file=$1
+subject=$(sed -n '1p' "$msg_file")
+
+case "$subject" in
+	Merge\ *|Revert\ *|fixup!\ *|squash!\ *|amend!\ *) exit 0 ;;
+esac
+
+if [ -f "$(git rev-parse --git-dir)/MERGE_HEAD" ]; then
+	exit 0
+fi
+
+if printf '%s' "$subject" | grep -qE '\(V-[0-9]+\)$'; then
+	exit 0
+fi
+
+echo "commit-msg: subject must end with a Vikunja task ref." >&2
+echo "  got:  $subject" >&2
+echo "  want: router: narrow the single-token rule (V-359)" >&2
+echo "  No task yet? Create one. Work without a task is work nobody can resume." >&2
+exit 1

.githooks/pre-commit

@@ -0,0 +1,30 @@
+#!/bin/sh
+# Two guards, both bypassable with --no-verify when you mean it.
+#   1. master is not a working branch.
+#   2. a code commit stays under 300 changed lines.
+# Markdown is exempt from the size cap on purpose: docs land as one batch.
+
+branch=$(git symbolic-ref --short HEAD 2>/dev/null)
+
+case "$branch" in
+	master|main)
+		echo "pre-commit: refusing to commit on $branch." >&2
+		echo "  task start <vikunja-id>   # branch off origin/master, write TASK.md" >&2
+		exit 1
+		;;
+esac
+
+# Added + deleted lines across staged files that are not markdown.
+# numstat prints "-\t-\t<path>" for binaries; those count 0 and that is fine,
+# a binary blob is not the kind of diff this cap exists to stop.
+loc=$(git diff --cached --numstat -- . ':(exclude)*.md' |
+	awk '$1 ~ /^[0-9]+$/ { a += $1 } $2 ~ /^[0-9]+$/ { d += $2 } END { print a + d + 0 }')
+
+if [ "$loc" -gt 300 ]; then
+	echo "pre-commit: $loc changed lines in non-markdown files, cap is 300." >&2
+	echo "  Split it. Each commit should be one reviewable idea." >&2
+	echo "  git reset <path> to unstage, or --no-verify if this genuinely cannot split." >&2
+	exit 1
+fi
+
+exit 0

.gitignore

@@ -9,6 +9,7 @@
 /mavwaked
 /mavmaild
 /mavupdate
+/mavgpud
 
 # Certs (private keys, don't commit)
 certs/
@@ -40,6 +41,9 @@ deploy/telegram.env
 deploy/zenmoney.token
 # IMAP password, read by mavmaild (never in argv, never committed)
 deploy/imap.password
+# Compose interpolation secrets — MAVEN_AMBIENT_TOKEN today. docker compose
+# reads this file itself; it is not an env_file on any service.
+/.env
 
 # Temp files
 /tmp/
@@ -50,7 +54,19 @@ opencode.json
 # Test coverage output
 coverage.out
 
-# Agent worktrees and local agent state
-.claude/
+# Agent worktrees and local agent state. The workflow itself is tracked: the
+# hooks, the skills and the prose dictionary are how a session behaves, so they
+# get reviewed like code. Everything else under .claude/ is scratch.
+/.claude/*
+!/.claude/settings.json
+!/.claude/prose-dictionary.yaml
+!/.claude/skills/
+
+# The disposable handoff. One session, then deleted. Never committed:
+# anything worth keeping belongs in Vikunja, CLAUDE.md or docs/.
+/HANDOFF.md
 /models/stt
 /models/tts
+
+# root .env — MAVEN_AMBIENT_TOKEN and friends, same class as deploy/telegram.env
+.env

CLAUDE.md

@@ -28,6 +28,14 @@ model is a one-line change to `phraser.model_path` in `deploy/mavend.json`.
 See `docs/rearchitecture.md` for the target architecture, `docs/design.md` for the folded design spec, and
 `AGENTS.md` for local-preview + model-download recipes.
 
+**Model work is moving to the workstation** (owner's call, 2026-08-02). homesrv cannot grow a
+GPU and the workstation has 16GB of VRAM. So the resident model, STT and TTS become preferred
+remotes with a floor on homesrv. The workstation is never assumed up. Fall back silently when
+it would only do the job better. Name the gap when the 1.7B cannot do it at all. The embedder
+stays on homesrv permanently, because it backs that floor. Read `docs/offload.md` before
+touching a daemon seam or adding a model caller. Vikunja #483 is the umbrella, #484 to #487
+are the work.
+
 ## Build & test
 
 CGO daemons (`mavend`, `mavsttd`, `mavttsd`, `mavenclient`) need the vendored toolchain
@@ -127,11 +135,13 @@ on in deploy** — this section used to say it was wired `nil`, which stopped be
 Cascade order: `stage0.go` exact-match fast-path → LLM router (when non-nil) → classifier
 fallback. Any LLM error falls through to the classifier so a turn never breaks on the model.
 
-Measured on the 77-case RU fixture (`docs/evals/2026-07-31-model-bakeoff.md`): the classifier scores
-36.8% full accuracy at p50 31ms; Qwen3-1.7B scores 67.5% intent-only / 72.7% through the
-cascade at p50 ≈825ms. Accuracy roughly doubled, latency is ~27× worse, and that trade was
-accepted deliberately. **The ≈2.7s figure that stood here until 2026-08-02 was contention,
-not the model.** See `docs/evals/2026-07-31-routing.md` line 61, which measures the LLM router at
+Measured on the 77-case RU fixture. **Re-measured 2026-08-02: the classifier scores 68.8%
+full accuracy at p50 16.6µs**, not the 36.8% at p50 31ms that stood here from
+`docs/evals/2026-07-31-model-bakeoff.md`. That older figure predates the stage 0 rules and the
+seed additions, both of which now score inside the classifier baseline. Qwen3-1.7B scores
+77.9% intent-only / 72.7% through the cascade. So the router buys about 4 points of accuracy,
+not a doubling, and the trade is worth re-arguing rather than assuming. **The ≈2.7s figure
+that stood here until 2026-08-02 was contention, not the model.** See `docs/evals/2026-07-31-routing.md` line 61, which measures the LLM router at
 p50 825ms / p95 1.2s / max 3.0s and the full cascade at p50 0.80-1.04s. Do not plan latency
 work off the bakeoff table. `Confidence: 1.0` used to be hardcoded in `llmrouter.go`, so the LLM
 path could never ask for clarification (6/6 refusal cases missed on the fixture) — Vikunja
@@ -213,3 +223,60 @@ data pans on a phone. Local preview + headless screenshot recipe is in `AGENTS.m
 
 This repo is project **Maven** (ID 2) in Vikunja. MCP: `http://localhost:9100/mcp` (or
 `http://192.168.1.104:9100/mcp` from workpc). Feature/bug/deploy tasks go there.
+
+Vikunja is the durable task store. A task holds the goal, the constraints and the
+assumption ledger. Work without a task id is work nobody can resume, so a session that
+has no id asks for one before it starts.
+
+## Session workflow
+
+`~/.local/bin/task` owns the branch, the commit identity and the PR. One task, one
+session, one PR.
+
+```sh
+task start <vikunja-id>   # branch off origin/master, write TASK.md, fetch review comments
+task pr                   # push, open or refresh the PR, label Vikunja, notify
+task comments             # re-pull this branch's review comments into .task/
+```
+
+Around that, `/pickup` opens a session and `/wrap` closes it. Wrap at roughly half
+context rather than letting the session compact.
+
+Five stores, and each one owns something the others must not hold:
+
+| Store | Holds | Lifetime |
+|---|---|---|
+| Vikunja task | goal, constraints, assumption ledger, status | durable |
+| `CLAUDE.md`, `AGENTS.md` | what an agent must know before touching code | durable |
+| `docs/` | design, measurements, decisions | durable |
+| `TASK.md` | the brief for this branch, written by `task start`, immutable | one branch |
+| `HANDOFF.md` | only what the next agent needs to resume | one session |
+
+`TASK.md` and `.task/` are excluded through `.git/info/exclude`. `HANDOFF.md` is
+gitignored and injected at session start. If a line in the handoff would still matter
+next week, it is in the wrong file.
+
+Docs are tiered by path, so staleness is visible from the filename. Files directly under
+`docs/` are living and carry a `Last verified: <date> @ <sha>` line. Files under
+`docs/evals/` are dated measurements and are never edited after the day, so a newer
+number is a new file. Files under `docs/archive/` are dead and read by nobody by default.
+
+## Git guards
+
+Two hooks in `.githooks/`, tracked, wired with `core.hooksPath`. Fresh clone:
+
+```sh
+git config core.hooksPath .githooks
+```
+
+- `pre-commit` refuses master, and refuses more than 300 changed lines in non-markdown
+  files. Markdown is exempt and may land as one batch.
+- `commit-msg` requires the subject to end with `(V-<id>)`. `V-` and not `#`, because
+  Gitea autolinks `#123` to a Gitea issue, which is the wrong tracker.
+
+Two more guards live outside the repo, in `~/.claude/hooks/`. `diff-budget.sh` blocks
+further edits past 600 changed lines on a `task/` branch. `prose_lint_hook.py` checks
+prose on every write. Both measure against `origin/master`, so a local master that is
+ahead of the remote makes the diff budget read high.
+
+`--no-verify` exists. Using it means saying why in the commit body.

Makefile

@@ -16,11 +16,11 @@ PIPER_BIN     := $(shell pwd)/deps/piper/piper
 PIPER_MODEL   := $(shell pwd)/models/tts/ru_RU-irina-medium.onnx
 PIPER_ESPEAK  := $(shell pwd)/deps/piper/espeak-ng-data
 
-.PHONY: simulate stt-fixtures test-stt-golden all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall eval-phrasing eval-models
+.PHONY: simulate stt-fixtures test-stt-golden all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall eval-phrasing eval-models build-gpud
 
 all: build
 
-build: build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav build-mail build-update
+build: build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav build-mail build-update build-gpud
 
 build-stt:
 	CGO_CFLAGS="$(CGO_CFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
@@ -59,6 +59,12 @@ build-mail:
 build-update:
 	$(GO) build $(GOFLAGS) -o mavupdate ./cmd/mavupdate/
 
+# mavgpud runs on the workstation, not here. It is built with the rest so a
+# broken supervisor is caught by `make build` on homesrv rather than by the
+# workstation refusing to serve. Copy the binary over, do not `make deploy` it.
+build-gpud:
+	$(GO) build $(GOFLAGS) -o mavgpud ./cmd/mavgpud/
+
 run-web: build-web
 	./mavweb -addr :9200 -voice 127.0.0.1:9100
 

cmd/mavend/confirm.go

@@ -107,14 +107,18 @@ func (h *reactiveHandler) confirmResolvers(ctx context.Context) []confirmResolve
 				return pr != nil && !h.now().After(pr.expiry)
 			},
 			yes: func() string {
-				// Only record the acceptance. The tick loop reads accepted
-				// routines and nudges on their own interval. Building a
-				// reminder here made a routine fire exactly once (Vikunja #366).
-				if err := h.dataStore.AcceptProposedRoutine(ctx, pr.routineID, h.now()); err != nil {
-					log.Printf("voice: accept proposed routine: %v", err)
-					return "не получилось запомнить рутину."
-				}
-				return "буду напоминать."
+				// Voice does NOT accept (Vikunja #367). Accepting hands the
+				// tick loop a standing new reason to speak, which is the same
+				// tier as enabling a tool — and DESIGN.md § "surface caps
+				// authority" says a room mic, reachable by anyone present, is
+				// structurally incapable of layer 3. So a spoken "да" leaves
+				// the row 'proposed' and points at the authed page, where the
+				// accept button is gated at step-up. The convenience of
+				// answering out loud stays; the authority does not move.
+				//
+				// Acceptance itself is recorded by /routines, and the tick
+				// loop nudges on the interval from there (Vikunja #366).
+				return "поняла — подтверди на странице рутин, и начну напоминать."
 			},
 			no: func() string {
 				if err := h.dataStore.DismissProposedRoutine(ctx, pr.routineID); err != nil {

cmd/mavend/followup.go

@@ -12,13 +12,21 @@ import (
 // which waits on voice-print attribution (see PROGRESS multi-user deferral).
 const voiceDialogueID = "voice"
 
-// toDialogueSlots projects the router's slots onto the dialogue layer's subset
-// (everything except the fact Value, which the dialogue layer doesn't carry).
+// toDialogueSlots and applyDialogueSlots are the only bridge between
+// router.Slots and dialogue.Slots. dialogue must not import router (import
+// cycle), so the two structs are hand-kept copies and every field has to be
+// carried by hand here. Adding a field to either struct without adding it to
+// BOTH functions loses a slot silently — nothing fails to build. The tests in
+// slotsparity_test.go fail when the field sets or the converters stop matching;
+// when they do, fix these two functions, not the tests.
+
+// toDialogueSlots projects the router's slots onto the dialogue layer's copy.
 func toDialogueSlots(s router.Slots) dialogue.Slots {
 	return dialogue.Slots{
 		Time:    s.Time,
 		HasTime: s.HasTime,
 		Key:     s.Key,
+		Value:   s.Value,
 		HasKey:  s.HasKey,
 		Text:    s.Text,
 		Fn:      s.Fn,
@@ -27,11 +35,10 @@ func toDialogueSlots(s router.Slots) dialogue.Slots {
 	}
 }
 
-// applyDialogueSlots writes inherited dialogue slots back onto router slots,
-// preserving router-only fields (Value) the dialogue layer never touched.
+// applyDialogueSlots writes dialogue slots back onto router slots.
 func applyDialogueSlots(base router.Slots, d dialogue.Slots) router.Slots {
 	base.Time, base.HasTime = d.Time, d.HasTime
-	base.Key, base.HasKey = d.Key, d.HasKey
+	base.Key, base.Value, base.HasKey = d.Key, d.Value, d.HasKey
 	base.Text = d.Text
 	base.Fn, base.Args, base.HasFn = d.Fn, d.Args, d.HasFn
 	return base

cmd/mavend/patterns_test.go

@@ -9,6 +9,7 @@ import (
 
 	"github.com/kami/maven/internal/config"
 	"github.com/kami/maven/internal/delivery"
+	"github.com/kami/maven/internal/ipc"
 	"github.com/kami/maven/internal/loop"
 	"github.com/kami/maven/internal/pattern"
 	"github.com/kami/maven/internal/store"
@@ -283,3 +284,81 @@ func TestTickProposalCooldownSpacesAnnouncements(t *testing.T) {
 		}
 	}
 }
+
+// TestVoiceYesDoesNotAcceptRoutine — Vikunja #367. Accepting a routine hands
+// the tick loop a standing new reason to speak, which DESIGN.md puts at layer
+// 3, and voice is structurally incapable of layer 3. A spoken "да" must park
+// the decision for the authed page, not flip the row itself.
+func TestVoiceYesDoesNotAcceptRoutine(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedRefillEvents(t, st, ctx, now, pattern.MinEvents-1)
+
+	h := &reactiveHandler{api: ipc.NewStoreAPI(st), dataStore: st, now: func() time.Time { return now }}
+
+	// The MinEvents'th event is the one that makes the pattern detectable, and
+	// it goes through the voice path so the proposal is parked for a y/n.
+	last := now.Add(time.Duration(pattern.MinEvents-1) * 7 * 24 * time.Hour)
+	factID, err := st.WriteFact(ctx, last, store.KindSelf, "cat_water", "refill", "voice", 1.0, sql.NullInt64{})
+	if err != nil {
+		t.Fatalf("write fact: %v", err)
+	}
+	if phrase := h.detectPattern(ctx, factID, "cat_water", "refill", last); phrase == "" {
+		t.Fatal("expected a parked routine proposal")
+	}
+
+	reply, handled := h.resolveConfirm(ctx, "да")
+	if !handled {
+		t.Fatal("the spoken yes should be consumed by the routine confirm")
+	}
+	if !strings.Contains(reply, "рутин") {
+		t.Fatalf("reply should send him to the routines page, got %q", reply)
+	}
+
+	rows, err := st.ListProposedRoutinesByStatus(ctx, store.RoutineAccepted)
+	if err != nil {
+		t.Fatalf("list accepted: %v", err)
+	}
+	if len(rows) != 0 {
+		t.Fatalf("voice accepted a routine: %+v", rows)
+	}
+	proposed, err := st.ListProposedRoutinesByStatus(ctx, store.RoutineProposed)
+	if err != nil {
+		t.Fatalf("list proposed: %v", err)
+	}
+	if len(proposed) != 1 {
+		t.Fatalf("proposed routines = %d, want 1 (still waiting for the page)", len(proposed))
+	}
+}
+
+// TestVoiceNoStillDismissesRoutine — declining does not move the boundary
+// outward, so voice keeps it. Only acceptance is gated.
+func TestVoiceNoStillDismissesRoutine(t *testing.T) {
+	st := newTestStore(t)
+	ctx := context.Background()
+	now := refNow()
+	seedRefillEvents(t, st, ctx, now, pattern.MinEvents-1)
+
+	h := &reactiveHandler{api: ipc.NewStoreAPI(st), dataStore: st, now: func() time.Time { return now }}
+
+	last := now.Add(time.Duration(pattern.MinEvents-1) * 7 * 24 * time.Hour)
+	factID, err := st.WriteFact(ctx, last, store.KindSelf, "cat_water", "refill", "voice", 1.0, sql.NullInt64{})
+	if err != nil {
+		t.Fatalf("write fact: %v", err)
+	}
+	if phrase := h.detectPattern(ctx, factID, "cat_water", "refill", last); phrase == "" {
+		t.Fatal("expected a parked routine proposal")
+	}
+
+	if _, handled := h.resolveConfirm(ctx, "нет"); !handled {
+		t.Fatal("the spoken no should be consumed by the routine confirm")
+	}
+	rows, err := st.ListProposedRoutinesByStatus(ctx, store.RoutineDismissed)
+	if err != nil {
+		t.Fatalf("list dismissed: %v", err)
+	}
+	if len(rows) != 1 {
+		t.Fatalf("dismissed routines = %d, want 1", len(rows))
+	}
+}

