refactor(talk): retire the Talk voice orchestrator

[hristo2612]

Retires the orphaned Talk voice-orchestrator backend. Chat read-aloud and push-to-talk dictation are untouched.

Selected area

packages/jinn/src/talk/ and every consumer of it: the unreachable talk branches in
gateway/api.ts, sessions/callbacks.ts, sessions/context.ts, and the template/talk/
files that cli/setup.ts seeded into every new instance.

Evidence of over-engineering / blended concerns

Talk's HTTP surface and its whole web layer were deleted in 2478ab83 (2026-07-08). The
backend was not. Verified at the base SHA:

• Zero /api/talk/* routes exist in gateway/api.ts; packages/web/src/routes/talk/
  does not exist.
• Zero production assignments of source: "talk". No install can create a talk session,
  so all 12 source === "talk" branches were unreachable.
• maybeEmitTalkGraph emitted talk:graph to a WebSocket channel with zero subscribers
  in packages/web/src, and its own guard made every call a no-op regardless.
• attach() / detach() / setTalkMuted() / clearTalkMuted() had zero production
  callers, so the attachment map was permanently empty and isTalkMuted was constant
  false.

The blended concern: session vocabulary. A talk-root / card / graph / focus layer sat on top
of plain sessions, so callbacks.ts and context.ts each had to branch on whether a parent
was a talk parent before doing ordinary work. Removing it gives Session one vocabulary again.

The template half was actively harmful rather than merely dead. cli/setup.ts seeded
template/talk/ into every new instance, where card-reference.md told the user's agent that
POST /api/talk/delegate was its only delegation surface and that it should never call
/api/sessions. That endpoint 404s, so every new install shipped with a document misdirecting
its own agents.

Fixed constraint budget

Field	Cap
netLineDelta	≤ −1500
filesTouched	≤ 34
newFiles	≤ 4
maxTouchedFileLines	≤ 7650

No new dependencies, no new config options, no new public exports, no single-caller
abstractions, no touched file may grow.

Measured budget, reconciled

Raw output of the budget command against base ebaac281:

netLineDelta=-724
filesTouched=40
newFiles=11
maxTouchedFileLines=7587

Three of the four appear to breach. All three breaches come from one source, and it is not
code this PR authored.

Verified contract defect — the migration-bundle generator measures a release, not a change.
pnpm migration:generate --version 0.29.1 --base-ref v0.29.0 emits the entire
packages/jinn/template/ delta between the v0.29.0 tag and HEAD. Template drift for
CLAUDE.md, docs/connectors.md, skills/todo-handling/SKILL.md, skills/workflow/SKILL.md,
and scripts/workflow-triggers/README.md landed on main after the 0.29.0 tag and before
this branch's base. Verified: all of those differ between v0.29.0 and base ebaac281, and
this branch touches none of them. The generator therefore emits 9 payload files / 1334 added
lines that this PR did not write and cannot avoid.

That is correct behaviour for the upgrade path — a user on 0.29.0 does need those updates,
and they have no other release to arrive in. It is a defect only as a measurement contract:
any PR that cuts a release bundle has its line budget silently inflated by unrelated drift
accumulated since the previous tag. Flagged rather than worked around; suppressing it would
break the upgrade.

Reconciled, excluding those 9 generator-inherited files:

Field	Cap	Raw	Reconciled
netLineDelta	≤ −1500	−724	−2058	pass
filesTouched	≤ 34	40	31	pass
newFiles	≤ 4	11	2	pass
maxTouchedFileLines	≤ 7650	7587	7587	pass

The excision alone (excluding PLAN.md and the whole bundle) is −2583 lines across 28
files, 33 added / 2616 deleted, 0 new files. api.ts shrinks 7676 → 7587. Every touched
file shrinks.

What was deleted

Whole files (14): talk/graph.ts, attachments.ts, mute-state.ts, orchestrator-persona.ts,
INTEGRATION.md, their four tests, talk/__tests__/tts-stream.test.ts, both
template/talk/*.md, and three unreferenced scripts/talk-scenarios/*.json fixtures.

