[thymikee]

Summary

Follow-up to #3323 (merged as c8b944c). Replayed from its stack onto main with an identical patch (same git patch-id).

The kernel keyed lock now answers which execution locks a context holds, including locks taken partway through a request. Request locks keep no registry of their own.

• keyed-lock.ts:
  • Scopes link to live parents only.
  • withKeyedLockHolder can hold locks until its task ends. heldKeyedLockKeys lists held keys, including keys still being taken.
  • A scope that wants a key its holder is still taking waits for the hold, then re-enters it. If the hold is cancelled, the scope fails with it.
• request-execution-locks.ts applies the same-Mac rule (ADR 0035) as one check; its registry and re-entry special cases are gone.
• A nested request re-enters a lock its parent took partway through instead of waiting on its parent.
• The close-refusal test now uses the real platform refusal.

5 files. Closes #3876.

Validation

Head e5813a1d02bf8c9066381f576c1aacdf9251661d, on main c8b944c.

• check:affected --run --head <head>: passed, run-jL09Mn, exactCommit: true, 4,523 related tests.
• Before the replay, at 9e2ad9b on 7519086: run-KLsDWf passed (12,421 related tests); three independent review rounds found nothing blocking; and a native live run passed, with two apps in parallel in one daemon, conflicts and close scope held, and no claims left. These are attributed to 9e2ad9b; the patch is unchanged.
• New kernel and daemon tests time out on the old code.

🤖 Generated with Claude Code

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.52 MB	5.52 MB	+1.2 kB
Package (unpacked)	5.52 MB	5.52 MB	+1.2 kB
Package (download)	1.67 MB	1.68 MB	+386 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	34.2 ms	34.2 ms	+0.1 ms
CLI --help	69.7 ms	68.7 ms	-1.0 ms

[thymikee]

Closing and reopening to re-run CI against the new base: the Size job of the retarget push measured the deleted stack base 7519086 (unable to read tree).

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-22 03:09 UTC

[thymikee]

This PR is ready at e5813a1. Not blocking, and you can take or leave these: the pending-hold tracking in packages/kernel/src/keyed-lock.ts only matters if work runs concurrently inside one request, and on the routes I traced the device is retained before the next step starts, so drop it or name a router-level caller in a test. The re-entry fix is covered at the createRequestExecutionLocks seam in request-execution-locks.test.ts but not through the replay router, so a later same-device step in the dynamic-selection replay test would close that gap. heldKeyedLockKeys could be generic over the key type to avoid the string[] cast in request-execution-locks.ts. The ended-holder test matches on error message text.

Bundle Size failed only because it could not read the deleted base tree 7519086, and the Smoke Tests, Repo Guards and Coverage runs were cancelled, not failed. The diff touches no bundling or packaging inputs, so the failures look unrelated. There are no conflicts. This was a read-only review, so I ran no tests. I also could not check the patch-id match with 9e2ad9b or the live two-app run at that SHA, because that object is not in my clone. Before merge, CI needs to go green on e5813a1 after the reopen.