cmd/mavend/slotsparity_test.go

@@ -0,0 +1,71 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/dialogue"
+	"github.com/kami/maven/internal/router"
+)
+
+// TestSlotsParity — dialogue.Slots is a hand-kept copy of router.Slots
+// (dialogue must not import router: import cycle). Drift is silent, so this
+// test compares the two field sets by name and type. If it fails, add the new
+// field to both structs AND to toDialogueSlots/applyDialogueSlots in
+// followup.go — do not relax the test.
+func TestSlotsParity(t *testing.T) {
+	fields := func(v any) map[string]string {
+		rt := reflect.TypeOf(v)
+		out := make(map[string]string, rt.NumField())
+		for i := 0; i < rt.NumField(); i++ {
+			f := rt.Field(i)
+			out[f.Name] = f.Type.String()
+		}
+		return out
+	}
+	rf, df := fields(router.Slots{}), fields(dialogue.Slots{})
+	for name, typ := range rf {
+		dt, ok := df[name]
+		if !ok {
+			t.Errorf("router.Slots.%s (%s) missing from dialogue.Slots", name, typ)
+			continue
+		}
+		if dt != typ {
+			t.Errorf("field %s: router has %s, dialogue has %s", name, typ, dt)
+		}
+	}
+	for name, typ := range df {
+		if _, ok := rf[name]; !ok {
+			t.Errorf("dialogue.Slots.%s (%s) missing from router.Slots", name, typ)
+		}
+	}
+}
+
+// TestSlotsRoundTrip — the converters carry every field. A field the parity
+// test accepts can still be dropped in transit, so round-trip a fully
+// populated value and compare.
+func TestSlotsRoundTrip(t *testing.T) {
+	full := router.Slots{
+		Time:    time.Date(2026, 8, 2, 11, 0, 0, 0, time.UTC),
+		HasTime: true,
+		Fn:      "restart",
+		Args:    []string{"nginx"},
+		HasFn:   true,
+		Key:     "water",
+		Value:   `"drank"`,
+		HasKey:  true,
+		Text:    "выпил воды",
+	}
+	// Every field must be non-zero, or the round-trip proves nothing.
+	rv := reflect.ValueOf(full)
+	for i := 0; i < rv.NumField(); i++ {
+		if rv.Field(i).IsZero() {
+			t.Fatalf("field %s is zero: extend this fixture so the round-trip covers it",
+				rv.Type().Field(i).Name)
+		}
+	}
+	if got := applyDialogueSlots(router.Slots{}, toDialogueSlots(full)); !reflect.DeepEqual(got, full) {
+		t.Errorf("round-trip lost a slot:\n got %+v\nwant %+v", got, full)
+	}
+}

cmd/mavgpud/gpu.go

@@ -0,0 +1,117 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+)
+
+// The card is an AMD 7900 GRE with 16GB, driven by amdgpu and ROCm. Everything
+// here reads sysfs and forks nothing: rocm-smi is not even installed on the
+// workstation, and a poll that costs a subprocess every second is a poll that
+// gets tuned down until it is useless.
+
+// gpuProc — one process holding the compute engine.
+type gpuProc struct {
+	PID  int
+	Comm string
+	VRAM int64 // bytes, as the kernel accounts them to this process
+}
+
+// probe reads the two sysfs trees the supervisor decides from.
+//
+// kfdRoot is /sys/class/kfd/kfd/proc, one directory per ROCm process. The
+// directory appears when the process initialises HIP, which is well before it
+// allocates anything large. That is the whole reason this works: the job that
+// is about to want the card announces itself while it is still starting up,
+// so we see the contender rather than only the winner of an allocation race.
+//
+// drmDev is /sys/class/drm/cardN/device, which reports total and used VRAM for
+// the card as a whole.
+type probe struct {
+	kfdRoot string
+	drmDev  string
+}
+
+// foreign lists every ROCm process that is not ours. selfPID is the supervisor's
+// llama-server child, or 0 when it is not running.
+//
+// An unreadable kfd tree returns no processes and no error. That is deliberate
+// and it is the safe direction only because startVRAM also has to agree before
+// anything launches: a supervisor that cannot see the KFD never sees free VRAM
+// either, because the CPT run holding the card shows up in the drm totals.
+func (p probe) foreign(selfPID int) []gpuProc {
+	entries, err := os.ReadDir(p.kfdRoot)
+	if err != nil {
+		return nil
+	}
+	var out []gpuProc
+	for _, e := range entries {
+		pid, err := strconv.Atoi(e.Name())
+		if err != nil || pid == selfPID {
+			continue
+		}
+		out = append(out, gpuProc{
+			PID:  pid,
+			Comm: readComm(pid),
+			VRAM: p.procVRAM(e.Name()),
+		})
+	}
+	return out
+}
+
+// procVRAM sums the per-node vram_* files under one process directory. The
+// suffix is the KFD topology node id (vram_35881 on this card), so it is
+// globbed rather than named, and a machine with two cards sums both.
+func (p probe) procVRAM(pid string) int64 {
+	matches, err := filepath.Glob(filepath.Join(p.kfdRoot, pid, "vram_*"))
+	if err != nil {
+		return 0
+	}
+	var total int64
+	for _, m := range matches {
+		total += readInt(m)
+	}
+	return total
+}
+
+// freeVRAM reports the bytes the card has left. Used only to decide whether to
+// start: a shortfall here means llama-server would refuse to load anyway. It is
+// never used to decide to stop, because by the time free VRAM has dropped the
+// other job has already failed its allocation, which is exactly the outcome
+// yielding exists to prevent.
+func (p probe) freeVRAM() int64 {
+	total := readInt(filepath.Join(p.drmDev, "mem_info_vram_total"))
+	used := readInt(filepath.Join(p.drmDev, "mem_info_vram_used"))
+	if total <= 0 {
+		return 0
+	}
+	if free := total - used; free > 0 {
+		return free
+	}
+	return 0
+}
+
+func readInt(path string) int64 {
+	b, err := os.ReadFile(path)
+	if err != nil {
+		return 0
+	}
+	n, err := strconv.ParseInt(strings.TrimSpace(string(b)), 10, 64)
+	if err != nil {
+		return 0
+	}
+	return n
+}
+
+// readComm names the contender for the log. The log is the instrument for the
+// open question in Vikunja #488: whether a process can want this card without
+// ever registering on the KFD, which a Vulkan or video-decode job would.
+func readComm(pid int) string {
+	b, err := os.ReadFile(filepath.Join("/proc", strconv.Itoa(pid), "comm"))
+	if err != nil {
+		return "?"
+	}
+	return strings.TrimSpace(string(b))
+}

cmd/mavgpud/gpu_test.go

@@ -0,0 +1,103 @@
+package main
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"strconv"
+	"testing"
+)
+
+// fakeKFD builds the sysfs shape the workstation actually has: one directory
+// per ROCm process, each holding a vram_<node> file. Sampled from the live box
+// on 02-08-2026, where the CPT run appeared as proc/478104/vram_35881.
+func fakeKFD(t *testing.T, vramByPID map[int]int64) string {
+	t.Helper()
+	root := t.TempDir()
+	for pid, vram := range vramByPID {
+		dir := filepath.Join(root, strconv.Itoa(pid))
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		f := filepath.Join(dir, "vram_35881")
+		if err := os.WriteFile(f, []byte(strconv.FormatInt(vram, 10)+"\n"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return root
+}
+
+func TestForeignExcludesOurChild(t *testing.T) {
+	root := fakeKFD(t, map[int]int64{478104: 12791693312, 999: 4096})
+	p := probe{kfdRoot: root}
+
+	all := p.foreign(0)
+	if len(all) != 2 {
+		t.Fatalf("with no child running, both processes are foreign, got %d", len(all))
+	}
+
+	ours := p.foreign(999)
+	if len(ours) != 1 || ours[0].PID != 478104 {
+		t.Fatalf("our own llama-server must not count as a contender, got %+v", ours)
+	}
+	if ours[0].VRAM != 12791693312 {
+		t.Errorf("per-process VRAM = %d, want the value from vram_35881", ours[0].VRAM)
+	}
+}
+
+// An empty KFD tree is the state that permits a start, so it must read as empty
+// rather than as an error the caller has to interpret.
+func TestForeignEmptyAndMissing(t *testing.T) {
+	if got := (probe{kfdRoot: t.TempDir()}).foreign(0); len(got) != 0 {
+		t.Errorf("empty kfd tree: got %d processes, want 0", len(got))
+	}
+	if got := (probe{kfdRoot: "/nonexistent"}).foreign(0); got != nil {
+		t.Errorf("missing kfd tree: got %+v, want nil", got)
+	}
+}
+
+func TestFreeVRAM(t *testing.T) {
+	dev := t.TempDir()
+	write := func(name, v string) {
+		if err := os.WriteFile(filepath.Join(dev, name), []byte(v), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	// The live numbers from the workstation while the CPT run held the card.
+	write("mem_info_vram_total", "17163091968\n")
+	write("mem_info_vram_used", "13396389888\n")
+	p := probe{drmDev: dev}
+	if got, want := p.freeVRAM(), int64(3766702080); got != want {
+		t.Errorf("freeVRAM = %d, want %d", got, want)
+	}
+	if got := (probe{drmDev: "/nonexistent"}).freeVRAM(); got != 0 {
+		t.Errorf("unreadable card reports %d free, want 0 so nothing starts", got)
+	}
+}
+
+// With no model loaded the supervisor must still answer, and it must answer 503
+// rather than hanging or proxying into a closed port. Maven reads this endpoint
+// on a timer forever, including while the workstation is busy.
+func TestHealthAndProxyRefuseWhenNotReady(t *testing.T) {
+	s := &supervisor{run: newRunner("/bin/true", nil, "")}
+	h := s.handler(mustURL(t, "http://127.0.0.1:1"))
+
+	for _, path := range []string{"/health", "/v1/chat/completions"} {
+		w := httptest.NewRecorder()
+		h.ServeHTTP(w, httptest.NewRequest(http.MethodGet, path, nil))
+		if w.Code != http.StatusServiceUnavailable {
+			t.Errorf("%s with no model: got %d, want 503", path, w.Code)
+		}
+	}
+}
+
+func mustURL(t *testing.T, s string) *url.URL {
+	t.Helper()
+	u, err := url.Parse(s)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return u
+}

cmd/mavgpud/main.go

@@ -0,0 +1,247 @@
+// mavgpud — the workstation's GPU supervisor.
+//
+// It runs on the workstation (an AMD 7900 GRE, 16GB), not on homesrv, and it is
+// deployed separately from the Maven daemons. Maven does not participate in any
+// of this and never asks for a start: it reads /health through internal/llm.Pair
+// and either gets the big model or falls back to the resident 1.7B.
+//
+// The rule, from Vikunja #488: keep llama-server loaded whenever the card is
+// free, unload it when it has been idle too long or when another process needs
+// the card. Not on demand, because a 7-14B takes tens of seconds to load and a
+// world question would be answered by a gap every time the card had been quiet.
+// Not always on, because that holds 16GB against the owner's own jobs.
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"flag"
+	"log"
+	"net/http"
+	"net/http/httputil"
+	"net/url"
+	"os"
+	"os/signal"
+	"sync/atomic"
+	"syscall"
+	"time"
+)
+
+type config struct {
+	Listen    string `json:"listen"`     // what Maven talks to
+	LlamaAddr string `json:"llama_addr"` // where llama-server binds
+	LlamaBin  string `json:"llama_bin"`
+	// LlamaArgs must include the flags that bind LlamaAddr. They are passed
+	// through untouched so the model, context size and layer count stay the
+	// owner's business and not this daemon's schema.
+	LlamaArgs []string `json:"llama_args"`
+
+	KFDRoot   string `json:"kfd_root"`
+	DRMDevice string `json:"drm_device"`
+
+	Poll        duration `json:"poll"`
+	IdleTimeout duration `json:"idle_timeout"`
+	StopGrace   duration `json:"stop_grace"`
+	MinFreeVRAM int64    `json:"min_free_vram_bytes"`
+	// EvictAfter and StartAfter are counted in polls, not seconds. Both exist
+	// to damp flapping: a one-tick blip from a short-lived rocm process must
+	// not evict the model, and a card that has just been released must not be
+	// grabbed before the previous job has finished unmapping.
+	EvictAfter int `json:"evict_after_polls"`
+	StartAfter int `json:"start_after_polls"`
+}
+
+func defaults() config {
+	return config{
+		Listen:      ":8080",
+		LlamaAddr:   "127.0.0.1:8081",
+		KFDRoot:     "/sys/class/kfd/kfd/proc",
+		DRMDevice:   "/sys/class/drm/card1/device",
+		Poll:        duration(time.Second),
+		IdleTimeout: duration(15 * time.Minute),
+		StopGrace:   duration(20 * time.Second),
+		MinFreeVRAM: 15 << 30,
+		EvictAfter:  2,
+		StartAfter:  5,
+	}
+}
+
+// duration lets the config file say "15m" instead of counting nanoseconds.
+type duration time.Duration
+
+func (d *duration) UnmarshalJSON(b []byte) error {
+	var s string
+	if err := json.Unmarshal(b, &s); err != nil {
+		return err
+	}
+	v, err := time.ParseDuration(s)
+	if err != nil {
+		return err
+	}
+	*d = duration(v)
+	return nil
+}
+
+func main() {
+	path := flag.String("config", "/etc/mavgpud.json", "config file")
+	flag.Parse()
+
+	cfg := defaults()
+	b, err := os.ReadFile(*path)
+	if err != nil {
+		log.Fatalf("mavgpud: read config: %v", err)
+	}
+	if err := json.Unmarshal(b, &cfg); err != nil {
+		log.Fatalf("mavgpud: parse config: %v", err)
+	}
+	if cfg.LlamaBin == "" {
+		log.Fatal("mavgpud: llama_bin is required")
+	}
+
+	base := "http://" + cfg.LlamaAddr
+	run := newRunner(cfg.LlamaBin, cfg.LlamaArgs, base+"/health")
+	sup := &supervisor{
+		cfg:   cfg,
+		probe: probe{kfdRoot: cfg.KFDRoot, drmDev: cfg.DRMDevice},
+		run:   run,
+	}
+	sup.touch()
+
+	ctx, cancel := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
+	defer cancel()
+
+	target, err := url.Parse(base)
+	if err != nil {
+		log.Fatalf("mavgpud: llama_addr: %v", err)
+	}
+	srv := &http.Server{Addr: cfg.Listen, Handler: sup.handler(target)}
+	go func() {
+		log.Printf("mavgpud: listening on %s, model %s", cfg.Listen, cfg.LlamaBin)
+		if err := srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
+			log.Fatalf("mavgpud: listen: %v", err)
+		}
+	}()
+
+	sup.loop(ctx)
+
+	// The card must come back before we do. A supervisor that exits leaving
+	// llama-server holding 14GB is worse than one that never ran.
+	shut, done := context.WithTimeout(context.Background(), 5*time.Second)
+	defer done()
+	_ = srv.Shutdown(shut)
+	run.stop(time.Duration(cfg.StopGrace))
+}
+
+type supervisor struct {
+	cfg   config
+	probe probe
+	run   *runner
+
+	lastReq atomic.Int64 // unix nanos of the last request Maven sent
+
+	foreignStreak int
+	clearStreak   int
+}
+
+func (s *supervisor) touch() { s.lastReq.Store(time.Now().UnixNano()) }
+
+func (s *supervisor) idle() time.Duration {
+	return time.Since(time.Unix(0, s.lastReq.Load()))
+}
+
+// handler serves the two things the workstation exposes.
+//
+// /health is answered locally and always, with no GPU cost and no round trip,
+// because it is the only thing Maven reads and Maven reads it on a timer
+// forever. Everything else is llama-server's API, reverse-proxied. Proxying
+// rather than pointing Maven straight at llama-server is what makes the idle
+// window measurable: the supervisor cannot otherwise know when the model was
+// last used.
+func (s *supervisor) handler(target *url.URL) http.Handler {
+	proxy := httputil.NewSingleHostReverseProxy(target)
+	mux := http.NewServeMux()
+	mux.HandleFunc("/health", func(w http.ResponseWriter, r *http.Request) {
+		if !s.run.isReady() {
+			http.Error(w, "model not loaded", http.StatusServiceUnavailable)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"status":"ok"}`))
+	})
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		if !s.run.isReady() {
+			http.Error(w, "model not loaded", http.StatusServiceUnavailable)
+			return
+		}
+		s.touch()
+		proxy.ServeHTTP(w, r)
+	})
+	return mux
+}
+
+func (s *supervisor) loop(ctx context.Context) {
+	t := time.NewTicker(time.Duration(s.cfg.Poll))
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			s.tick(ctx)
+		}
+	}
+}
+
+// tick is the whole decision. Yielding is checked before starting, and presence
+// on the KFD is what triggers it — not a VRAM threshold. A ROCm process
+// registers under /sys/class/kfd/kfd/proc when it initialises HIP, before it
+// allocates, so we see a contender during its startup rather than after it has
+// already failed to get the memory it wanted.
+func (s *supervisor) tick(ctx context.Context) {
+	others := s.probe.foreign(s.run.pid())
+	if len(others) > 0 {
+		s.foreignStreak++
+		s.clearStreak = 0
+	} else {
+		s.foreignStreak = 0
+		s.clearStreak++
+	}
+
+	if s.run.running() {
+		s.run.refreshReady(ctx)
+		switch {
+		case s.foreignStreak >= s.cfg.EvictAfter:
+			log.Printf("mavgpud: yielding the card to %s", describe(others))
+			s.run.stop(time.Duration(s.cfg.StopGrace))
+		case s.idle() > time.Duration(s.cfg.IdleTimeout):
+			log.Printf("mavgpud: idle for %s, unloading", s.idle().Round(time.Second))
+			s.run.stop(time.Duration(s.cfg.StopGrace))
+		}
+		return
+	}
+
+	if s.clearStreak < s.cfg.StartAfter {
+		return
+	}
+	if free := s.probe.freeVRAM(); free < s.cfg.MinFreeVRAM {
+		return
+	}
+	s.touch() // the idle clock starts at load, not at the last request before it
+	if err := s.run.start(); err != nil {
+		log.Printf("mavgpud: start llama-server: %v", err)
+	}
+}
+
+// describe names the contenders in the log. This log is the instrument for the
+// open question in #488: whether polling the KFD misses a job that wants the
+// card without registering there.
+func describe(procs []gpuProc) string {
+	out := ""
+	for i, p := range procs {
+		if i > 0 {
+			out += ", "
+		}
+		out += p.Comm
+	}
+	return out
+}

