v1.5.3

OMK v1.5.3

OMK 1.5.3 is a patch release: three runtime fixes, plus test-fixture repairs that make the
suite green on a loaded host and restore CI.

Fixed

• Skill search phase. SkillSearchGate now passes the active TODO kind into metacon
  ranking. The ranker and the skill-select run log already accepted a phase, but the
  gate never supplied one, so ranking and phaseSource always reported the task default.
• Concurrent attachment previews. Two previews of the same active attachment now share
  one in-flight Photon decode instead of each allocating a full-resolution image and a
  resized RGBA buffer. A dropped draft still releases its entry through the existing
  WeakMap.
• Journal reads during a live append. A reader that opens the journal before a writer
  appends and reads after it could observe a partial final line and report integrity.
  readRunJournal now retries that bounded window as a concurrent change, and still reports
  journal_truncated when the retries are exhausted. Middle-file corruption is unchanged.

Tests and CI

• Two compaction tests pin a 1 ms retry backoff. The default policy (5 attempts, 2 s base)
  waited about 62 s for a 503, past the 30 s Vitest timeout, which failed CI on every run
  since retry.maxRetries changed to 5 in 1.5.1. Retry counts and assertions are unchanged.
• verified-run.test.ts raises its fixture work/verify budget to 30 s and derives the
  "past the work deadline" clock from that budget. The old 3 s budget also bounded process
  spawn, so a loaded host failed the suite with deadline. The timeout behavior keeps its
  own test at workMs: 100.
• verified-run-authority-restart.test.ts uses a 60 s grant TTL. Its fixtures assert restart
  quarantine, not expiry. Expiry remains pinned deterministically in
  verified-run-authority-store.test.ts.
• memory-factorial allows 30 s for the host clock to recover instead of 5 s, so a WSL2 time
  resync no longer fails the offline experiment. The store's real-clock guard and the
  future-dated-memory assertions are unchanged.
• The journal-read-retry mutator retries a short pwrite, so the test only produces the two
  valid journal states it claims to produce.

Notes

• No CLI, protocol, or configuration keys were added or removed. There are no intentional
  breaking changes.
• The full suite (./test.sh) passed twice in a row on this host while it was under load
  (load average about 26). npm run check and npm run build pass.
• Harness² (google-research/harness2) is not a dependency of this release. The evaluation
  in notes/2026-10-19-harness2-assessment.md concludes it is a Python 3.12 research harness
  for OpenCode/Codex substrates with no OMK substrate, so only its static validation rules are
  worth porting, under a separate spec.