fix(sleep): revalidate transient unlinked latest pointers

[Yifan Yang (Yif-Yang)]

Summary

Concurrent staging publications can fail when .latest is atomically replaced while its metadata is being inspected. A snapshot of the replaced regular inode can report zero links, which the publisher currently treats as permanently unsafe.

• Revalidate only the derived .latest pointer after a zero-link regular-file observation.
• Require a non-link regular inode with exactly one link before publication; repeat link/junction and file-type checks on every attempt.
• Keep hard links, nonregular files, missing paths, and persistent zero-link states rejected.
• Bound retries to 21 observations / 20 backoffs (at most 1.05 seconds of requested sleep, not a wall-clock deadline).

The atomic writer, pointer reader, adoption-target authorization, receipt/backup validation, and transaction recovery code are unchanged. This is not a claim to eliminate every filesystem replacement race.

Validation

• Reproduced the existing concurrent-publication failure on unmodified main and preserved that failure.
• A deterministic transient-zero regression fails before the fix and passes after it, with the imported implementation verified.
• All 19 deterministic boundary cases pass, including unsafe type/link transitions, persistent invalid states, and recovery on the last permitted observation.
• A fixed 200-iteration run of the original concurrent-publication test completed with 200 passes and zero failures; each iteration uses 16 publications / 8 workers.
• On the audited combination of fix(model): wire openai_compatible configuration across trainer, eval, and config pipeline #262, fix: harden security beyond PR #249 — command injection, deps, path injection #264, docs(guide): document Cheaper Inference as an openai_compatible provider example #297 and docs: clarify current and best state under strict gating #301: 1713 passed, 13 skipped, 359 subtests; focused staging/adoption/engine suite: 292 passed, 2 skipped.
• This PR was then independently cherry-picked onto current upstream main fa4ca184573e42ec11472959dd57422381418096 and revalidated: 1576 passed, 13 skipped, 359 subtests.
• Changed production and test file contents match the independently reviewed patch. Ruff and diff checks pass.

Tests used isolated synthetic data on Linux with no real user transcripts or paid model calls. Junction detection was simulated; native Windows validation is not claimed. Local evidence does not replace official CI.