cmd/mavgpud/runner.go

@@ -0,0 +1,132 @@
+package main
+
+import (
+	"context"
+	"log"
+	"net/http"
+	"os/exec"
+	"sync"
+	"syscall"
+	"time"
+)
+
+// runner owns one llama-server process. Owning it is the point of the daemon:
+// the workstation cannot keep a 7-14B resident, because that holds 16GB against
+// the owner's CPT runs, Correx and the manga-recap pipeline. So the thing that
+// stays up is this, which costs no VRAM, and the model comes and goes under it.
+type runner struct {
+	bin  string
+	args []string
+	// ready is llama-server's own /health, which answers "is a model loaded".
+	// Loading a 7-14B takes tens of seconds, so started is not ready.
+	readyURL string
+
+	mu    sync.Mutex
+	cmd   *exec.Cmd
+	ready bool
+	http  *http.Client
+}
+
+func newRunner(bin string, args []string, readyURL string) *runner {
+	return &runner{
+		bin: bin, args: args, readyURL: readyURL,
+		http: &http.Client{Timeout: 2 * time.Second},
+	}
+}
+
+// pid is the child's, or 0. The GPU probe needs it to tell our own model apart
+// from a contender.
+func (r *runner) pid() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.cmd == nil || r.cmd.Process == nil {
+		return 0
+	}
+	return r.cmd.Process.Pid
+}
+
+func (r *runner) running() bool { return r.pid() != 0 }
+
+// isReady reports the cached readiness. The supervisor loop refreshes it; the
+// health handler only reads, so answering /health never costs a round trip.
+func (r *runner) isReady() bool {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.ready
+}
+
+// start launches llama-server. It returns as soon as the process exists, not
+// when the model is loaded.
+func (r *runner) start() error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.cmd != nil {
+		return nil
+	}
+	cmd := exec.Command(r.bin, r.args...)
+	// Own process group, so stop kills anything llama-server spawned rather
+	// than leaving it holding VRAM after we have declared the card yielded.
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	r.cmd, r.ready = cmd, false
+	log.Printf("mavgpud: started llama-server pid=%d", cmd.Process.Pid)
+	go func() {
+		err := cmd.Wait()
+		r.mu.Lock()
+		r.cmd, r.ready = nil, false
+		r.mu.Unlock()
+		log.Printf("mavgpud: llama-server exited: %v", err)
+	}()
+	return nil
+}
+
+// stop ends llama-server and waits for the VRAM to come back. SIGTERM first so
+// it unmaps cleanly, SIGKILL after the grace window. Returning before the
+// process is gone would let the supervisor report a free card while 14GB is
+// still mapped, which is the one lie that would make yielding useless.
+func (r *runner) stop(grace time.Duration) {
+	r.mu.Lock()
+	cmd := r.cmd
+	r.ready = false
+	r.mu.Unlock()
+	if cmd == nil || cmd.Process == nil {
+		return
+	}
+	pgid := -cmd.Process.Pid
+	_ = syscall.Kill(pgid, syscall.SIGTERM)
+	deadline := time.Now().Add(grace)
+	for time.Now().Before(deadline) {
+		if !r.running() {
+			return
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+	log.Printf("mavgpud: llama-server did not exit in %s, killing", grace)
+	_ = syscall.Kill(pgid, syscall.SIGKILL)
+}
+
+// refreshReady asks llama-server whether the model is loaded. Called once per
+// supervisor tick, never per request.
+func (r *runner) refreshReady(ctx context.Context) {
+	if !r.running() {
+		return
+	}
+	ok := false
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, r.readyURL, nil)
+	if err == nil {
+		resp, err := r.http.Do(req)
+		if err == nil {
+			ok = resp.StatusCode == http.StatusOK
+			resp.Body.Close()
+		}
+	}
+	r.mu.Lock()
+	was := r.ready
+	r.ready = ok
+	r.mu.Unlock()
+	if ok && !was {
+		log.Printf("mavgpud: model ready")
+	}
+}

cmd/mavweb/main.go

@@ -1210,13 +1210,10 @@ func routineRows(rs []ipc.ProposedRoutine) []routineRow {
 	return out
 }
 
-// acceptRoutine creates the recurring reminder for a proposal, then marks the
-// proposal accepted and links the reminder to it. Weekly patterns get a cron
-// expression; any other interval fires once.
-//
-// TODO(vikunja#46): this mirrors the voice accept path in cmd/mavend/voice.go.
-// When the tick loop learns to read accepted proposals directly, both callers
-// should hand off to one place in core instead of each building a reminder.
+// acceptRoutine marks a proposal accepted. This page is the ONLY surface that
+// may do it (Vikunja #367): accepting gives the tick loop a standing new
+// reason to speak, which DESIGN.md puts at layer 3, and the button here is
+// behind step-up. Voice can park the question and dismiss, never accept.
 func acceptRoutine(ctx context.Context, core ipc.CoreAPI, id int64) error {
 	proposed, err := core.ListProposedRoutines(ctx)
 	if err != nil {

deploy/mavend.json

@@ -76,6 +76,56 @@
     "snippet_runes": 1500
   },
 
+  "//morning_routines": [
+    "The daily checklist (Vikunja #280). Each item is done when its fact_key",
+    "gets a non-voided fact inside the window, so 'выпил воды' closes water and",
+    "nothing has to be ticked by hand. nudge_at fires once, at the end of the",
+    "window, and only for what is still open. Weekdays empty = every day."
+  ],
+  "morning_routines": [
+    {
+      "name": "утро",
+      "window_start": "08:00",
+      "window_end": "11:00",
+      "nudge_at": "10:30",
+      "severity": 1,
+      "items": [
+        { "key": "medicine", "fact_key": "medicine", "label": "лекарство" },
+        { "key": "water",    "fact_key": "water",    "label": "вода" },
+        { "key": "pets",     "fact_key": "pets",     "label": "покормить кота" }
+      ]
+    }
+  ],
+
+  "//feeds": [
+    "RSS reading (Vikunja #258). Every item lands as a note with source",
+    "rss:<name>, which is also what puts entries in the intake journal that",
+    "/events reads. Only the feed URL leaves the box.",
+    "This is a starting pair, not a curated set — trim or extend it."
+  ],
+  "feeds": {
+    "poll_interval": "30m",
+    "max_items": 5,
+    "max_age": "24h",
+    "sources": [
+      { "name": "lwn",       "url": "https://lwn.net/headlines/newrss", "category": "технологии" },
+      { "name": "archlinux", "url": "https://archlinux.org/feeds/news/", "category": "технологии" }
+    ]
+  },
+
+  "//crawl": [
+    "Reading a web page (Vikunja #259). on_demand answers 'посмотри <URL>'.",
+    "No allow_hosts, so any public host he names is readable; private",
+    "addresses are refused unconditionally by internal/webfetch and do not",
+    "need listing. Setting allow_hosts here would also narrow on-demand,",
+    "which is the point of leaving it empty."
+  ],
+  "crawl": {
+    "on_demand": true,
+    "timeout": "10s",
+    "max_runes": 4000
+  },
+
   "digest": {
     "enabled": true,
     "window": "30m",
@@ -119,7 +169,7 @@
     "timeout": "400ms",
     "rate": 100,
     "max_hosts": 256,
-    "enabled": false
+    "enabled": true
   },
 
   "nexus":  { "url": "http://nexus:9740" },

deploy/mavgpud.json

@@ -0,0 +1,32 @@
+{
+  "listen": ":8080",
+  "llama_addr": "127.0.0.1:10000",
+  "llama_bin": "llama-server",
+  "llama_args": [
+    "-m", "/mnt/D/AI/gemma4/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf",
+    "-md", "/mnt/D/AI/gemma4/mtp-gemma-4-12B-it-BF16.gguf",
+    "-ngl", "99",
+    "-fa", "on",
+    "-np", "1",
+    "--host", "127.0.0.1",
+    "--port", "10000",
+    "--ctx-size", "32768",
+    "--threads", "6",
+    "--batch-size", "2048",
+    "--ubatch-size", "512",
+    "--jinja",
+    "--chat-template-kwargs", "{\"enable_thinking\":false}",
+    "--spec-type", "draft-mtp",
+    "--spec-draft-n-max", "2"
+  ],
+
+  "kfd_root": "/sys/class/kfd/kfd/proc",
+  "drm_device": "/sys/class/drm/card1/device",
+
+  "poll": "1s",
+  "idle_timeout": "15m",
+  "stop_grace": "20s",
+  "min_free_vram_bytes": 10737418240,
+  "evict_after_polls": 2,
+  "start_after_polls": 5
+}

deploy/mavgpud.service

@@ -0,0 +1,24 @@
+[Unit]
+# Runs on the workstation (bugmachine), not on homesrv. Install as a systemd
+# user unit and turn on lingering, so the card is supervised after a reboot
+# with nobody logged in:
+#
+#   scp mavgpud workpc:~/.local/bin/mavgpud
+#   scp deploy/mavgpud.json workpc:~/.config/mavgpud.json
+#   scp deploy/mavgpud.service workpc:~/.config/systemd/user/mavgpud.service
+#   ssh workpc 'systemctl --user daemon-reload && systemctl --user enable --now mavgpud'
+#   sudo loginctl enable-linger kami
+Description=Maven GPU supervisor (holds llama-server while the card is free)
+After=network.target
+
+[Service]
+ExecStart=%h/.local/bin/mavgpud -config %h/.config/mavgpud.json
+Restart=always
+RestartSec=5
+# The card must come back when the supervisor goes down. mavgpud stops
+# llama-server on SIGTERM, so give it longer than stop_grace to do that.
+KillSignal=SIGTERM
+TimeoutStopSec=60
+
+[Install]
+WantedBy=default.target

docker-compose.yml

@@ -79,7 +79,16 @@ services:
     <<: *image
     # voice.bind is 0.0.0.0:9100 in deploy/mavend.json so mavweb can reach it
     # cross-container. Verified 2026-07-06.
+    # -ambient-token turns on POST /api/ambient (Vikunja #126): the phone posts
+    # notification text, mavweb keeps only a meeting time. Empty ⇒ no route at
+    # all, which is what a missing MAVEN_AMBIENT_TOKEN gives. The value comes
+    # from the gitignored .env docker compose reads for interpolation, NOT from
+    # an env_file — flags are interpolated before any service env exists.
+    # Weakness worth naming: mavweb takes this as a flag, so it is visible in
+    # `ps` inside this container, unlike the zenmoney and IMAP secrets which are
+    # read from files.
     command: ["mavweb", "-addr", ":9201", "-voice", "mavend:9100", "-core", "/run/maven/mavend.sock",
+              "-ambient-token", "${MAVEN_AMBIENT_TOKEN:-}",
               "-nexus", "http://nexus:9740", "-praxis", "http://praxis:8989", "-hexis", "http://hexis:9741"]
     depends_on: [mavend]
     # loopback-only on purpose: /tools defines+executes arbitrary argv and

docs/offload.md

@@ -0,0 +1,142 @@
+# Offloading model work to the workstation
+
+*Last verified: 2026-08-02 @ 5c05163. Living doc: correct it in place, do not append.*
+
+Owner's call, 2026-08-02. Vikunja #483 is the umbrella. Tasks #484 to #487 are the
+work, and this file holds the shape and the rules all four must obey.
+
+## The goal
+
+homesrv cannot grow a GPU. The workstation has 16GB of VRAM. Move the model work
+to the workstation and leave homesrv running the logic that must be always-on,
+deterministic and cheap.
+
+## Why this is tractable
+
+The split already exists structurally. `mavsttd` and `mavttsd` are separate
+daemons that core reaches over a socket, not linked libraries. Moving them off-box
+is a transport change, not a redesign.
+
+The microphone is at the workstation, because that is where the owner sits and
+homesrv is headless. So speech-to-text and the wake word are already on the
+workstation side by construction. Audio never has to cross the LAN. Only the core
+turn does.
+
+## The constraint that shapes everything
+
+The workstation's GPU is often busy: CPT runs, experiments, Correx, the manga-recap
+pipeline. It also sleeps. homesrv does not.
+
+So an offloaded model is never *the* model. It is the preferred one, with a floor
+on homesrv. That is the shape the cascade already has, where a router error falls
+through to the classifier.
+
+## The degradation rule
+
+Two cases, and the line between them is sharp.
+
+**Fall back silently** when the workstation model would only do the job *better*:
+routing, phrasing, a nudge. Falling back costs nothing that exists today, because
+the resident Qwen3-1.7B is today's production quality. The owner should not be told
+that his reply was phrased by the smaller model.
+
+**Name the gap** when the resident model cannot do the job *at all*. A world
+question that a 1.7B answers by inventing is the case. A wrong answer is worse
+than "не могу сейчас". This is the rule CLAUDE.md already states for a sibling
+service being down.
+
+Nothing in between. A turn never breaks on the workstation being asleep.
+
+## Admission control, not a scheduler
+
+There is no GPU arbiter. That is a service with its own failure modes, and nothing
+here needs work *distributed*. It needs admission control. The workstation
+advertises free VRAM over a health endpoint, and Maven treats it as one more query
+source that claims a turn or passes. llama-server also refuses to load when VRAM is
+short, so the failure is detectable without cooperation from the owner's other
+jobs.
+
+The caller must be able to ask "is this peer usable right now" without a turn
+hanging on a timeout. A dead remote is a normal state, not an error state.
+`internal/llm.Pair` is that check on the Maven side. A prober caches the answer,
+so `Available()` is an atomic read and no turn pays for a health check.
+
+llama-server does not stay up on the workstation. It cannot: a resident 7-14B
+would hold 16GB against the owner's CPT runs. So a supervisor there owns its
+lifecycle, keeps it loaded while the card is free, and unloads it on idle or
+when another process needs the card (owner's call, 2026-08-02, Vikunja #488).
+
+That supervisor is still not a scheduler, and the distinction is worth holding.
+It arbitrates nothing between callers. It reports whether it can take work and
+manages one process to back that answer. Maven never asks it to start anything
+and never learns that it did.
+
+Contention is decided by presence under `/sys/class/kfd/kfd/proc`, not by a VRAM
+threshold. A ROCm process registers there when it initialises HIP, before it
+allocates anything. So the supervisor sees a contender during that job's startup,
+and yields before the job loses the memory it asked for. A
+threshold reads the card too late. By the time free VRAM has dropped, the other
+job has already lost the allocation race. Free VRAM is still read, but only as a
+precondition for loading, never as the eviction signal. One blind spot is known.
+A job can take the card without registering on the KFD, as a Vulkan or a
+video-decode job would. `describe()` logs every contender's comm, and that log is
+how we find out whether the blind spot is real.
+
+`mavgpud` runs from a systemd unit on the workstation with
+`deploy/mavgpud.json` as its config, and `llama_args` is passed to llama-server
+untouched. The model, the context size, the layer count and the MTP flags are the
+owner's business and not this daemon's schema.
+
+## What stays on homesrv, permanently
+
+The **embedder** (multilingual-e5-small, ONNX, CPU). It backs the classifier, which
+must answer while the GPU is saturated. It is also cheap enough on CPU that moving
+it buys nothing. Four callers:
+
+| Caller | What for |
+|---|---|
+| `internal/router/classifier.go` | the routing floor |
+| `cmd/mavend/actions_query.go` (`queryEmbed`) | memory recall |
+| `cmd/mavend/feeds.go` | ingest embedding for every RSS item |
+| `internal/crawl/watch.go` | ingest embedding for every crawled page |
+
+`internal/speaker` becomes a fifth once it lands.
+
+## Inventory: what runs a model on homesrv today
+
+The **resident model** is one llama-server with seven callers:
+
+| Caller | What for |
+|---|---|
+| `cmd/mavend/voicewire.go` | routing |
+| `cmd/mavend/replier_llm.go` | replies |
+| `cmd/mavend/tick.go` | digestion worker: `PhraseNudge`, `PhraseReminder` |
+| `cmd/mavend/capture.go` | capture summarisation (unreachable, see #480) |
+| `cmd/mavend/mail.go` | mail extraction (off, no IMAP) |
+| `cmd/mavend/kiwixwire.go` | answering from a Kiwix, search or crawl passage |
+| `memoryeval.go`, `modelswap.go` | admin and evals |
+
+Then the embedder above, **whisper.cpp** in `mavsttd`, and **piper** in `mavttsd`.
+`mavwaked` uses no model at all: an energy-threshold VAD over 30ms frames.
+
+## Order
+
+1. **Transport** (#484). Nothing else is possible until a seam can cross a host.
+   `internal/netaddr` landed in PR #92. A seam address now carries its own scheme,
+   and a scheme-less one is still unix. A tcp seam requires a shared token, because
+   the filesystem permission that authenticated the unix socket is gone.
+2. **The resident model** (#485). Biggest quality delta. A 16GB card runs a 7-14B,
+   which fixes what the 1.7B gets wrong: world knowledge, and the persona the CPT
+   targets. The degradation path is already written and measured, since the
+   classifier scores 68.8% full accuracy at p50 16.6µs on its own.
+3. **Speech-to-text and text-to-speech** (#486). They gain a real margin, but on
+   quality alone, and both already work.
+4. **The wake word** (#487). Independent of all of the above.
+
+## Assumptions
+
+- The LAN is trusted enough that wireguard is supported but not required (owner's
+  call). What crosses the wire is still his utterances. That is why the tcp seam
+  carries its own token instead of assuming a network boundary.
+- The workstation is not expected to be up. Every child task must still serve a
+  turn while it is down.

docs/qa.md

@@ -1,19 +1,62 @@
 # QA plan: checking Maven properly
 
-*Last verified: 2026-08-02 @ 7079a24. Living doc: correct it in place, do not append.*
+*Last verified: 2026-08-02 @ 20aa2d5. Living doc: correct it in place, do not append.*
 
 Written 2026-08-01, after the 35-PR stack landed and the box came back up.
+Refreshed 2026-08-02 against the live list, after PRs #85-#90.
 