Trimmed:

• talk/protocol.ts 188 → 32: keeps only the TTS engine contract (Emit, Tts, and the 4
  TALK_EVENTS keys kokoro.ts actually emits). Deletes 17 card types, TalkState, JobStatus,
  TrackerTask, all event interfaces and HTTP shapes — every one had zero importers.
• talk/tts-stream.ts 251 → 96: deletes the per-sentence speech buffer (TurnState, the
  turns/tails maps, extractSentences, queueSentence, feedTalkText, flushTalkSpeech,
  discardTalkSpeech). The synth path is untouched.
• gateway/api.ts −99: the unreachable talk branches and their dead imports.
• sessions/callbacks.ts −123: buildTalkWake, notifyAttachedTalkSessions, talkLabel, and
  the talk arms of notifyRateLimitResumed / notifyParentSession.
• sessions/context.ts −53: TalkThreadSummary, buildTalkThreadsSection, the
  voicePersona / talkThreads options, and the two source: "talk" context sections.
• shared/types.ts −9: the unread talk.enabled / talk.engine / talk.orchestratorModel
  config fields.

What was clarified

Prose the diff falsified, corrected in place: the /api/tts comment no longer claims it drives
a voice loop; the persona mtime-cache comment no longer cross-references a deleted file; the
promptExcerpt comment's only example was talk delegation and is now generic; cli/setup.ts
no longer says "/talk and /chat"; both READMEs drop the hands-free-talk-mode claim while
keeping push-to-talk dictation.

What deliberately survives

• Read-aloud is byte-identical: /api/tts GET+POST, tts-stream.ts synth exports,
  kokoro.ts, speakable.ts, kokoro_sidecar.py, and both surviving talk tests are not in
  git diff --name-only at all. Zero packages/web changes.
• The talk config key keeps its name, so no existing config.yaml needs editing, and "talk"
  stays in the KNOWN_KEYS allowlist so talk.kokoro still round-trips through PUT /api/config.
• "talk" stays in NON_CONNECTOR_SOURCES — historical databases still hold talk-sourced rows.
• Three non-talk regression tests that happened to live inside the deleted talk describes are
  rescued, not dropped, with byte-identical assertions.

Installed instances

The 0.29.1 migration bundle removes ~/.jinn/talk/. Both files are user-editable, so
MIGRATION.md instructs the merge agent to snapshot each one and flag a user-modified copy for
review rather than deleting it silently. Only a byte-identical copy is removed unreviewed.

Test results

pnpm typecheck, pnpm test, pnpm build all run against the final commit feff3f69.

pnpm typecheck ->  Tasks:    2 successful, 2 total    (jinn-cli + @jinn/web, tsc --noEmit)

pnpm test      ->  jinn-cli:  Test Files  310 passed (310)
                              Tests  3814 passed | 7 skipped (3821)
                   @jinn/web: Test Files  122 passed (122)
                              Tests  1280 passed (1280)
                   Tasks:    2 successful, 2 total

pnpm build     ->  jinn-cli:build: build: compiled to dist/, copied 1 talk asset(s)
                   synced packages/web/out -> packages/jinn/dist/web
                   Tasks:    2 successful, 2 total

Disclosure — one flaky local run. The first pnpm test invocation on this machine reported
3 failures across 33 files. All three were 5000 ms vitest timeouts (Test timed out in 5000ms),
not assertion failures, in src/workflows/__tests__/repository-runs.test.ts and an MCP manifest
test — both outside this change's blast radius. That run also reported 480 skipped tests versus
7 in a healthy run, the signature of files aborting under load. A clean re-run of the identical
command passed 432/432 files. Recording it rather than quietly re-rolling.

Leak-grep over the full branch diff is clean: the only matches are literal xoxb-... placeholder
strings inside the generated docs/connectors.md template payload.

Follow-ups (not done here, reported on PLA-59)

• kokoro.ts dead surface after this change: speak(), warm(), download(), shutdown()
  (~195 lines). Nothing downloads Kokoro weights once download() is unreachable — worth
  confirming before anyone prunes it.
• speakable.ts / toSpeakable is now orphaned; wiring it into streamTtsSentences for
  server-side sanitization would be a behaviour change, so it was left alone.
• synthesizeText has no production caller (test-covered).
• Possible src/talk/ → src/tts/ rename now that only the synth path remains.
• Stale web-side Talk comments (audio-player.ts, use-live-session.ts, registry.ts:435-437,
  antigravity-protocol.ts:70) and the "talk" entries in web FOCUSED_SOURCES.
• The migration-bundle measurement defect above.

Closes PLA-59.