fix(gsd): make auto mode complete a milestone end to end

[jeremymcs]

Auto mode could not finish a planned milestone. Nine defects, each found by running a real milestone against a scratch project and measuring the result, each with a test that fails without its fix.

Progress across acceptance runs

run	outcome
1	exit 10, zero files, zero commits — died at startup migration
2	stuck 25 min on one task, looping on a bogus verification failure
5	4/4 files, S01 closed — stalled at S02 close and exited 0 anyway
8	milestone complete, but 4 HARD BLOCKs per run
14	milestone complete, 0 blocks, 0 pauses, 17k events (fastest)

What was broken

Startup and dispatch

• migrateToFlatPhase held no cross-process lock. Headless runs the extension in two processes and both fire session_start, so two migrations raced and the loser died on ENOENT. Repro went 3/3 fail → 5/5 clean.
• Dispatch raced that migration's move-aside window — when slice plans exist in neither layout — and re-planned an already-planned milestone on every run. Scoped to a migration actually in flight so a settled legacy project still reaches the rule (#4671).

Verification and gates

• The verification gate executed a task's prose verify field as a shell command: greet/hello.txt exists and contains "hello" tried to run the .txt file, exited 126, and failed a task that had succeeded. The prose detector only caught capitalised prose.
• Browser-UAT escalation fired on a negated mention — no runtime behavior, server, UI, or browser interaction is involved reads as a browser requirement when the negator is a list away.

Reporting

• Blocked stops exited 0, reporting success over an unfinished milestone. Two independent breaks in stop-notice.ts — the module written to keep emitter and detector in lockstep. Either alone loses the exit code.

Projection and identity

• Slice plans now state their DB identity; under flat-phase layout an agent inferred milestoneId: "01-m001" from the directory name and was hard-blocked.

Unit tool contracts

• Contracts were enforced but not advertised: 15 of 24 units with an allowedGsdTools list had no tool guidance at all. The allowed set is now derived from the same table the enforcer reads, and repeated at the dispatch seam — the ## Tool Surface section alone sat 4% into a 14K-character prompt and had no measurable effect. HARD BLOCKs went 4 → 0.
• The reminder is phrased as a reference, not a menu: an earlier iteration advertising gsd_reassess_roadmap invited validate-milestone to add slices to a completed milestone until runs stopped terminating.

Worktree closeout

• Under git.isolation=worktree the closeout gate computed the verification source revision from the project root while validation ran in the worktree — different trees by design — so a milestone the DB had marked complete was refused at merge. Measured: worktree sha256:074a8c04… matched the recorded authorization exactly; root sha256:a86aa329… did not.

Verification

• A 2-slice/4-task milestone completes unattended: exit=0, both slices and the milestone complete in the DB, byte-exact artifacts, 0 HARD BLOCKs.
• Each fix has a test that fails without it.
• Full gsd suite: 137 failures both with and without these changes — no regressions introduced.

Known limits

• The worktree merge is not confirmed end-to-end. The closeout gate fix is verified against the exact blocked project (gate flips to ok) and introduces no regressions, but every full acceptance attempt under worktree isolation was cut short. The gate fix also has no regression test — one needs a project with a registered git worktree plus recorded validation events.
• Two of the nine fixes were wrong on first attempt and corrected after measurement (a dispatch guard that broke 5 tests; a prompt fix that changed nothing). Both are noted here so reviewers know where to look hardest.
• A tenth fix — stamp-insensitive projection drift comparison — is deliberately excluded: it depends on the state-version stamping mechanism, which does not exist on main.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.}

[github-actions]

🔴 PR Risk Report — CRITICAL

Files changed	15
Systems affected	3
Overall risk	🔴 CRITICAL

Affected Systems

Risk	System
🔴 critical	Auto Engine
🟠 high	GSD Workflow
🟢 low	Headless Mode

File Breakdown

Risk	File	Systems
🔴	src/resources/extensions/gsd/auto-dispatch.ts	Auto Engine
🔴	src/resources/extensions/gsd/auto/loop.ts	Auto Engine
🔴	src/resources/extensions/gsd/stop-notice.ts	Auto Engine, Headless Mode
🟠	src/resources/extensions/gsd/verification-gate.ts	GSD Workflow
⚪	src/resources/extensions/gsd/browser-evidence.ts	(unclassified)
⚪	src/resources/extensions/gsd/closeout-consistency-gate.ts	(unclassified)
⚪	src/resources/extensions/gsd/flat-phase-migration.ts	(unclassified)
⚪	src/resources/extensions/gsd/markdown-renderer.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/browser-evidence.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/dispatch-during-migration.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/flat-phase-migration.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/stop-notice.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/tool-surface-affordance.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/verification-gate.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/unit-context-composer.ts	(unclassified)

⚠️ 🔴 Critical risk — the following systems require verification before merge:

• 🔴 Auto Engine: validate auto-mode trigger conditions and loop termination
• 🟠 GSD Workflow: verify GSD workflow state transitions end-to-end

⛔ This PR should not be merged without executing this follow-up prompt.

Ask your coding agent to verify before submitting:

Review this PR for risks in: Auto Engine, GSD Workflow. Verify:

1. validate auto-mode trigger conditions and loop termination
2. verify GSD workflow state transitions end-to-end

Before modifying any code, assess the scope of this fix:

- Identify the root cause, not just the reported symptom.
- Search the codebase for other call sites, similar patterns, or duplicated logic that may share the same bug.
- List affected tests, documentation, and any downstream consumers that depend on the current behavior.
- Flag any changes that extend beyond the immediate file or function.

Report findings first. Then propose a fix scoped to the actual root cause, and wait for confirmation before applying changes outside the originally reported location.

💡 Have a Codex subscription? Get an independent second opinion: codex review --adversarial