-44 of the 50 open Vikunja tasks are `QA:` tasks. They are verification work, not
+42 of the 50 open Vikunja tasks are `QA:` tasks. They are verification work, not
 build work. Most sat unverifiable while Maven was down for 11 days. That
 blocker is gone.
 
+The plan as written on 2026-08-01 named 40 task numbers. Ten open `QA:` tasks were
+missing and two of the named ones had closed. Every open task now appears below,
+the eight non-QA ones in the last two sections.
+
 This plan orders them by what unblocks what. Do sessions 1 and 2 first. Almost everything
 downstream assumes the voice loop works, and nobody has confirmed that since
 the redeploy.
 
 ---
 
+## What the 02-08-2026 run found
+
+Sessions 1, 2 and 3 all ran. Read these five before picking anything up.
+
+- **470: a question writes invented knowledge into memory.** Recall then serves
+  it back. `что дальше?` lands on `IntentFact` and stores the model's answer as a
+  `self` fact at confidence 1.00. Two junk rows then claimed seven unrelated
+  world questions through recall, outranking the search leg. A question about the
+  capital of Australia was answered `какая последняя версия языка Go?`. Two bad
+  writes silently disabled world answering, with nothing logged.
+- **466: a pending clarify is global.** One unanswerable clarify swallowed the
+  next three utterances from three separate sessions. With ntfy, telegram and
+  voice all live, a clarify raised on web chat eats the next telegram message.
+- **467: spoken task capture is dead.** The router calls the capture marker an
+  `act`, and capture is reachable only from the `note` intent.
+- **The classifier baseline in this repo was wrong**, and it flattered the
+  router. See session 2 and **464**.
+- **477: the model swap and the self-update cannot be triggered on this box.**
+  Both are built and both are correct in test. The swap needs a passkey and
+  WebAuthn is unconfigured. `mavupdate` needs to reach a socket that only an
+  in-container uid can open.
+
+- **479: an unconfigured capability lets the question escape to web search.**
+  Netscan off, asked `какие устройства в сети?`. She answered from the live web
+  with a general article about network hardware. A question about his LAN went to
+  an upstream engine. The crawler fails the same way.
+
+Twenty-one defects were filed on 02-08-2026: 462 through 482. Six tasks this plan
+had written off as blocked turned out to be ready to check. All six ran. Every
+one of them is code-correct and stops at the deploy.
+
+Three of the five config blockers in **472** were then cleared. The morning
+routine, ambient ingest, feeds, the crawler and netscan are all live. Two remain,
+and both are the owner's call: a token for each ecosystem sibling, and seed data
+in Nexus and Praxis.
+
+---
+
 ## Before you start
 
 Two things bite anyone running these checks on homesrv.
@@ -36,45 +79,75 @@ Nothing here has been confirmed since the redeploy, and everything else assumes
 it works. Do this first.
 
 Closes or advances: **44** (conversation), **45** (text chat), **287** (voice
-session quality), **321** steps 3-5 (quiet mode), **288** (STT fixtures).
+session quality), **321** steps 3-5 (quiet mode), **288** (STT golden audio).
+
+**288 is not blocked.** The fixtures are committed under `cmd/mavsttd/testdata/`
+and `make test-stt-golden` runs today. This plan said otherwise until 02-08-2026.
+
+Steps 1 and 3-6 were run on 02-08-2026 and pass. Steps 2 and 7-9 still need a
+person at the box, because they need a microphone or a nudge to arrive.
+
+Steps 1 and 3-6 do not need a browser. `POST /api/chat` takes a form-encoded
+`text=` field and a cookie jar, and answers with the rendered `/chat` page:
+
+```sh
+curl -s --noproxy '*' -c jar -b jar -L -X POST \
+  http://127.0.0.1:9201/api/chat --data-urlencode 'text=привет'
+```
+
+Parse the whole page, not the last text node. The page carries nav and footer
+text. A naive tail of the Cyrillic nodes returns the wrong string, which makes
+turns look misaligned when they are not.
 
 1. Open `http://127.0.0.1:9201/chat` and hold a short conversation in Russian.
    Watch for three things: she answers in feminine forms (`рада`, `поняла`), she
    says `ты` and never `вы`, and no pet names appear.
+   **Passes** (02-08-2026, five turns): `я рада`, `поняла`, `помогла`,
+   `проверила`, `записала`, `грустна`, `ты` throughout, no pet names.
 2. Press push-to-talk on `/dash`. Say `привет`. Confirm a spoken reply comes
    back. This is the only check that covers mic to STT to core to TTS to
    speaker as one path. It is also the path the eleven-day outage most likely
    broke.
-3. Say `тихий режим`. Expect `тихий режим включён. буду реже напоминать.`
-4. Say `выключи тихий режим`. Expect `тихий режим выключен.` Negation must win.
+3. Say `тихий режим`. Expect `тихий режим включён. буду реже напоминать.` **Passes.**
+4. Say `выключи тихий режим`. Expect `тихий режим выключен.` Negation must win. **Passes.**
 5. Say `в комнате тихо`. Quiet mode must NOT flip. Confirm on `/history` that no
-   `quiet_hours` fact was written.
+   `quiet_hours` fact was written. **Passes**: no row written. She answers `пока
+   не умею отвечать на этот вопрос.`, so it lands on `IntentSystem` with no arm.
 6. Say `включи режим тишины`, then `сделай потише`. Both must flip quiet mode
-   on. These are the noun form and the comparative, added 01-08-2026.
+   on. These are the noun form and the comparative, added 01-08-2026. **Both pass.**
 7. Wait for a nudge, then say `потом` within twenty minutes. Expect `хорошо,
    вернусь к этому позже.` and the nudge row on `/notifications` reading
    `snoozed`. Say `потом` again with nothing pending: it must route as an
    ordinary utterance, not be swallowed.
 8. Wait for the water nudge, then say `выпил воды`. Expect the ordinary fact
-   reply and nothing extra — she must not congratulate you. Check
+   reply and nothing extra. She must not congratulate you. Check
    `/notifications`: the row reads `acted`. Then trigger another nudge and say
-   `готово`; expect `отлично, отметила.` and the same outcome.
+   `готово`. Expect `отлично, отметила.` and the same outcome.
 9. Note anything where she is slow, cuts off, or talks over herself. That is
    287's whole content and it has no written acceptance criteria yet.
+   **First evidence, in text** (02-08-2026): nothing breaks, but answers wander
+   and stitch unrelated topics. Asked whether he should move flats, she opened
+   with the weather. That is 287, and it is a phrasing problem, not a loop problem.
 
-**319 is fixed** (01-08-2026). Single-word Russian utterances no longer come
+**The wake path cannot be checked as deployed.** `mavwaked` and `mavenclient`
+appear in no compose file and run as no host process. Step 2 covers only
+push-to-talk, from `/dash` through mavsttd and mavttsd. Wake word and VAD
+are untested by construction. Decide whether they belong in compose or on a
+client machine, and say which in the deploy docs. Tracked as **463**.
+
+**319's single-token bug is fixed** (01-08-2026). Single-word Russian utterances no longer come
 back as `не совсем поняла — можешь переформулировать?`. `привет` and `поужинал`
 both pass now: `thinSingleToken` spares social singles and any token carrying a
-verb ending, and only thins a bare nominal like `вода`. If a one-word utterance
-still gets clarified during the smoke test, that is a new case for the lexicon,
-not the old bug.
+verb ending, and only thins a bare nominal like `вода`. A one-word utterance that
+still gets clarified in this session is a new case for the lexicon, not the old bug.
 
 ---
 
 ## Session 2: measurement (half a day, mostly waiting)
 
-Closes or advances: **320** items 2-4, **278** (make the eval lab routine),
-**319** (gate recalibration).
+Closes or advances: **320** items 2-4, **278** (make the eval lab routine).
+Also **248** (memory evaluation), **319** (the margin gate) and **323** (the
+startup timeout arm).
 
 The resident llama-server cannot be reached by the eval harness. It binds
 `--host 127.0.0.1 --port 0` inside the container, so the port is kernel-assigned
@@ -99,14 +172,67 @@ make eval-recall
 
 A large miss against 72.7% means the deploy differs from the bench harness.
 
-Two things to decide while the numbers are in front of you:
+**Run on 02-08-2026 @ af9d213. The deploy matches the bench.** `eval-models`
+scored 56 of 77: 72.7% full, 77.9% intent-only, 2 false clarifies and 1 missed.
+That is the recorded figure to the decimal, and calendar sat at 2 of 2, so the
+stage 0 agenda rules hold. `eval-phrasing` scored 21 of 27 on the talk fixture
+against a recorded 20, and the 15 nudge templates passed every check.
 
-- **319's gate recalibration.** The single-token rule needs narrowing or
-  dropping. This needs your judgement, not a threshold sweep. The fixture and the
-  daemon disagree about what is correct on two of the three false clarifies.
+Two numbers in this repo were wrong, and both flattered the resident model.
+
+- **The classifier is not 36.8% and not 31ms.** `make eval-router` reports
+  `classifier+onnx: 53/77 (68.8% full)` at p50 16.6µs. The figure repeated here
+  and in `CLAUDE.md` predates the stage 0 rules and the seed additions. Both now
+  score inside that baseline. The accuracy gap the router buys is
+  roughly 4 points, not 36. Re-argue the trade on the real numbers: **464**.
+- **Router latency was measured under contention again.** p50 1.126s, p95 1.58s,
+  max 3.24s, against a recorded p50 825ms. The resident model was serving the
+  daemon on the same iGPU throughout. Do not record this as a regression, and do
+  not record it as a measurement either. Stop the stack before timing the router.
+
+`classifier+hash` scores 19.5%, which is the no-ONNX degraded path and is not the
+failure floor the deploy uses. Do not quote it as the classifier baseline.
+
+Then three things to decide while the numbers are in front of you:
+
+- **319 is done.** 359 gave the LLM path a real confidence signal.
+  `thinSingleToken` was narrowed on 01-08-2026, and agenda questions moved to
+  stage 0. Missed clarify sits at 1 of 6 and false clarifies at 2. Item 2 point 2
+  closed on 02-08-2026: the `make eval-recall` margin sweep is the distribution
+  that was asked for, and `0.008` sits at the knee.
+
+  | delta | answered | false recall |
+  |---|---|---|
+  | 0.005 | 18/27 | 2/5 |
+  | **0.008** | **18/27** | **1/5** |
+  | 0.010 | 16/27 | 1/5 |
+
+  It removes four of five false recalls at no cost in answers, and the next step
+  costs two answers for nothing. The hand-picked value survives on evidence.
 - **278's real ask** is making the eval lab routine rather than building it. It
   is built. Decide whether it runs on a timer, on every merge, or on demand, and
   the task can close.
+- **248** is the memory evaluation loop. It ships, it writes notes, and it cannot
+  speak. `make eval-recall` covers the retrieval half. The open question is whether
+  a written evaluation nobody reads is worth the tick.
+
+**323 is down to one check.** PR #90 covered the spawn path and took phraser
+coverage to 76.9%. Only the 60s startup timeout arm is untested, because testing it
+needs a `StartupTimeout` field on `Config` rather than a test-only hack. While you
+are on the box, time a cold 1.7B load off spinning disk. If it runs near 60s, the
+default is too tight and the field earns itself twice.
+
+Warm, it is nowhere near. A second llama-server answered `/health` 1.8s after
+launch at `n_ctx 4096` on 02-08-2026. That is page cache, so it does not settle
+the question. A cold read needs a cache drop, which needs root.
+
+**`CheckFeminine` has a false positive.** On 02-08-2026 it failed
+`query-notes-do-not-answer` for `ты заплатил`, calling it masculine
+self-reference. Masculine second person is correct, because the owner is male.
+The check matches a masculine
+past-tense verb before `за` without confirming the subject is `я`. Fix it in
+`internal/phraser/eval/checks.go` before trusting a phrasing score to the case.
+The real talk-fixture score on that run is 22 of 27, not 21. Tracked as **462**.
 
 Item 4 of **320** needs a permission I do not have. Kill the `llama-server`
 pid under `maven-mavend-1`, post a turn, and confirm it still completes
@@ -115,47 +241,269 @@ check that the failure floor catches a mid-session model death.
 
 ---
 
-## Session 3: the interaction batch (a day, or three sittings)
+## Session 3: the interaction batch (a day, or five sittings)
 
 These need real use rather than a command, grouped by what one sitting covers.
 
-**Morning and delivery** (**280**, **281**, **128**, **282**): open `/morning`,
-walk the seven required behaviours, then check the four interruption outcomes
-and the digest gap. **282** needs the `desk_active` script enabled on the desk
-PC first, which is **15** and needs you at that machine.
+**Morning and delivery** (**280**, **281**, **128**, **282**, **283**, **285**):
+open `/morning`, walk the seven required behaviours, then check the four
+interruption outcomes and the digest gap. **282** needs the `desk_active` script
+enabled on the desk PC first, which is **15** and needs you at that machine.
+**283** is the event intake envelope every reach shares, so a delivery check
+exercises it whether you name it or not. **285** is not verification: the bridge
+framework works and the remaining ask is more adapters. Decide which reach comes
+next, or park it.
+
+Run 02-08-2026. **280 is blocked.** No morning routine is configured (**472**).
+`morning.Item` also has no required-versus-optional field, so behaviour 1 cannot
+hold whatever you configure (**473**). **281's digest gap is closed**, and
+its presence rule passes on inspection. Three of its five items need traffic the
+box has not had. **283 is blocked**: nothing feeds the intake journal. **128
+found the worst defect of the whole session, see below.**
+
+Three of 472's five blockers were cleared the same day, in `deploy/mavend.json`
+and `docker-compose.yml`.
+
+- A `morning_routines` block, one routine `утро` 08:00-11:00 with medicine,
+  water and pets. It is live: the dispatcher logged `dropped morning:утро (sev1,
+  presence=away)`, so the plan builds and the nudge is proposed. 280's
+  behaviours and 128 step 11 are checkable now. 473 still stands.
+- `-ambient-token` on mavweb, value in a gitignored `/.env` that docker compose
+  reads for interpolation. `/api/ambient` answers 401 without the token and 201
+  with it, storing `calendar_event_20260802_Standup`. 283 step 5 and 128 step 8
+  are unblocked. The token is a flag, so it shows in `ps` inside that container.
+  The zenmoney and IMAP secrets are read from files instead. Ingest also
+  reads the notification's wall clock as UTC and stores a 14:30 meeting at 18:30
+  (**482**).
+- `feeds` (two sources), `crawl.on_demand` and `netscan.enabled`. The intake
+  journal now fills: `/events` holds `scan:lan` and `ambient:notif` rows.
+
+Two are not mine to clear. No sibling has a `token` in `deploy/mavend.json`, so
+273 steps 6 and 8 need a credential decision. Nexus has no entities and Praxis no
+attention items, so 272 step 3 needs seed data whose content is the owner's call.
+
+For **285**, two facts bear on the choice. Synapse is already running on this box
+and healthy, so a Matrix reach has a live target and needs no new service. And
+mavweb is already a PWA with a service worker, which 285 itself calls the highest
+value adapter left. Today's reaches are ntfy, telegram and voice.
+
+**Query sources** (**258**, **286**): ask her something the RSS feeds answer and
+something only a ZIM answers, with the search block on. Live search leads and the
+ZIMs are the fallback since 02-08-2026. **286**'s remaining half is doc and
+git ingestion, which is build work, not a check.
+
+**Do not read `/trace` for this.** `/trace` is the nudge-rule trace: rule,
+severity, predicate, gate, selected. No query-source field exists anywhere in the
+codebase. The only evidence of which query source claimed a turn is the
+`voice: search:` and `voice: kiwix:` lines in `docker compose logs mavend`
+(`actions_query.go:589` and `:660`).
+
+Run 02-08-2026, 20 turns. **Search leads and the personal boundary holds.** Every
+world question that reached the boundary was claimed by search. All three
+personal questions produced no search and no kiwix line at all.
+
+The rest of this sitting went badly. **Kiwix has zero live coverage.** SearXNG
+returns four results for everything, including two invented nonsense terms. So
+`querySearch` always claims, and Kiwix is unreachable code as deployed. The ZIM
+half of the 02-08-2026 decision is unverified. A ZIM answer cannot signal a
+silent search failure, because a ZIM answer cannot happen.
+**Ordering defects** in feeds and calendar, plus 258 step 1's utterance not
+working: **474**. And the sitting independently found stage 2 of **470**.
 
