boshu a8267ddc8f feat(storage): parent-dir fsync after rename + consolidate dir-fsync (age-y9jy) ...

Add storage.FsyncDir — the single dir-fsync impl — and call it after the rename
in canonical storage.AtomicWriteFile, so the rename itself (not just the file
body) is durable across power loss. Every delegator (quest, cmd/ao inject,
vendorimage codexruntime, llmwiki.AtomicWriteFile) inherits it.

Close the same gap in the separate-rename writers: pool.atomicMove,
llmwiki.AtomicWriteFromReader, storage.FileStorage.atomicWrite. Dedup the 4th
private copy (evalsubstrate.WriteAtomic's fsyncDir) onto storage.FsyncDir.

FsyncDir swallows only the directory-fsync-unsupported errnos (EINVAL/ENOTSUP/
ENOTTY — macOS APFS) and propagates genuine I/O errors, not the prior
swallow-all shape that would hide a real durability failure.

Note: the bead's '3 private copies (inject/vendorimage/pool)' framing was
stale — inject and vendorimage already delegate through quest->storage; the
real remaining work was the parent-dir fsync durability gap across all rename
sites.

2026-06-26 08:31:16 -04:00

..

errors.go

evolve: cycle 97 -- modernize sentinel errors: fmt.Errorf -> errors.New

2026-02-22 04:44:28 -05:00

ingest_orchestrator_test.go

refactor(pool): extract shared file-iteration orchestrator

2026-05-01 10:37:48 -04:00

ingest_orchestrator.go

refactor(pool): extract shared file-iteration orchestrator

2026-05-01 10:37:48 -04:00

ingest_parse.go

refactor(cli): extract metrics_golden + metrics_health + notebook (wave 6.3)

2026-04-06 08:21:33 -04:00

pool_test.go

feat(flywheel): cross-writer dedup contract + advisory promotion-volume gate (4-A + 4-B)

2026-04-30 08:41:11 -04:00

pool.go

feat(storage): parent-dir fsync after rename + consolidate dir-fsync (age-y9jy)

2026-06-26 08:31:16 -04:00

promoted_content_test.go

fix(pool): block recreation of archived knowledge bodies

2026-05-01 12:27:28 -04:00

promoted_content.go

fix(pool): block recreation of archived knowledge bodies

2026-05-01 12:27:28 -04:00