Adversarial re-review of df275bb found the two-guard ready-flip fix still left a
narrow double-race open, plus test-hygiene drift from the banUser param removal.
MED — residual stale-keepsake race the `export_released_at` guard alone missed
The flip relied on two guards defeating two clearers: the `release_seq` EXISTS
check (vs a re-release bump) and `export_released_at IS NOT NULL` (vs open_event's
clear). But release_gallery re-arms `export_released_at = NOW()` and bumps
`release_seq` as SEPARATE statements, so a stale worker's flip landing in the gap
between them satisfies BOTH guards (released re-armed, seq not yet bumped) and
resurrects a pre-reopen keepsake — the next re-release then skips regeneration and
serves an archive missing the reopen-window uploads.
Root-cause fix: `open_event` now bumps every `export_job.release_seq`, making a
reopen a supersession point symmetric with a re-release. A worker that captured the
pre-reopen seq can never match its `release_seq`-guarded finalize/flip again,
regardless of when the re-release re-arms `export_released_at`. The released-anchor
guard stays as defense-in-depth. Added a deterministic e2e asserting the reopen
bumps the seq (the invariant that closes the race without depending on
sub-millisecond worker timing).
LOW
- Gate the `export-progress: 100` SSE + `prune_stale_export_files` on the ready-flip
actually flipping (rows_affected > 0). A superseded worker no longer advertises a
misleading 100% or prunes on a fresh generation's behalf.
- moderation.spec.ts: the two ban tests had become byte-identical after the
hide_uploads param removal; drop the one whose "hide_uploads=true" title no longer
matched what it exercised, keep the accurate "always hides" test.
- host-dashboard page-object: drop the dead `banUser(hideUploads)` param + its
checkbox branch (the UI no longer renders that checkbox — a latent hang trap).
- sse-eviction.spec.ts: rename the test whose title referenced the removed flag.
Verified: backend 40 tests, e2e 156 passed / 1 skipped on chromium-desktop
(incl. the new reopen-supersession regression).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1.4 KiB
1.4 KiB