 **Tasks and calendar** (**129**, **130**, **127**, **126**, **246**): capture a
 task by voice, confirm it lands, check prioritisation ordering is not nonsense.
 **246** (mail reader) also exercises the `IngestMail` rung that moved to
 `AuthWrite` this morning.
 
+Run 02-08-2026. **129 passes.** The page and the spoken answer agree on ordering.
+The undistinguished task carries no invented reason on either surface, which is
+the thing 129 asks for. **130 fails outright** and **127 half fails**:
+**467**, **469**. **246 cannot be run**: `mavmaild` is commented out in
+`docker-compose.yml` and there is no `email` block, so nothing in steps 4-13 is
+reachable. The `IngestMail` rung does sit at `AuthWrite`
+(`internal/auth/policy.go:96`, asserted in `auth_test.go:421`), verified by
+reading only.
+
 **Routines and patterns** (**43**, **46**, **247**, **254**): these need history
 to detect against. If the database is thin after the outage, they may have
 nothing to propose, which is not a failure. Check `/routines` before
 concluding anything.
 
+Run 02-08-2026. The answer is the middle case: **the detector ran and found
+nothing.** The tick loop is live, and `detectPatterns` is called unconditionally
+at `cmd/mavend/tick.go:227`. It has run about 25 times since the restart. It
+finds nothing because the events table is empty upstream of it. Rows land there
+only from `pattern.Extract` at fact-write time, and `Extract` requires the fact
+value to match a closed 7-action lexicon. All 200 facts on `/history` are
+`page_heartbeat`, `netdata_alarm`, `quiet_hours`, `name`, `service_down` and
+`рост`. Not one lexicon hit, so no event can exist, let alone the four one pair
+needs. **46 step 5 passes**: `/routines` renders `noticed 0` with the empty state
+and the hint string.
+
+Two things block this sitting, and both are build work. The seeding recipe on
+**43** goes through `sqlite3` and cannot work. And `pattern.Detect` has no
+minimum-interval floor, so seeding by hand mints a permanent false routine
+(**468**). Do not try to seed a pattern with four fast chat turns.
+
 **Ecosystem** (**272**, **273**, **276**): nexus, hexis and praxis are wired and
-logged clean at boot. **276** is the degraded-mode suite, which means taking
-siblings down on purpose. Worth doing while you are already in there.
+logged clean at boot.
+
+Run 02-08-2026, read-only half. All three answer `/health` 200 and `/ecosystem`
+lists 18 Hexis capabilities with correct read-only and mutating badges. **272 and
+273 are blocked on empty data**, not on code. Nexus holds no entities, Praxis
+holds no attention items, and the Calls panel has never recorded a call. See
+**472**, and read its warning first. 273's trace fix has never been validated
+here. An empty Calls panel is exactly what the old bug looked like. The page is
+`/ecosystem`, not `/siblings`.
+
+**276 ran 02-08-2026 and the suite is sound.** 17 `TestEcosystem_` cases pass
+under `-race`, not the 10 the task describes. The mutation check bites: patching
+the Nexus-error branch of `handleHexisAct` to `return ""` fails
+`TestEcosystem_MalformedNexusResponseFailsClosed` on the expected line.
+
+Steps 4 and 6 could not be checked through chat, because no utterance reaches
+Praxis (**475**). «что требует внимания» routes to `intent=query` and is answered
+by the search leg, identically whether `ecosystem-praxis-1` is up or stopped. The
+degraded string never appears because its branch is never entered. Step 5 is
+blocked the same way: `перезапусти muzick indexer` clarifies on
+`HasFn:false`, and the router had already rewritten the entity name to
+`музик индексер` (**476**).
+
+Both steps were checked on `/ecosystem` instead, which reads Praxis directly.
+With Praxis stopped the card reads `praxis — unreachable` while Nexus and Hexis
+keep rendering. On `docker start` the card returns to `nothing needs attention.`
+with no mavend restart. Independent degradation and recovery both hold.
+
+**Operations** (**249**, **250**): both ran 02-08-2026. The code is correct and
+neither lever can be pulled on this box. See **477**.
+
+**250** passes steps 1, 2, 3, 9 and 10 on the deploy. The capability announces
+itself. `/models` names the model llama-server reports, not the config filename.
+Asking her to switch models does nothing. Removing `swap_models` renders `swap
+not configured`. Step 4's refusal half passes at HTTP 403, and the 403 comes from
+mavend rather than mavweb. WebAuthn is unconfigured, so the web gate fails open
+and the wire gate fails closed. Steps 5 to 8 need a passkey assertion nothing on
+this box can produce. They pass in test: 13 swap cases and 7 page cases covering
+drain, mid-swap refusal, rollback, failed rollback and the not-owned refusal.
+
+**249** passes steps 1 and 2. Step 3 stops it. `mavupdate` health-checks over
+`/run/maven/mavend.sock`, which is `srw------- 1 10001 999` inside a docker
+volume. The host owner cannot traverse `/var/lib/docker/volumes` and cannot
+connect to a socket owned by an in-container uid. `mavupdate` assumes a
+host-installed daemon and the deploy is containers. Do not sudo around this.
 
 ---
 
-## Housekeeping (one sitting, no box needed)
+## Housekeeping (done 02-08-2026, and this section was mostly wrong)
 
-Four QA tasks will not close no matter how long they sit, because they are
-gated on something that does not exist:
+This section claimed eleven tasks were not verification work. **Three were not.
+The other eight are.** Every one of the eight has shipped, tested code behind it.
+The error ran one way: it wrote off work that is ready to check. Do not trust a
+"nothing is built" line in this plan without grepping for the package first.
 
-- **125** zenmoney: needs a token you have not minted.
-- **256** Home Assistant: needs HA configured.
-- **257** Bluetooth: BLOCKED, no bluez on the box. Says so in the title.
-- **288** STT golden audio: needs fixtures generated.
+Relabelled to `Blocked:`, claim verified:
 
-Relabel these so they stop reading as backlog. They are not verification work
-that is pending, they are work that has not started.
+- **125** zenmoney. `internal/zenmoney/` ships and is tested against a fixture.
+  `deploy/zenmoney.token` does not exist and the compose mount is commented out.
+  One token unblocks it.
+- **256** Home Assistant. `internal/smarthome/` ships, the `smarthome` block sits
+  in `deploy/mavend.json` at `enabled: false`, and 8123 and 1883 are closed.
+- **14** cold-start unlock. The seam is real at `cmd/mavend/main.go:128` and
+  `internal/webauthn/prf.go` is in place. `lockedAPI` is gone, replaced by
+  `Server.Check` in `internal/ipc/server.go`. Gated on an authenticator that
+  implements the WebAuthn PRF extension, which is hardware, not code.
 
-Same treatment for the five plan-only tasks (**251** MCP, **252** vision,
-**253** hearing, **255** speaker recognition, **259** crawler). A `QA:` prefix on
-a plan is misleading.
+Left alone, because the claim here was false:
+
+- **284** simulator. `cmd/mavend/simulator_test.go`, three scenarios under
+  `cmd/mavend/testdata/scenarios/`, and a `simulate` target at `Makefile:98`.
+  **Run 02-08-2026: all three scenarios pass**, plus the determinism and
+  backwards-step guards. One defect found, see below.
+- **288** STT golden audio. Four WAVs and `golden_v1.json` are committed under
+  `cmd/mavsttd/testdata/`, the make targets exist, and `models/stt/ggml-small.bin`
+  is on the box. Session 1 lists 288 as blocked on fixtures, which is wrong.
+  **Run 02-08-2026: all four pass**, WER at or under ceiling with no drift.
+
+| fixture | transcript | WER | ceiling |
+|---|---|---|---|
+| ru_reminder | `Напомни мне через час позвонить маме.` | 0.00 | 0.10 |
+| ru_fact | `А отметь, что я выпил воды.` | 0.20 | 0.25 |
+| ru_query | `Что у меня сегодня по календарю?` | 0.00 | 0.10 |
+| en_act | `Restart the web server and check the disk space.` | 0.00 | 0.10 |
+
+That also settles a session 1 worry indirectly: whisper.cpp works on Vulkan
+after the redeploy. Only the mic and the wake path remain unproven.
+
+**The simulator routes with an empty seed set.** Every `make simulate` run logs
+`loaded 0 seed examples from models/seeds`, seven times per scenario. The test
+runs from `cmd/mavend`, and the seed path is relative to the repo root. The
+scenarios still pass, which means they pass without the classifier having any
+seeds to match against. Whatever 284 is proving, it is not proving the routing
+the deploy runs. Fix the path before trusting a green simulator.
+- **257** Bluetooth. The bluez half is genuinely absent. The LAN-scan half shipped
+  (`internal/netscan/`), and steps 1-9 run today. Only step 10 is Bluetooth, so
+  relabelling the whole task would bury real pending work.
+- **251** MCP, **253** hearing, **259** crawler. All three ship
+  (`internal/mcp/`, `internal/capture/`, `internal/crawl/`) with no external gate.
+  Fully checkable. `259`'s step 1 wants no `crawl` block in `deploy/mavend.json`,
+  and there is none, so it is already set up correctly.
+- **252** vision and **255** speaker recognition. Both ship. Each is blocked only
+  on a model download: a vision gguf with mmproj, and a speaker embedding model.
+  Neither is present under `/mnt/hdd1`. Their refusal-path steps run today.
+
+So the honest split is three blocked on a credential or hardware, two blocked on
+a download, and six ready to check. That is roughly a session of real QA this
+plan had written off as backlog.
+
+**All six ran on 02-08-2026.** Every one of them is code-correct and stops at the
+deploy. The pattern repeats often enough to be the headline: the packages pass,
+and the box cannot reach them.
+
+**251, MCP.** Steps 1, 2, 3, 4 and 13 pass. Package tests green under `-race`.
+Off-by-default is clean, and the SSRF refusal is exact: without `allow_private`
+the log reads `refusing to connect to a private address: 127.0.0.1` and `/tools`
+shows the server down with zero proposals. Steps 5 to 12 are blocked. `ss -lntp`
+shows the Vikunja MCP server on `127.0.0.1:9100` only, so no container reaches it
+at any address (**478**). `allow_private` does work, measured both ways.
+
+**253, hearing.** Steps 1, 2 and 17 pass. `internal/capture` covers 90.3%. Steps
+7 to 16 are blocked on something nobody can work around: no shipped client calls
+`CaptureStart`. There is no `cmd/mavheard`, no mavweb route, and `mavenclient`
+never calls it (**480**). Two of its QA steps are also stale.
+
+**257, netscan.** Steps 2, 3 and 9 pass at unit level. Step 1 fails. Steps 4 to 8
+need the block enabled. Step 10 is Bluetooth and stays skipped.
+
+**259, crawler.** Steps 1 and 15 pass. Step 2 fails. Steps 3 to 14 need a `crawl`
+block that nobody has written.
+
+Both were configured later the same day, and both work. `netscan.enabled: true`
+answers `какие устройства в сети?` with `нашла 3 устройства, из них 2 с вебом, 2 с
+ssh. список записала.` and the scan lands in the intake journal as `scan:lan`.
+`crawl.on_demand: true` answers `посмотри https://lwn.net — что там пишут?` from
+the real page. So **479** is one defect, not the routing defect it was filed as.
+An unconfigured capability declines its own turn instead of naming the gap.
+Nothing is wrong with the routing.
+
+257 step 1 and 259 step 2 fail the same way and share a task (**479**). An
+unconfigured capability does not name the gap, so the question escapes to web
+search. `какие устройства в сети?` was answered with a general article about
+network hardware. That is his LAN going to an upstream engine.
+
+**252 vision and 255 speaker.** Both confirmed blocked. The disk claim was
+re-verified rather than taken on trust: 16 text-only ggufs under `/mnt/hdd1`, no
+mmproj and no speaker embedding model. Everything not needing the model passes,
+including the two refusals that matter. `TestNewLocalRefusesNonPrivateEndpoints`
+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.
 
 ---
 
@@ -171,16 +519,39 @@ Not QA. These are blocked on a decision or a credential only you have.
 | 355 | Deploy the Hexis auth change. Was blocked on Maven being under construction, which it no longer is. The client half is vendored and wired. |
 | 357 | Decide whether entity-existence validation is the permanent target guard or whether blessing lands in Nexus. |
 | 275 | Hexis native API and MCP parity. |
-| — | Decide on `-require-stepup`. Making it the default needs WebAuthn configured first, or it locks you out of your own admin surfaces. See **317**. |
-| — | Three nginx sites bind wildcard `:80` (`acme.conf`, `matrix`, `panel`), so the ecosystem's bind-level protection is not in effect and `allow`/`deny` is carrying it alone. See **354**. |
+| — | Decide on `-require-stepup`. Making it the default needs WebAuthn configured first, or it locks you out of your own admin surfaces. |
+
+317 and 354 closed on 01-08-2026. The step-up gate now covers `POST /api/chat` and
+`/routines`, and the nginx template is locked down with a `maven.<domain>` block for
+mavweb. The `-require-stepup` default is still your call.
+
+---
+
+## Not this repo
+
+Two open tasks sit on the Maven board and are not Maven work. Move them or note
+where they land, so the board stops reading as 50 things Maven owes.
+
+- **358** replace the rowid execution cursor with a real seq column. This is Hexis,
+  and it must land before any execution retention or pruning does.
+- **362** mirror the router prompt reorder into the relabelling prompt. This is the
+  training workspace, enforced by `llm/check_prompt_parity.py` there, not here.
 
 ---
 
 ## Suggested order
 
 1. Session 1. If the voice loop is broken, nothing else matters.
-2. The `-require-stepup` and Kuma decisions. Five minutes, unblocks **317** fully
-   and **16**.
-3. Session 2. The numbers tell you whether the router is worth its 90x latency.
+2. The `-require-stepup` and Kuma decisions. Five minutes, and it unblocks **16**.
+3. Session 2. **Run on 02-08-2026.** The numbers came back worse for the router
+   than the docs claimed. The classifier is 68.8%, not 36.8%, and 16.6µs, not
+   31ms. The router buys about 4 points of accuracy for four orders of magnitude
+   of latency. Whether that still earns its place is now an open question.
 4. Housekeeping. Cheap, and it makes the remaining backlog honest.
-5. Session 3, split whichever way suits you.
+5. Session 3, split whichever way suits you. All five sittings ran on
+   02-08-2026. Read the per-sitting notes before repeating any of them.
+
+The next thing to fix is not in this plan. Four defects say the same sentence:
+a capability is built and no utterance reaches it. **466** (a clarify is global),
+**467** (capture is act-routed), **475** (attention is act-routed), **476** (the
+router rewrites entity names). Routing is where the work is.

internal/ipc/client.go

@@ -8,13 +8,15 @@ import (
 	"net"
 	"sync"
 	"time"
+
+	"github.com/kami/maven/internal/netaddr"
 )
 
-// Client — the module side of the boundary. Wraps a unix-socket connection
-// and satisfies CoreAPI, so a module imports ipc, holds a CoreAPI, and is
-// agnostic to whether it's been wired in-process (tests / daemon-embedded)
-// or over this socket (full topology). The swappability is the seam auth
-// will insert into without touching module code.
+// Client — the module side of the boundary. Wraps a connection to core and
+// satisfies CoreAPI, so a module imports ipc, holds a CoreAPI, and is
+// agnostic to whether it's been wired in-process (tests / daemon-embedded),
+// over a local unix socket, or over tcp to another host. The swappability is
+// the seam auth will insert into without touching module code.
 //
 // One Client ⇒ one conn ⇒ one concurrent request at a time. A module that
 // wants parallel requests opens one Client per goroutine; the store is the
@@ -22,7 +24,8 @@ import (
 // per-Client lock keeps frame interleaving impossible by construction.
 type Client struct {
 	conn net.Conn
-	path string // kept so a dropped conn can be re-dialed (core restart)
+	path string       // the address as configured, kept for errors and logs
+	addr netaddr.Addr // parsed, so a dropped conn can be re-dialed (core restart)
 	mu   sync.Mutex
 }
 
@@ -81,14 +84,22 @@ var readOnlyMethods = map[Method]bool{
 	MethodPing:                  true,
 }
 
