[Juliusolsson05]

Closes #1341. It also covers #1296 items 3 and 7, which were moved here from #1353.

Problem

GoalLoopService started work nothing could await: void this.persist(), void this.maybeContinue(), and zero-delay setTimeout(requestContinue) deferrals.

• afterEach removed each test's temp dir while a persist was still renaming into it. The GoalLoopService worker on main CI run 36156560842 logged 13 persisting loop state failed: ENOENT.
• goalLoopTurnHooks.system.test.ts slept 20 ms before rm, and still hit ENOTEMPTY under load (Machine- and time-dependent tests beyond #1107 #1296 item 3).
• Fixed 10–300 ms sleeps followed by "not delivered" passed whenever the work simply had not run yet (Machine- and time-dependent tests beyond #1107 #1296 item 7).

Change

• whenSettled() resolves when no tracked work is in flight. Tracked work is every promise the service starts without awaiting, plus every zero-delay deferral until it has run. It loops, because settling work can start more.
  • The future polls (1 s hold poll, retry backoff) are deliberately not in flight. Waiting for them would mean waiting for a hold to end.
• dispose() stops new work (a delivery already in flight records its outcome but does not retry; defer schedules nothing; a racing start() attaches nothing; startLoop, complete and carry reject; control returns null; hooks are ignored), detaches the manager listeners start attached, cancels every timer, and drains. maybeContinue returns once disposed, so no path (including a continuation queued behind one in flight) delivers after dispose begins.
• All raw setTimeouts go through one owned defer, so dispose can clear every timer.
• Tests: all three goal-loop files dispose their services before removing directories. The vacuous sleeps became whenSettled(). The two sleeps that wait on real timers stay, each with the reason:
  • the 1.5 s one is a positive wait on the 1 s hold poll;
  • the 400 ms one is sound because Node fires timers in expiry order.
• Not here (decision): wiring dispose() into the app's shutdown inventory. That composition has its own owner and exhaustive inventory, and a lost final persist at quit is pre-existing. Filed as goal-loop: dispose GoalLoopService in the app shutdown inventory (last persist can be lost at quit) #1372.
• Plan: docs/plans/2026-09-27-goal-loop-drain.md.

Tests

• New drain cases:
  • dispose waits for a persist parked on a gated store, then starts nothing;
  • whenSettled covers a zero-delay continuation check with no waitFor;
  • a disposed service ignores a held loop's direct quiet-edge continuation.
• Review round 1 regressions:
  • a queued continuation behind a gated persist does not deliver once dispose begins;
  • manager events after dispose change nothing and write nothing;
  • dispose waits for a control action's persist;
  • dispose cancels a held loop's poll timer.
• Mutations killed:
  • whenSettled resolving at once;
  • a zero-delay deferral not tracked;
  • dispose not draining;
  • a disposed service still continuing;
  • maybeContinue ignoring dispose;
  • listeners not detached;
  • control's persist untracked;
  • dispose keeping timers.
  • Round 2: the retry ignoring dispose, start ignoring dispose, defer after dispose, public mutators after dispose, control after dispose.
• src/main/goalLoop: 113/113, with 0 persisting loop state failed warnings.
  • origin/main also logged 0 locally. The 13 warnings are from CI under load, so this PR's CI log is the real check.
• npx tsc -b: 0 lines.

🤖 Generated with Claude Code

[Juliusolsson05]

Review disposition (round 1: a, b, c codex at 4145f1cf)

#	Finding	Verdict	Disposition
a1 / b1 / c1	A queued continuation, replayed from finally via maybeContinue, delivered after dispose() began	valid	Fixed: maybeContinue returns when disposed. The reviewers' sequence (gated persist, provider hooks close the turn, queued Resume) is a test; the "maybeContinue ignores dispose" mutant is red.
b2 / c2	Manager listeners outlived dispose(); a late removed paused the loop and wrote into a removed directory	valid	Fixed: dispose detaches the listeners start attached (off on the port). A test covers events after dispose (no change, no write, 0 listeners); the "not detached" mutant is red.
a / c surviving mutation	control()'s persist was not pinned as tracked	valid	Pinned: dispose waits for a gated control persist; the untracked mutant is red.
b surviving mutation	Timer cancellation was not asserted	valid	Pinned with fake timers: a held loop's poll is cancelled; the mutant is red.

Verification:

• src/main/goalLoop: 110/110, with 0 ENOENT warnings.
• b's three concurrent runs at 4145f1cf were already at 0 warnings.
• npx tsc -b is clean.
• The verification pass (a/b/c) is queued under the reviewer cap.

Round 2 (verification at 7a1d9c0d: a, b, c FIX-BEFORE-MERGE):

#	Finding	Verdict	Disposition
a1 / b1 / c1	A delivery in flight when dispose() began retried and could schedule a backoff	valid	Fixed in 8fd8b3f5: the retry is skipped once disposed, and defer is inert after dispose. Pinned with a gated delivery plus fake timers.
b2	start() racing dispose() attached listeners and wrote	valid	Fixed: start returns after its read if disposal began. Pinned with a gated read.
a2 / b3 / c2 / a3 / c3	Public entry points wrote or scheduled after dispose()	valid	Fixed: startLoop, complete and carry reject; control returns null; hooks are ignored. Pinned.

Five mutations killed. 113/113, with 0 ENOENT warnings. Round 2 is the cap, so acceptance is the manager's call.