test(chat-routes): prove the autonomous lane's final write is authoritative

[drewstone]

What

Mutation testing over the incremental-persistence suite (#250) found one invariant no test could see: replacing runDetachedTurn's draft.finalize(values) with draft.close() — dropping the completion-time settlement entirely — left all 11 tests green.

The existing autonomous-lane test hides it by coincidence. At intervalMs: 0 over a turn whose tools all complete, the last draft happens to equal the final projection, so a lane that never settled authoritative values still produced a correct-looking row.

This test removes both coincidences:

• the shipped 2 s cadence (only the first draft escapes the floor, so the last draft is stale), and
• a tool that never completes (only finalizeAssistantParts terminalizes it; draftParts deliberately leaves it running).

Proof the check can fail

With await draft.finalize(values) -> await draft.close() in detached-turn.ts:

FAIL tests/chat-routes/incremental-persistence.test.ts > incremental assistant persistence
  > crash-safe convergence > autonomous lane: the FINAL write is authoritative over the last
    draft (stale cadence + dangling tool)
→ expected 'Looking' to be 'Looking, then gave up.' // Object.is equality

Restored: 12 passed (12).

Full mutation table

#	invariant broken	before this PR	after
M1	coalescing interval floor removed	KILLED	KILLED
M2	row id made non-deterministic	KILLED	KILLED
M3	finalize() skips the authoritative write	KILLED	KILLED
M4	discard() made a no-op	KILLED	KILLED
M5	drafts bypass transformFinalText	KILLED	KILLED
M6	detached lane never settles the row	SURVIVED	KILLED
M7	fail-loud opt-in guard disabled	KILLED	KILLED
M8	retry discriminator always 'idle'	KILLED	KILLED

Verification

• pnpm typecheck — clean
• pnpm test — 3171 passed / 1 skipped (195 files)
• agent-docs --check — docs are fresh
• Live proof against a real gemini-2.5-pro sandbox turn (76 raw events): durable row grew 28 -> 244 -> 783 -> 2136 chars mid-run at the shipped 2 s cadence, final row 3134 chars byte-identical to the single-write path, and a crash at 1009 chars re-entered onto the same row id with no duplicated parts.

[tangletools]

✅ Auto-approved drewstone PR — c3b6714b

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-25T02:43:57Z}

[tangletools]

✅ No Blockers — c3b6714b

Review health 100/100 · Reviewer score 95/100 · Confidence 65/100 · 0 findings (none)

glm: Correctness 95 · Security 95 · Testing 95 · Architecture 95

Reviewer score is advisory once the run is complete and the verdict has no blockers.

Full multi-shot audit completed 1/1 planned shots over 1 changed files. Global verifier still owns final merge decision.

No findings.

---

_{tangletools · 2026-07-25T02:49:03Z · trace}

[tangletools]

✅ Clean — c3b6714b

Full multi-shot audit completed 1/1 planned shots over 1 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

---

_{tangletools · 2026-07-25T02:49:03Z · immutable trace}

[tangletools]

🟢 Value Audit — sound

Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	184.5s (2 bridge agents)
Total	184.5s

💰 Value — sound

Adds one test that closes a verified mutation gap in runDetachedTurn's completion-time settlement — reproducible, targeted, no duplication.

• What it does: Adds a single regression test to tests/chat-routes/incremental-persistence.test.ts:474 asserting that runDetachedTurn's final write is authoritative over the last throttled draft — using the shipped 2s cadence (so only the first draft escapes the floor) and a tool part left permanently 'running' (only finalizeAssistantParts terminalizes it).
• Goals it achieves: Kill the one surviving mutation (M6) from the #250 mutation-testing pass: replacing draft.finalize(values) with draft.close() in detached-turn.ts left all 11 existing tests green because the existing autonomous-lane test hid the invariant behind two coincidences (intervalMs:0 + completed tools). The test makes the completion-time settlement observable.
• Assessment: High-quality, minimal, in-grain. It reuses the suite's existing real-SQLite harness (freshStore/rows/waitUntil/partUpdated) and sits in the same describe block. I reproduced the mutation: finalize->close fails ONLY the new test (expected 'Looking' to be 'Looking, then gave up.'), other 11 stay green — exactly the PR's claim. Reverted; 12/12 green; clean tree. The two coincidence-removers are load-
• Better / existing approach: none — this is the right approach. Searched for an existing equivalent: incremental-persistence.live.test.ts:76 covers the interactive lane (createChatTurnRoutes), is CI-skipped under LIVE_SANDBOX, and asserts the opposite (mid-stream snapshots must NOT terminalize) — so it neither covers the autonomous lane nor duplicates this invariant. No mutation-testing harness (stryker/mutate) exists in pack
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: bridge stream ended without value-audit content

🎯 Usefulness — sound

A mutation-found test that closes the one gap (#250's M6) where dropping the autonomous lane's authoritative final write left all prior tests green; verified to kill the mutation.

• Integration: The test exercises the real, shipped, exported runDetachedTurn (src/chat-routes/detached-turn.ts:261) through its real persist seam against the real SQLite-backed createChatStore — the same harness and fixtures (freshStore/rows/waitUntil/partUpdated) as the other 11 tests in this file. It runs in the existing pnpm test suite (confirmed: 12 passed). No new surface, no new wiring — i
• Fit with existing patterns: Follows the exact pattern of the neighboring crash-convergence test (tests/chat-routes/incremental-persistence.test.ts:419): same runDetachedTurn shape, same persist seam, same assertion style, same waitUntil mid-run probe. It sits naturally in the crash-safe convergence describe block. The distinction it draws (draftAssistantParts leaves a tool running vs finalizeAssistantParts term
• Real-world viability: The scenario is realistic and non-coincidental: default 2 s cadence (the shipped default, so later drafts are throttled to staleness) plus a tool that never completes (the dangling-tool case finalizeAssistantParts exists for, src/stream/stream-normalizer.ts:495-501). The mid-run waitUntil probe proves the row is created by a DRAFT — so a lane that skipped finalize would leave a stale row, not
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260725T024915Z}