-// Dial connects to a core socket at path and returns a Client. The module
-// owns its Client lifecycle; Close on shutdown.
+// Dial connects to core at path and returns a Client. The module owns its
+// Client lifecycle; Close on shutdown.
+//
+// path is a netaddr seam address: a bare path is the unix socket it has
+// always been, and "tcp://host:port?token=..." reaches a core on another
+// host. See internal/netaddr.
 func Dial(path string) (*Client, error) {
-	c, err := net.Dial("unix", path)
+	addr, err := netaddr.Parse(path)
 	if err != nil {
-		return nil, fmt.Errorf("ipc: dial %s: %w", path, err)
+		return nil, err
 	}
-	return &Client{conn: c, path: path}, nil
+	c, err := netaddr.Dial(addr)
+	if err != nil {
+		return nil, fmt.Errorf("ipc: dial %s: %w", addr, err)
+	}
+	return &Client{conn: c, path: path, addr: addr}, nil
 }
 
 func (c *Client) Close() error {
@@ -189,9 +200,9 @@ func (c *Client) call(ctx context.Context, m Method, params, result any) error {
 // re-dials clean. Caller holds c.mu.
 func (c *Client) roundtrip(m Method, raw json.RawMessage, resp *Response) error {
 	if c.conn == nil {
-		conn, err := net.Dial("unix", c.path)
+		conn, err := netaddr.Dial(c.addr)
 		if err != nil {
-			return fmt.Errorf("%w: dial %s: %v", errWriteLost, c.path, err)
+			return fmt.Errorf("%w: dial %s: %v", errWriteLost, c.addr, err)
 		}
 		c.conn = conn
 	}

internal/ipc/server.go

@@ -8,11 +8,11 @@ import (
 	"fmt"
 	"log"
 	"net"
-	"os"
 	"sync"
 	"sync/atomic"
 	"time"
 
+	"github.com/kami/maven/internal/netaddr"
 	"github.com/kami/maven/internal/store"
 	"golang.org/x/sys/unix"
 )
@@ -446,6 +446,7 @@ func mapErr(err error) error {
 type Server struct {
 	api  atomic.Value // stores CoreAPI
 	path string
+	addr netaddr.Addr
 
 	ln     net.Listener
 	wg     sync.WaitGroup
@@ -610,31 +611,29 @@ type CheckFunc func(ctx context.Context, m Method, params json.RawMessage) error
 // MethodAssertStepUp dispatch calls this instead of going through CoreAPI.
 type StepUpFunc func(ctx context.Context) error
 
-// Listen creates a Server bound to path. path's parent dir must exist and be
-// 0700 (we chmod it if we own it); the socket file itself is created 0600 so
-// only the same unix user can connect — the current "auth floor", same radius
-// as wg at the network boundary. Removing a stale socket at path first lets
-// the daemon restart cleanly.
+// Listen creates a Server bound to path.
+//
+// A bare path is a unix socket, unchanged: its parent dir is 0700 and the
+// socket file itself is 0600, so only the same unix user can connect — the
+// current "auth floor", same radius as wg at the network boundary. A stale
+// socket is removed first so the daemon restarts cleanly.
+//
+// A "tcp://host:port?token=..." address binds a network listener instead, for
+// a module that lives on another host. There is no filesystem there to be the
+// auth floor, so netaddr checks the shared token before this package sees the
+// connection and a token is mandatory. See internal/netaddr.
 func Listen(path string, api CoreAPI) (*Server, error) {
-	_ = os.Remove(path) // stale socket from a crashed daemon; ignore missing
-	if err := os.MkdirAll(parentDir(path), 0o700); err != nil {
-		return nil, fmt.Errorf("ipc: mkdir socket dir: %w", err)
-	}
-	// umask could widen the perms on socket creation; tighten then chmod to
-	// be explicit. 0600 ⇒ read+write by owner only.
-	oldMask := unix.Umask(0o077)
-	ln, err := net.Listen("unix", path)
-	unix.Umask(oldMask)
+	addr, err := netaddr.Parse(path)
 	if err != nil {
-		return nil, fmt.Errorf("ipc: listen %s: %w", path, err)
+		return nil, err
 	}
-	if err := os.Chmod(path, 0o600); err != nil {
-		_ = ln.Close()
-		_ = os.Remove(path)
-		return nil, fmt.Errorf("ipc: chmod socket: %w", err)
+	ln, err := netaddr.Listen(addr)
+	if err != nil {
+		return nil, err
 	}
 	s := &Server{
 		path: path,
+		addr: addr,
 		ln:   ln,
 		done: make(chan struct{}),
 	}
@@ -1268,7 +1267,7 @@ func (s *Server) Close() error {
 		// missing the seal costs every write since the last clean shutdown.
 		log.Printf("ipc: %d connection(s) still busy after %s, closing anyway", s.liveConns(), closeGrace)
 	}
-	_ = os.Remove(s.path)
+	netaddr.Cleanup(s.addr)
 	return err
 }
 
@@ -1338,25 +1337,6 @@ func (s *Server) Path() string { return s.path }
 // while the server is serving (dispatch loads api once per request via atomic).
 func (s *Server) SetAPI(api CoreAPI) { s.api.Store(api) }
 
-func parentDir(p string) string {
-	if i := lastIndexByte(p, '/'); i >= 0 {
-		if i == 0 {
-			return "/"
-		}
-		return p[:i]
-	}
-	return "."
-}
-
-func lastIndexByte(s string, b byte) int {
-	for i := len(s) - 1; i >= 0; i-- {
-		if s[i] == b {
-			return i
-		}
-	}
-	return -1
-}
-
 // peerCaller — read SO_PEERCRED off a unix conn to identify the connecting
 // process. Returns ok=false on a non-unix conn or a platform without
 // SO_PEERCRED; the caller then proceeds without a Caller (the socket perms

internal/llm/remote.go

@@ -0,0 +1,193 @@
+package llm
+
+import (
+	"context"
+	"errors"
+	"log"
+	"net/http"
+	"sync/atomic"
+	"time"
+)
+
+// Pair — a preferred model on another host, with the resident one as the floor.
+//
+// homesrv cannot grow a GPU and the workstation has 16GB of VRAM, so the big
+// model runs there and the resident Qwen3-1.7B stays here. See docs/offload.md.
+// The workstation is never assumed up: its GPU is often busy with CPT runs and
+// the manga-recap pipeline, and the machine sleeps. So the remote is preferred,
+// never required, and Pair is what makes "preferred" mean something precise.
+//
+// This is admission control, not a scheduler. There is no arbiter deciding who
+// gets the card. A prober asks the remote whether it will take work, caches the
+// answer, and every request reads that cached answer in nanoseconds. Routing
+// sits on the hot path at p50 825ms and must never wait on a machine that may
+// be asleep, so no request ever pays for a health check itself.
+//
+// Pair satisfies nothing by itself. Callers pick a method by which half of the
+// degradation rule they live under:
+//
+//   - Complete falls back silently. For routing, replies, and nudge phrasing,
+//     where the big model is only better and the 1.7B is today's shipping
+//     quality. He is not told which model phrased his reply.
+//   - CompleteRemote returns ErrRemoteUnavailable instead of falling back. For
+//     a world question, or a long Kiwix or search passage, where a 1.7B
+//     confabulates rather than summarises. A named gap beats an invented
+//     answer.
+type Pair struct {
+	remote *Client
+	floor  *Client
+
+	// up — the cached admission answer, written only by the prober goroutine
+	// and read by every request. Atomic so the read costs nanoseconds and no
+	// request ever contends with the prober.
+	up atomic.Bool
+
+	health   string
+	interval time.Duration
+	http     *http.Client
+	stop     chan struct{}
+}
+
+// ErrRemoteUnavailable — the workstation model was required and is not
+// answering. Callers on the naming half of the degradation rule turn this into
+// a gap in the reply ("не могу сейчас"), never into a guess from the floor.
+var ErrRemoteUnavailable = errors.New("llm: workstation model unavailable")
+
+// ErrNoFloor — a Pair was built with no resident model to fall back to. A
+// configuration mistake: the floor is the whole point.
+var ErrNoFloor = errors.New("llm: no floor client")
+
+// NewPair builds the two-model arrangement. remote may be nil, which is the
+// unconfigured deploy and must behave exactly as the box behaves today: every
+// call goes to the floor and nothing probes anything.
+//
+// health is the URL the prober asks. llama-server's /health answers "is a model
+// loaded and ready", which is the useful signal here, because llama-server
+// refuses to load at all when VRAM is short. That makes a busy card detectable
+// without any cooperation from the owner's other jobs.
+func NewPair(remote, floor *Client, health string, interval time.Duration) *Pair {
+	p := &Pair{
+		remote:   remote,
+		floor:    floor,
+		health:   health,
+		interval: interval,
+		http:     &http.Client{Timeout: probeTimeout},
+		stop:     make(chan struct{}),
+	}
+	return p
+}
+
+// probeTimeout — a remote that cannot answer /health this fast is not going to
+// serve a turn either. Short on purpose: the prober runs on its own goroutine,
+// but a slow probe still delays the moment Maven notices the card came back.
+const probeTimeout = 2 * time.Second
+
+// Start begins probing. It returns immediately, and the first probe runs before
+// the first tick so a remote that is already up is used on the first turn
+// rather than after one interval of falling back. Safe to call with a nil
+// remote; it does nothing.
+func (p *Pair) Start(ctx context.Context) {
+	if p.remote == nil || p.health == "" {
+		return
+	}
+	go func() {
+		p.probe(ctx)
+		t := time.NewTicker(p.interval)
+		defer t.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case <-p.stop:
+				return
+			case <-t.C:
+				p.probe(ctx)
+			}
+		}
+	}()
+}
+
+// Stop ends the prober. Idempotent.
+func (p *Pair) Stop() {
+	select {
+	case <-p.stop:
+	default:
+		close(p.stop)
+	}
+}
+
+// Available reports whether the workstation will take work right now. It reads
+// a cached flag, so it is safe to call per turn on the hot path. A false answer
+// is never stale in the direction that matters: the worst case is that Maven
+// falls back for up to one probe interval after the card frees up.
+func (p *Pair) Available() bool {
+	return p.remote != nil && p.up.Load()
+}
+
+func (p *Pair) probe(ctx context.Context) {
+	ctx, cancel := context.WithTimeout(ctx, probeTimeout)
+	defer cancel()
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, p.health, nil)
+	if err != nil {
+		p.set(false)
+		return
+	}
+	resp, err := p.http.Do(req)
+	if err != nil {
+		p.set(false)
+		return
+	}
+	defer resp.Body.Close()
+	p.set(resp.StatusCode == http.StatusOK)
+}
+
+// set records the admission answer and logs only the transitions. A machine
+// that sleeps every night would otherwise write one line per interval forever.
+func (p *Pair) set(up bool) {
+	if p.up.Swap(up) == up {
+		return
+	}
+	if up {
+		log.Printf("llm: workstation model available at %s", p.health)
+	} else {
+		log.Printf("llm: workstation model unavailable, falling back to the resident model")
+	}
+}
+
+// Complete runs r on the workstation when it will take work, and on the
+// resident model otherwise. A remote that fails mid-request falls back too: the
+// admission answer is a cache and can be one interval out of date, so an error
+// here is expected rather than exceptional.
+//
+// This is the silent half of the degradation rule. It must be indistinguishable
+// from today's behaviour when the workstation is down.
+func (p *Pair) Complete(ctx context.Context, r Req) (string, error) {
+	if p.floor == nil {
+		return "", ErrNoFloor
+	}
+	if p.Available() {
+		out, err := p.remote.Complete(ctx, r)
+		if err == nil {
+			return out, nil
+		}
+		// The cached answer was wrong. Correct it now rather than sending the
+		// next request into the same hole, then fall back.
+		p.set(false)
+	}
+	return p.floor.Complete(ctx, r)
+}
+
+// CompleteRemote runs r on the workstation or refuses. It never falls back,
+// because for a world question the resident 1.7B does not answer worse, it
+// invents. Callers turn ErrRemoteUnavailable into a named gap.
+func (p *Pair) CompleteRemote(ctx context.Context, r Req) (string, error) {
+	if !p.Available() {
+		return "", ErrRemoteUnavailable
+	}
+	out, err := p.remote.Complete(ctx, r)
+	if err != nil {
+		p.set(false)
+		return "", errors.Join(ErrRemoteUnavailable, err)
+	}
+	return out, nil
+}

internal/llm/remote_test.go

@@ -0,0 +1,210 @@
+package llm
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// completionServer stands in for a llama-server. It counts what reached it, so
+// a test can say which of the two models answered.
+func completionServer(t *testing.T, reply string, hits *atomic.Int64) *httptest.Server {
+	t.Helper()
+	s := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hits.Add(1)
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"` + reply + `"}}]}`))
+	}))
+	t.Cleanup(s.Close)
+	return s
+}
+
+func healthServer(t *testing.T, ok *atomic.Bool) *httptest.Server {
+	t.Helper()
+	s := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if !ok.Load() {
+			w.WriteHeader(http.StatusServiceUnavailable)
+			return
+		}
+		w.WriteHeader(http.StatusOK)
+	}))
+	t.Cleanup(s.Close)
+	return s
+}
+
+// waitFor polls until cond holds or the deadline passes. The prober runs on its
+// own goroutine, so a test has to wait for it rather than assume it has run.
+func waitFor(t *testing.T, cond func() bool) bool {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if cond() {
+			return true
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	return false
+}
+
+// The unconfigured deploy. No remote, no probing, every call to the floor —
+// exactly what the box does today.
+func TestNoRemoteGoesToTheFloor(t *testing.T) {
+	var floorHits atomic.Int64
+	floor := completionServer(t, "floor", &floorHits)
+
+	p := NewPair(nil, New(floor.URL, time.Second), "", time.Second)
+	p.Start(context.Background())
+	defer p.Stop()
+
+	if p.Available() {
+		t.Fatal("a Pair with no remote reports available")
+	}
+	out, err := p.Complete(context.Background(), Req{User: "привет"})
+	if err != nil {
+		t.Fatalf("complete: %v", err)
+	}
+	if out != "floor" || floorHits.Load() != 1 {
+		t.Fatalf("out = %q, floor hits = %d", out, floorHits.Load())
+	}
+}
+
+// The workstation is up, so it answers and the resident model is not touched.
+func TestAvailableRemoteAnswers(t *testing.T) {
+	var remoteHits, floorHits atomic.Int64
+	remote := completionServer(t, "remote", &remoteHits)
+	floor := completionServer(t, "floor", &floorHits)
+	up := &atomic.Bool{}
+	up.Store(true)
+	health := healthServer(t, up)
+
+	p := NewPair(New(remote.URL, time.Second), New(floor.URL, time.Second), health.URL, 20*time.Millisecond)
+	p.Start(context.Background())
+	defer p.Stop()
+	if !waitFor(t, p.Available) {
+		t.Fatal("prober never saw the remote come up")
+	}
+
+	out, err := p.Complete(context.Background(), Req{User: "привет"})
+	if err != nil {
+		t.Fatalf("complete: %v", err)
+	}
+	if out != "remote" || floorHits.Load() != 0 {
+		t.Fatalf("out = %q, floor hits = %d", out, floorHits.Load())
+	}
+}
+
+// The card is busy, so /health refuses and Complete degrades silently. This is
+// the constraint from 483: the workstation being down is indistinguishable from
+// today's behaviour.
+func TestBusyCardFallsBackSilently(t *testing.T) {
+	var remoteHits, floorHits atomic.Int64
+	remote := completionServer(t, "remote", &remoteHits)
+	floor := completionServer(t, "floor", &floorHits)
+	health := healthServer(t, &atomic.Bool{}) // never ok
+
+	p := NewPair(New(remote.URL, time.Second), New(floor.URL, time.Second), health.URL, 20*time.Millisecond)
+	p.Start(context.Background())
+	defer p.Stop()
+	time.Sleep(60 * time.Millisecond)
+
+	out, err := p.Complete(context.Background(), Req{User: "привет"})
+	if err != nil {
+		t.Fatalf("complete: %v", err)
+	}
+	if out != "floor" || remoteHits.Load() != 0 {
+		t.Fatalf("out = %q, remote hits = %d", out, remoteHits.Load())
+	}
+}
+
+// The cached admission answer can be one interval out of date, so a remote that
+// dies between probes must still not break the turn.
+func TestRemoteErrorMidRequestFallsBack(t *testing.T) {
+	var floorHits atomic.Int64
+	dead := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusInternalServerError)
+	}))
+	defer dead.Close()
+	floor := completionServer(t, "floor", &floorHits)
+	up := &atomic.Bool{}
+	up.Store(true)
+	health := healthServer(t, up)
+
+	p := NewPair(New(dead.URL, time.Second), New(floor.URL, time.Second), health.URL, time.Hour)
+	p.Start(context.Background())
+	defer p.Stop()
+	if !waitFor(t, p.Available) {
+		t.Fatal("prober never saw the remote come up")
+	}
+
+	out, err := p.Complete(context.Background(), Req{User: "привет"})
+	if err != nil {
+		t.Fatalf("complete: %v", err)
+	}
+	if out != "floor" || floorHits.Load() != 1 {
+		t.Fatalf("out = %q, floor hits = %d", out, floorHits.Load())
+	}
+	// The failed request must have corrected the cached answer, so the next
+	// one does not walk into the same hole.
+	if p.Available() {
+		t.Fatal("a failed remote request left the admission answer up")
+	}
+}
+
+// The naming half of the degradation rule. A world question must not be handed
+// to the resident model, because it answers by inventing.
+func TestCompleteRemoteNamesTheGap(t *testing.T) {
+	var floorHits atomic.Int64
+	floor := completionServer(t, "floor", &floorHits)
+	health := healthServer(t, &atomic.Bool{}) // never ok
+
+	p := NewPair(New("http://127.0.0.1:1", time.Second), New(floor.URL, time.Second), health.URL, 20*time.Millisecond)
+	p.Start(context.Background())
+	defer p.Stop()
+	time.Sleep(60 * time.Millisecond)
+
+	if _, err := p.CompleteRemote(context.Background(), Req{User: "почему небо голубое"}); !errors.Is(err, ErrRemoteUnavailable) {
+		t.Fatalf("err = %v, want ErrRemoteUnavailable", err)
+	}
+	if floorHits.Load() != 0 {
+		t.Fatalf("CompleteRemote fell back to the floor %d times", floorHits.Load())
+	}
+}
+
+// Routing sits on the hot path and must never pay for a health check. Available
+// reads a cached flag, so it costs no network at all.
+func TestAvailableDoesNotProbe(t *testing.T) {
+	var probes atomic.Int64
+	health := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		probes.Add(1)
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer health.Close()
+
+	p := NewPair(New("http://127.0.0.1:1", time.Second), New("http://127.0.0.1:1", time.Second), health.URL, time.Hour)
+	p.Start(context.Background())
+	defer p.Stop()
+	if !waitFor(t, p.Available) {
+		t.Fatal("prober never ran")
+	}
+
+	before := probes.Load()
+	for range 1000 {
+		p.Available()
+	}
+	if got := probes.Load(); got != before {
+		t.Fatalf("1000 Available calls made %d probes", got-before)
+	}
+}
+
+// A Pair with no floor is a configuration mistake, and it must say so rather
+// than silently having nowhere to degrade to.
+func TestNoFloorIsAnError(t *testing.T) {
+	p := NewPair(nil, nil, "", time.Second)
+	if _, err := p.Complete(context.Background(), Req{User: "привет"}); !errors.Is(err, ErrNoFloor) {
+		t.Fatalf("err = %v, want ErrNoFloor", err)
+	}
+}

internal/netaddr/netaddr.go

@@ -0,0 +1,277 @@
+// Package netaddr parses a daemon seam address and dials or binds it.
+//
+// Every seam between Maven's daemons used to be a unix socket with the
+// network hardcoded at the call site — two dials in internal/ipc, one listen,
+// and the same pair again in internal/worker. That is correct for co-located
+// daemons and it is the reason a module cannot live on another host. This
+// package moves the choice into the address string so a deploy picks the
+// transport, not a recompile:
+//
+//	/run/maven/stt.sock                   unix   (the default, unchanged)
+//	unix:///run/maven/stt.sock            unix   (explicit, same thing)
+//	tcp://workstation:9310?token=hunter2  tcp
+//
+// A scheme-less address is unix and behaves exactly as it did before this
+// package existed: same 0700 parent dir, same 0600 socket, same bytes on the
+// wire with no handshake in front of them.
+//
+// Over TCP the filesystem permission that authenticated the unix socket is
+// gone, and what crosses this seam is audio of the owner speaking and the
+// text of his turns. So a TCP seam carries a shared token, checked before the
+// first protocol frame is read. Wireguard is supported underneath and is not
+// required.
+package netaddr
+
+import (
+	"crypto/subtle"
+	"errors"
+	"fmt"
+	"net"
+	"net/url"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+// ErrUnauthorized — the peer presented a token the listener does not accept,
+// or presented none when one is required.
+var ErrUnauthorized = errors.New("netaddr: unauthorized")
+
+// Addr is a parsed seam endpoint.
+type Addr struct {
+	// Network is "unix" or "tcp".
+	Network string
+	// Address is the socket path (unix) or host:port (tcp).
+	Address string
+	// Token is the shared secret for a tcp seam. Empty for unix, where the
+	// filesystem does the same job.
+	Token string
+}
+
+// String renders the address for logs and errors. The token is never included.
+func (a Addr) String() string {
+	if a.Network == "unix" {
+		return a.Address
+	}
+	return a.Network + "://" + a.Address
+}
+
+// IsUnix reports whether this seam is a unix socket, and so is local, is
+// authenticated by file permissions, and needs no handshake.
+func (a Addr) IsUnix() bool { return a.Network == "unix" }
+
+// Parse reads a seam address. Anything without a "scheme://" prefix is a unix
+// socket path, which keeps every existing config and every default working
+// untouched.
+func Parse(s string) (Addr, error) {
+	if !strings.Contains(s, "://") {
+		return Addr{Network: "unix", Address: s}, nil
+	}
+	u, err := url.Parse(s)
+	if err != nil {
+		return Addr{}, fmt.Errorf("netaddr: parse %q: %w", s, err)
+	}
+	switch u.Scheme {
+	case "unix":
+		return Addr{Network: "unix", Address: u.Path}, nil
+	case "tcp":
+		if u.Host == "" {
+			return Addr{}, fmt.Errorf("netaddr: %q has no host:port", s)
+		}
+		return Addr{Network: "tcp", Address: u.Host, Token: u.Query().Get("token")}, nil
+	default:
+		return Addr{}, fmt.Errorf("netaddr: unsupported scheme %q", u.Scheme)
+	}
+}
+
+// MustParse is Parse for a literal known good at compile time. It panics on a
+// bad address, so use it in tests and constants, never on config input.
+func MustParse(s string) Addr {
+	a, err := Parse(s)
+	if err != nil {
+		panic(err)
+	}
+	return a
+}
+
+// handshakeTimeout bounds the token exchange. A peer that cannot write one
+// short line in this long is not going to serve a turn either.
+const handshakeTimeout = 5 * time.Second
+
+// greeting prefixes the token line. Versioned so a later mTLS seam can be
+// told apart from this one on the wire.
+const greeting = "MAVEN1 "
+
+// Dial connects to a. On a tcp seam it sends the token and waits for the
+// listener to accept it, so a returned conn is already authorized and the
+// caller can write its first protocol frame.
+func Dial(a Addr) (net.Conn, error) {
+	return DialTimeout(a, 0)
+}
+
+// DialTimeout is Dial with a bound on the connect. Zero means the operating
+// system default. The token exchange gets its own timeout either way.
+func DialTimeout(a Addr, timeout time.Duration) (net.Conn, error) {
+	var c net.Conn
+	var err error
+	if timeout > 0 {
+		c, err = net.DialTimeout(a.Network, a.Address, timeout)
+	} else {
+		c, err = net.Dial(a.Network, a.Address)
+	}
+	if err != nil {
+		return nil, err
+	}
+	if a.IsUnix() {
+		return c, nil
+	}
+	if err := clientHandshake(c, a.Token); err != nil {
+		_ = c.Close()
+		return nil, err
+	}
+	return c, nil
+}
+
+func clientHandshake(c net.Conn, token string) error {
+	_ = c.SetDeadline(time.Now().Add(handshakeTimeout))
+	defer c.SetDeadline(time.Time{})
+	if _, err := c.Write([]byte(greeting + token + "\n")); err != nil {
+		return fmt.Errorf("netaddr: send token: %w", err)
+	}
+	var reply [3]byte
+	if _, err := readFull(c, reply[:]); err != nil {
+		return fmt.Errorf("%w: %v", ErrUnauthorized, err)
+	}
+	if string(reply[:]) != "ok\n" {
+		return ErrUnauthorized
+	}
+	return nil
+}
+
+// Listener wraps a net.Listener so Accept performs the token check for a tcp
+// seam. A connection that fails the check is closed and never surfaces, so
+// the protocol above this layer only ever sees authorized peers.
+type Listener struct {
+	net.Listener
+	addr Addr
+}
+
+// Accept returns the next authorized connection. Unauthorized peers are
+// dropped and Accept keeps waiting: a bad token is a rejected stranger, not a
+// reason to stop serving.
+func (l *Listener) Accept() (net.Conn, error) {
+	for {
+		c, err := l.Listener.Accept()
+		if err != nil {
+			return nil, err
+		}
+		if l.addr.IsUnix() {
+			return c, nil
+		}
+		if err := serverHandshake(c, l.addr.Token); err != nil {
+			_ = c.Close()
+			continue
+		}
+		return c, nil
+	}
+}
+
+// Addr reports the parsed seam address this listener was built from.
+func (l *Listener) SeamAddr() Addr { return l.addr }
+
+func serverHandshake(c net.Conn, want string) error {
+	_ = c.SetDeadline(time.Now().Add(handshakeTimeout))
+	defer c.SetDeadline(time.Time{})
+	// The line is bounded: greeting, token, newline. Read a byte at a time so
+	// nothing of the first protocol frame is consumed when the token is short.
+	line := make([]byte, 0, 128)
+	var b [1]byte
+	for {
+		if _, err := readFull(c, b[:]); err != nil {
+			return err
+		}
+		if b[0] == '\n' {
+			break
+		}
+		line = append(line, b[0])
+		if len(line) > 512 {
+			return ErrUnauthorized
+		}
+	}
+	got, ok := strings.CutPrefix(string(line), greeting)
+	if !ok {
+		return ErrUnauthorized
+	}
+	if subtle.ConstantTimeCompare([]byte(got), []byte(want)) != 1 {
+		return ErrUnauthorized
+	}
+	if _, err := c.Write([]byte("ok\n")); err != nil {
+		return err
+	}
+	return nil
+}
+
+func readFull(c net.Conn, p []byte) (int, error) {
+	n := 0
+	for n < len(p) {
+		m, err := c.Read(p[n:])
+		n += m
+		if err != nil {
+			return n, err
+		}
+	}
+	return n, nil
+}
+
+// Listen binds a. A unix seam gets the perms it has always had: parent dir
+// 0700, socket 0600, and any stale socket from a crashed daemon removed
+// first. A tcp seam must carry a token, because there is no filesystem to
+// stand in for one.
+func Listen(a Addr) (*Listener, error) {
+	if a.IsUnix() {
+		ln, err := listenUnix(a.Address)
+		if err != nil {
+			return nil, err
+		}
+		return &Listener{Listener: ln, addr: a}, nil
+	}
+	if a.Token == "" {
+		return nil, fmt.Errorf("netaddr: listen %s: tcp seam requires a token", a)
+	}
+	ln, err := net.Listen("tcp", a.Address)
+	if err != nil {
+		return nil, fmt.Errorf("netaddr: listen %s: %w", a, err)
+	}
+	return &Listener{Listener: ln, addr: a}, nil
+}
+
+func listenUnix(path string) (net.Listener, error) {
+	_ = os.Remove(path) // stale socket from a crashed daemon; ignore missing
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		return nil, fmt.Errorf("netaddr: mkdir socket dir: %w", err)
+	}
+	// umask could widen the perms on socket creation; tighten then chmod to
+	// be explicit. 0600 ⇒ read+write by owner only.
+	oldMask := unix.Umask(0o077)
+	ln, err := net.Listen("unix", path)
+	unix.Umask(oldMask)
+	if err != nil {
+		return nil, fmt.Errorf("netaddr: listen %s: %w", path, err)
+	}
+	if err := os.Chmod(path, 0o600); err != nil {
+		_ = ln.Close()
+		_ = os.Remove(path)
+		return nil, fmt.Errorf("netaddr: chmod socket: %w", err)
+	}
+	return ln, nil
+}
+
+// Cleanup removes the socket file behind a unix seam. It is a no-op for tcp.
+func Cleanup(a Addr) {
+	if a.IsUnix() && a.Address != "" {
+		_ = os.Remove(a.Address)
+	}
+}

internal/netaddr/netaddr_test.go

@@ -0,0 +1,185 @@
+package netaddr
+
+import (
+	"errors"
+	"net"
+	"path/filepath"
+	"testing"
+)
+
+// A scheme-less address must stay unix. Every deploy in the tree writes a bare
+// path, so this is the test that says the transport change costs them nothing.
+func TestParseSchemelessIsUnix(t *testing.T) {
+	a, err := Parse("/run/maven/stt.sock")
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if !a.IsUnix() {
+		t.Fatalf("want unix, got %q", a.Network)
+	}
+	if a.Address != "/run/maven/stt.sock" {
+		t.Fatalf("address = %q", a.Address)
+	}
+	if a.Token != "" {
+		t.Fatalf("unix seam carries a token: %q", a.Token)
+	}
+}
+
+func TestParse(t *testing.T) {
+	cases := []struct {
+		in            string
+		net, addr, tk string
+		wantErr       bool
+	}{
+		{in: "", net: "unix", addr: ""},
+		{in: "unix:///run/maven/core.sock", net: "unix", addr: "/run/maven/core.sock"},
+		{in: "tcp://workstation:9310", net: "tcp", addr: "workstation:9310"},
+		{in: "tcp://workstation:9310?token=hunter2", net: "tcp", addr: "workstation:9310", tk: "hunter2"},
+		{in: "tcp://", wantErr: true},
+		{in: "udp://workstation:9310", wantErr: true},
+	}
+	for _, c := range cases {
+		a, err := Parse(c.in)
+		if c.wantErr {
+			if err == nil {
+				t.Errorf("Parse(%q) = %v, want error", c.in, a)
+			}
+			continue
+		}
+		if err != nil {
+			t.Errorf("Parse(%q): %v", c.in, err)
+			continue
+		}
+		if a.Network != c.net || a.Address != c.addr || a.Token != c.tk {
+			t.Errorf("Parse(%q) = %+v, want %s/%s/%s", c.in, a, c.net, c.addr, c.tk)
+		}
+	}
+}
+
+// The token must never reach a log line.
+func TestStringHidesToken(t *testing.T) {
+	a := MustParse("tcp://workstation:9310?token=hunter2")
+	if got := a.String(); got != "tcp://workstation:9310" {
+		t.Fatalf("String() = %q", got)
+	}
+}
+
+// A unix seam must round-trip with no handshake in front of the payload: the
+// first bytes the listener sees are the caller's, exactly as before.
+func TestUnixRoundTripHasNoHandshake(t *testing.T) {
+	a := MustParse(filepath.Join(t.TempDir(), "s.sock"))
+	ln, err := Listen(a)
+	if err != nil {
+		t.Fatalf("listen: %v", err)
+	}
+	defer ln.Close()
+	go echoOnce(ln)
+
+	c, err := Dial(a)
+	if err != nil {
+		t.Fatalf("dial: %v", err)
+	}
+	defer c.Close()
+	if got := roundTrip(t, c, "hello"); got != "hello" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestTCPRoundTripWithToken(t *testing.T) {
+	ln, addr := listenLoopback(t, "s3cret")
+	defer ln.Close()
+	go echoOnce(ln)
+
+	c, err := Dial(addr)
+	if err != nil {
+		t.Fatalf("dial: %v", err)
+	}
+	defer c.Close()
+	if got := roundTrip(t, c, "hello"); got != "hello" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestTCPWrongTokenIsRejected(t *testing.T) {
+	ln, addr := listenLoopback(t, "s3cret")
+	defer ln.Close()
+	// Accept keeps waiting past the bad peer, so nothing here should ever
+	// reach the echo. A conn that does means the token was not checked.
+	go echoOnce(ln)
+
+	bad := addr
+	bad.Token = "wrong"
+	if _, err := Dial(bad); !errors.Is(err, ErrUnauthorized) {
+		t.Fatalf("dial with wrong token: err = %v, want ErrUnauthorized", err)
+	}
+}
+
+// A stranger that speaks the protocol instead of the greeting is dropped, and
+// the listener stays up for the peer that follows it.
+func TestTCPUngreetedPeerDoesNotKillTheListener(t *testing.T) {
+	ln, addr := listenLoopback(t, "s3cret")
+	defer ln.Close()
+	go echoOnce(ln)
+
+	raw, err := net.Dial("tcp", addr.Address)
+	if err != nil {
+		t.Fatalf("raw dial: %v", err)
+	}
+	if _, err := raw.Write([]byte("GET / HTTP/1.1\n")); err != nil {
+		t.Fatalf("raw write: %v", err)
+	}
+	raw.Close()
+
+	c, err := Dial(addr)
+	if err != nil {
+		t.Fatalf("dial after stranger: %v", err)
+	}
+	defer c.Close()
+	if got := roundTrip(t, c, "still here"); got != "still here" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+// A tcp seam with no token is a misconfiguration, and it must fail at bind
+// rather than serve the owner's turns to anyone who connects.
+func TestTCPListenRequiresToken(t *testing.T) {
+	if _, err := Listen(MustParse("tcp://127.0.0.1:0")); err == nil {
+		t.Fatal("listen on a tokenless tcp seam succeeded")
+	}
+}
+
+func listenLoopback(t *testing.T, token string) (*Listener, Addr) {
+	t.Helper()
+	ln, err := Listen(Addr{Network: "tcp", Address: "127.0.0.1:0", Token: token})
+	if err != nil {
+		t.Fatalf("listen: %v", err)
+	}
+	return ln, Addr{Network: "tcp", Address: ln.Addr().String(), Token: token}
+}
+
+func echoOnce(ln *Listener) {
+	c, err := ln.Accept()
+	if err != nil {
+		return
+	}
+	defer c.Close()
+	buf := make([]byte, 256)
+	n, err := c.Read(buf)
+	if err != nil {
+		return
+	}
+	_, _ = c.Write(buf[:n])
+}
+
+func roundTrip(t *testing.T, c net.Conn, msg string) string {
+	t.Helper()
+	if _, err := c.Write([]byte(msg)); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	buf := make([]byte, 256)
+	n, err := c.Read(buf)
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	return string(buf[:n])
+}

internal/phraser/llmphraser.go

@@ -534,8 +534,14 @@ func (p *LLMPhraser) PhraseReminder(ctx context.Context, d loop.ReminderDecision
 		text = "reminder"
 	}
 
+	// Russian, like the other two prompts (Vikunja #404). Asking a model for a
+	// Russian reply in English is asking it to switch languages mid-prompt,
+	// and a 1.7B sometimes answers in the language it was asked in. The
+	// persona rules and the JSON contract are not repeated here: this call
+	// goes through chat(), so nudgeSystem already states both, and a second
+	// statement of the same contract is one more thing that can drift.
 	prompt := fmt.Sprintf(
-		`The user set a reminder: "%s". Rephrase it briefly as a gentle nudge. Respond as JSON: {"response": "...", "mood": "..."}`,
+		`Он поставил напоминание: "%s". Скажи это своими словами, коротко и мягко — одно предложение.`,
 		text,
 	)
 	resp, err := p.chat(ctx, prompt)
@@ -754,7 +760,7 @@ func (p *LLMPhraser) querySystemPrompt() string {
 	base := "Ты отвечаешь ему по источникам, которые тебе дали. Отвечай ТОЛЬКО по ним: всё, что ты говоришь, должно быть написано в источниках. " +
 		"Если ответа в них нет — так и скажи и на этом остановись; не добавляй ничего из своих знаний и не догадывайся. " +
 		"Не приплетай прошлые реплики разговора. " +
-		"Отвечай по-русски, коротко и своими словами, начинай с \"вот что я нашла: \". О себе — в женском роде, глаголы в прошедшем времени с окончанием -ла. Он мужчина, обращайся к нему на \"ты\". Respond ONLY with valid JSON: {\"response\": \"...\", \"mood\": \"neutral\"}."
+		"Отвечай по-русски, коротко и своими словами, начинай с \"вот что я нашла: \". О себе — в женском роде, глаголы в прошедшем времени с окончанием -ла. Он мужчина, обращайся к нему на \"ты\". Отвечай ТОЛЬКО одним объектом JSON: {\"response\": \"...\", \"mood\": \"neutral\"}."
 	return persona.Prepend(p.cfg.ContextBlock, base)
 }
 

internal/phraser/spawn_test.go

@@ -0,0 +1,256 @@
+package phraser
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"strings"
+	"syscall"
+	"testing"
+	"time"
+)
+
+// The spawn path (NewLLMPhraser, spawnLlamaServer, startLlamaProc, llamaProc.Close)
+// was at 0% coverage: every test built the phraser with NewLLMPhraserAt, which
+// starts no process. These tests drive the real spawn code against a fake
+// llama-server script, so the startup race arms and the reaping are exercised
+// without a model or a GPU.
+
+// fakeLlama writes an executable script standing in for llama-server and returns
+// its path. body runs after the script has recorded its own pid.
+func fakeLlama(t *testing.T, body string) string {
+	t.Helper()
+	dir := t.TempDir()
+	path := filepath.Join(dir, "fake-llama-server")
+	script := "#!/bin/sh\n" + body + "\n"
+	if err := os.WriteFile(path, []byte(script), 0o755); err != nil {
+		t.Fatalf("write fake server: %v", err)
+	}
+	return path
+}
+
+// listensThenSleeps prints the line startLlamaProc scrapes, then stays alive
+// until killed — the shape of a real llama-server that came up.
+const listensThenSleeps = `echo "srv    load_model: listening on http://127.0.0.1:18081" >&2
+while : ; do sleep 1 ; done`
+
+func testCfg(bin string) Config {
+	cfg := DefaultConfig("/nonexistent/model.gguf")
+	cfg.BinPath = bin
+	return cfg
+}
+
+func TestExtractPort(t *testing.T) {
+	for _, tc := range []struct{ in, want string }{
+		{"127.0.0.1:0", "0"},
+		{"127.0.0.1:8080", "8080"},
+		{"127.0.0.1:", "0"},
+		{"", "0"},
+		{"8080", "0"}, // no colon: Cut yields no port, so the caller gets the "any port" default
+	} {
+		if got := extractPort(tc.in); got != tc.want {
+			t.Errorf("extractPort(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+func TestStartLlamaProcScrapesPortAndReaps(t *testing.T) {
+	bin := fakeLlama(t, listensThenSleeps)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	p, err := startLlamaProc(ctx, testCfg(bin))
+	if err != nil {
+		t.Fatalf("startLlamaProc: %v", err)
+	}
+	if p.BaseURL() != "http://127.0.0.1:18081" {
+		t.Fatalf("BaseURL = %q, want the scraped address", p.BaseURL())
+	}
+	pid := p.cmd.Process.Pid
+
+	p.cancel = cancel
+	if err := p.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	// Close must Wait, otherwise the child lingers as a zombie.
+	if p.cmd.ProcessState == nil {
+		t.Fatal("Close did not reap the child: ProcessState is nil")
+	}
+	if err := syscall.Kill(pid, 0); err == nil {
+		t.Fatalf("child %d still exists after Close", pid)
+	}
+}
+
+func TestStartLlamaProcFailureArms(t *testing.T) {
+	t.Run("binary missing", func(t *testing.T) {
+		cfg := testCfg(filepath.Join(t.TempDir(), "does-not-exist"))
+		_, err := startLlamaProc(context.Background(), cfg)
+		if err == nil || !strings.Contains(err.Error(), "llm: start") {
+			t.Fatalf("err = %v, want a start failure", err)
+		}
+	})
+
+	t.Run("server exits without listening", func(t *testing.T) {
+		// stderr closes, so the reader goroutine reports EOF on errCh.
+		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)
+		}
+	})
+
+	t.Run("context cancelled during startup", func(t *testing.T) {
+		// Never prints the listen line and never exits: only ctx can end this.
+		bin := fakeLlama(t, `while : ; do sleep 1 ; done`)
+		ctx, cancel := context.WithCancel(context.Background())
+		go func() {
+			time.Sleep(150 * time.Millisecond)
+			cancel()
+		}()
+		defer cancel()
+		_, err := startLlamaProc(ctx, testCfg(bin))
+		if !errors.Is(err, context.Canceled) {
+			t.Fatalf("err = %v, want context.Canceled", err)
+		}
+	})
+}
+
+func TestNewLLMPhraserSpawns(t *testing.T) {
+	bin := fakeLlama(t, listensThenSleeps)
+	p, err := NewLLMPhraser(context.Background(), testCfg(bin))
+	if err != nil {
+		t.Fatalf("NewLLMPhraser: %v", err)
+	}
+	if p.BaseURL() != "http://127.0.0.1:18081" {
+		t.Fatalf("BaseURL = %q", p.BaseURL())
+	}
+	pid := p.be.(*llamaProc).cmd.Process.Pid
+	if err := p.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	if p.BaseURL() != "" {
+		t.Fatalf("BaseURL after Close = %q, want empty", p.BaseURL())
+	}
+	if err := syscall.Kill(pid, 0); err == nil {
+		t.Fatalf("llama-server %d survived Close", pid)
+	}
+}
+
+func TestNewLLMPhraserSpawnFailure(t *testing.T) {
+	cfg := testCfg(filepath.Join(t.TempDir(), "does-not-exist"))
+	p, err := NewLLMPhraser(context.Background(), cfg)
+	if err == nil {
+		p.Close()
+		t.Fatal("want an error when the server cannot start")
+	}
+	if p != nil {
+		t.Fatalf("want a nil phraser on failure, got %#v", p)
+	}
+}
+
+// TestPdeathsigKillsOrphan is the orphan test the task asked for. A SIGKILLed
+// mavend never runs Close, so nothing but the kernel's Pdeathsig can stop its
+// llama-server. Re-exec this test binary as the "daemon", let it spawn the fake
+// server, SIGKILL the daemon, and assert the grandchild died with it.
+func TestPdeathsigKillsOrphan(t *testing.T) {
+	bin := fakeLlama(t, listensThenSleeps)
+
+	cmd := exec.Command(os.Args[0], "-test.run=TestSpawnHelperProcess", "-test.v=false")
+	cmd.Env = append(os.Environ(), "MAVEN_SPAWN_HELPER=1", "MAVEN_FAKE_LLAMA="+bin)
+	out, err := cmd.StdoutPipe()
+	if err != nil {
+		t.Fatalf("stdout pipe: %v", err)
+	}
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start helper: %v", err)
+	}
+	defer func() { _ = cmd.Process.Kill(); _ = cmd.Wait() }()
+
+	buf := make([]byte, 256)
+	n, err := out.Read(buf)
+	if err != nil {
+		t.Fatalf("read child pid: %v", err)
+	}
+	childPID, err := strconv.Atoi(strings.TrimSpace(string(buf[:n])))
+	if err != nil {
+		t.Fatalf("helper printed %q, want a pid: %v", string(buf[:n]), err)
+	}
+	if err := syscall.Kill(childPID, 0); err != nil {
+		t.Fatalf("llama-server %d not running before the kill: %v", childPID, err)
+	}
+
+	// SIGKILL: the helper gets no chance to clean up, exactly like an OOM kill.
+	if err := cmd.Process.Signal(syscall.SIGKILL); err != nil {
+		t.Fatalf("kill helper: %v", err)
+	}
+	_, _ = cmd.Process.Wait()
+
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		if err := syscall.Kill(childPID, 0); err != nil {
+			return // gone: Pdeathsig did its job
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	_ = syscall.Kill(childPID, syscall.SIGKILL)
+	t.Fatalf("llama-server %d outlived the SIGKILLed parent", childPID)
+}
+
+// TestSpawnHelperProcess is not a test. It is the child half of
+// TestPdeathsigKillsOrphan: spawn a llama-server, print its pid, then block.
+func TestSpawnHelperProcess(t *testing.T) {
+	if os.Getenv("MAVEN_SPAWN_HELPER") != "1" {
+		t.Skip("helper for TestPdeathsigKillsOrphan")
+	}
+	cfg := testCfg(os.Getenv("MAVEN_FAKE_LLAMA"))
+	p, err := startLlamaProc(context.Background(), cfg)
+	if err != nil {
+		fmt.Println("spawn failed:", err)
+		os.Exit(1)
+	}
+	fmt.Println(p.cmd.Process.Pid)
+	os.Stdout.Sync()
+	select {} // wait to be killed
+}
+
+// TestKillMavenScriptMatchesRealCommandLine pins kill-maven.sh's fallback
+// pattern to the command line startLlamaProc actually builds. The script leaked
+// orphans twice already, both times because the pattern stopped matching: first
+// `llama-server.*maven`, then a hardcoded model name after the model was swapped.
+func TestKillMavenScriptMatchesRealCommandLine(t *testing.T) {
+	src, err := os.ReadFile("../../kill-maven.sh")
+	if err != nil {
+		t.Fatalf("read kill-maven.sh: %v", err)
+	}
+	m := regexp.MustCompile(`(?m)^\s*LLM='([^']+)'`).FindSubmatch(src)
+	if m == nil {
+		t.Fatal("no default LLM='...' pattern in kill-maven.sh")
+	}
+	pat, err := regexp.Compile(string(m[1]))
+	if err != nil {
+		t.Fatalf("LLM pattern %q does not compile: %v", m[1], err)
+	}
+
+	// Rebuild the command line from the production arg list, so a change to
+	// 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",
+	}, " ")
+	if !pat.MatchString(cmdline) {
+		t.Fatalf("kill-maven.sh pattern %q does not match %q — orphans would leak", m[1], cmdline)
+	}
+}

internal/router/knowledge.go

@@ -6,5 +6,5 @@ func KnowledgePrompt() string {
 	// No self-introduction here: the shared persona block already says who she
 	// is, and this line used to disagree with it — a different name ("Мавена")
 	// and a masculine noun ("ассистент") in front of a feminine persona.
-	return `Ответь кратко из своих знаний. Если не знаешь — скажи "не знаю". Не выдумывай. Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}.`
+	return `Ответь кратко из своих знаний. Если не знаешь — скажи "не знаю". Не выдумывай. Отвечай ТОЛЬКО одним объектом JSON: {"response": "...", "mood": "neutral"}.`
 }

internal/worker/client.go

@@ -24,6 +24,8 @@ import (
 	"net"
 	"sync"
 	"time"
+
+	"github.com/kami/maven/internal/netaddr"
 )
 
 // Client — one connection to one worker module. NOT goroutine-safe for
@@ -38,15 +40,25 @@ type Client struct {
 	dial func() (net.Conn, error)
 }
 
-// Dial opens a Client to the worker socket at path. The first call lazily
+// Dial opens a Client to the worker module at path. The first call lazily
 // dials; subsequent calls reuse the conn (a fresh dial happens on next call
 // after a teardown). Lazy dial keeps a worker that's restarting from
 // blocking core's startup; core attempts the dial on first use.
+//
+// path is a netaddr seam address. A bare path is the unix socket it has
+// always been; "tcp://workstation:9310?token=..." reaches a module on another
+// host, which is how stt and tts move to the machine with the GPU and the
+// microphone. A bad address surfaces on the first call, not here, because
+// Dial does not fail — see internal/netaddr.
 func Dial(path string) *Client {
+	addr, err := netaddr.Parse(path)
 	return &Client{
 		path: path,
 		dial: func() (net.Conn, error) {
-			return net.Dial("unix", path)
+			if err != nil {
+				return nil, err
+			}
+			return netaddr.Dial(addr)
 		},
 	}
 }

internal/worker/server.go

@@ -7,10 +7,11 @@
 // which module to dial; mixing the two is a config error caught cleanly by
 // the wire, not a runtime goroutine panic). One Server per module process.
 //
-// Socket perms mirror ipc.Server: dir 0700, socket 0600 ⇒ same unix user.
-// The module has no key, so the floor is "same user"; the wg/mTLS layers
-// are out of scope here (this socket never crosses the network radius —
-// it's local-only, point-to-point between two processes on the box).
+// The seam address decides the transport. On the default unix socket the
+// perms mirror ipc.Server — dir 0700, socket 0600 ⇒ same unix user — and that
+// is the whole auth floor, because the seam never leaves the box. A tcp
+// address moves the module to another host and takes that floor away, so
+// netaddr checks a shared token before the first frame. See internal/netaddr.
 package worker
 
 import (
@@ -18,11 +19,10 @@ import (
 	"encoding/json"
 	"fmt"
 	"net"
-	"os"
 	"sync"
 	"sync/atomic"
 
-	"golang.org/x/sys/unix"
+	"github.com/kami/maven/internal/netaddr"
 )
 
 // Server — a worker module process's listener. Wires either a Transcriber,
@@ -34,6 +34,7 @@ type Server struct {
 	s Synthesizer
 
 	path string
+	addr netaddr.Addr
 	ln   net.Listener
 
 	wg   sync.WaitGroup
@@ -61,25 +62,24 @@ func NewSynthesizerServer(path string, s Synthesizer) *Server {
 // two separate processes per the restart-free / fail-independent invariant).
 func (srv *Server) SetSynthesizer(s Synthesizer) { srv.s = s }
 
-// Listen binds the unix socket with 0700 dir + 0600 socket perms (same floor
-// as internal/ipc). A stale socket at path is removed first so the worker
-// process restarts cleanly after a crash, no manual cleanup needed.
+// Listen binds the seam address the Server was built with.
+//
+// A bare path is a unix socket with 0700 dir + 0600 socket perms, the same
+// floor as internal/ipc, and a stale socket is removed first so the worker
+// process restarts cleanly after a crash. A "tcp://host:port?token=..."
+// address binds a network listener instead, so this module can run on the
+// workstation while core stays on homesrv; the token is mandatory there,
+// because there is no filesystem to be the auth floor. See internal/netaddr.
 func (srv *Server) Listen() error {
-	_ = os.Remove(srv.path)
-	if err := os.MkdirAll(parentDir(srv.path), 0o700); err != nil {
-		return fmt.Errorf("worker: mkdir socket dir: %w", err)
-	}
-	oldMask := unix.Umask(0o077)
-	ln, err := net.Listen("unix", srv.path)
-	unix.Umask(oldMask)
+	addr, err := netaddr.Parse(srv.path)
 	if err != nil {
-		return fmt.Errorf("worker: listen %s: %w", srv.path, err)
+		return err
 	}
-	if err := os.Chmod(srv.path, 0o600); err != nil {
-		_ = ln.Close()
-		_ = os.Remove(srv.path)
-		return fmt.Errorf("worker: chmod socket: %w", err)
+	ln, err := netaddr.Listen(addr)
+	if err != nil {
+		return err
 	}
+	srv.addr = addr
 	srv.ln = ln
 	return nil
 }
@@ -193,7 +193,7 @@ func (srv *Server) Close() error {
 	}
 	err := srv.ln.Close()
 	srv.wg.Wait()
-	_ = os.Remove(srv.path)
+	netaddr.Cleanup(srv.addr)
 	return err
 }
 
@@ -214,15 +214,3 @@ func marshalResult(v any) json.RawMessage {
 	b, _ := json.Marshal(v)
 	return b
 }
-
-func parentDir(p string) string {
-	for i := len(p) - 1; i >= 0; i-- {
-		if p[i] == '/' {
-			if i == 0 {
-				return "/"
-			}
-			return p[:i]
-		}
-	}
-	return "."
-}