[magrhino]

Summary

Split out of #2759 at the maintainer's request. Hardcover edition hydration no longer overwrites book edits made while editions are being fetched, and it no longer silently drops its own results on rows whose updated_at was stored in a legacy shape.

Closes #2758.

Stacked on #2759. This branch comes from a fork, so it also carries the #2755 commits until #2759 merges. Only the last two commits are in scope:

1. fix(metadata): preserve concurrent edits during hydration (#2758)
2. fix(metadata): compare raw updated_at text in the hydration guard (#2758)

What changed

• Guarded metadata write. BookRepo.UpdateHydratedMetadata writes only the hydrated columns, and only if the row hasn't changed since the snapshot. When a concurrent edit wins, the book is reloaded, so the caller can't act on a discarded ASIN or discarded metadata.
• Guarded edition upsert. EditionRepo.UpsertMetadata checks that the parent still has the foreign ID and provider seen before the fetch. The check runs in the same statement as the insert or conflict update.
• Reload on every path. The caller's book is refreshed on the duration-only, empty, unchanged, and failed-fetch paths before callers search.
• SetLanguage now bumps updated_at, so a pending hydration sees an imported EPUB language.

How the precondition compares updated_at

The first version of the guard compared updated_at against a reformatted time.Time. books.updated_at holds several historical shapes, and for the two legacy ones the guard never matched: it returned updated=false, err=nil, so every refresh discarded its results.

The guard now keeps the column text exactly as stored, in Book.UpdatedAtRaw, and compares that. It reads the column through CAST(updated_at AS TEXT), because the SQLite driver turns DATETIME columns into timestamps even when the scan target is a string. There is no schema change.

Stored updated_at shape	Where it comes from	Before	After
2026-09-01 10:00:00	CURRENT_TIMESTAMP (migrations 082, 084)	never matched; results silently dropped	matches
2026-09-01 10:00:00.123456789 +0000 UTC	Go time.String() (#914, still written by author merge, #2792)	never matched; results silently dropped	matches
RFC3339Nano	current writers	matched	matches
unparseable	corrupt rows	hard invalid book snapshot error	matches, and the write normalises the value

I chose this over a row-version column because it needs no migration. Like any timestamp guard, two writes in the same clock tick would look identical. A version column is the only full fix for that, and that would be a separate change.

Tests

• TestBookRepoUpdateHydratedMetadataLegacyUpdatedAtShapes covers all four stored shapes:
  • an unmodified row accepts the hydration write;
  • a concurrent edit through each writer that bumps updated_at still wins: SetLanguage, SetImageURL, MarkWantedMonitored, SetExcluded, and author merge. It fails for a writer that stops bumping the timestamp.
• TestHydrateHardcoverEditionsPersistsOverLegacyUpdatedAt runs hydration end to end on both legacy shapes and checks that the result flags and persisted metadata survive. It fails against the old reformatted comparison.
• TestBookRepoFillMissingAudiobookDurationRequiresMatchingIdentity (from fix(metadata): retain hardcover audiobook runtime before search (#2755) #2759) now also checks that the reloaded duration comes with its own identity, not a stale one.

Follow-ups (not in this PR)

• Refs fix(db): author merge writes books.updated_at in the #914 time.String shape #2792: author merge writes books.updated_at in the Books page blank — unsupported Scan error on legacy date formats from Calibre import #914 shape.
• Refs fix(db): SetFilePath fallback changes status without bumping updated_at #2793: the SetFilePath fallback changes status without bumping updated_at.

Checklist

• Commits signed off with git commit -s
• Tests added or updated
• docs/DEPLOYMENT.md updated if env vars, config, or upgrade path changed (not applicable)
• Added a changelog fragment under changelog.d/ (2758-hydration-concurrent-edits.md)
• Wiki pages updated if user-facing behaviour changed (docs/User-Guide-Wiki.md)

Test plan

• go test -count=1 ./cmd/... ./internal/...
• go vet ./internal/...
• cd web && npm run build (not run; no web changes)

[codecov]

Codecov Report

❌ Patch coverage is 90.54054% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/db/books.go	86.79%	4 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[vavallee]

Early look since it's still a draft, no action needed yet. Reading `CAST(updated_at AS TEXT)` and comparing against that same raw text is the right fix for the problem that sank the earlier version: it's correct whatever shape the row holds, so none of the nine historical layouts can make it silently refuse.

The guard is only as good as "every writer of a guarded column bumps `updated_at`", so I went through main's `UPDATE books` statements that don't:

Writer	Column	Matters here?
`SetFilePath` fallback, `books.go:1118`	`status`	yes, `status` is in your guarded set. That's #2793, so it wants to land before or with this
`SetLanguage`, `books.go:1134`	`language`	yes, and you've already fixed it here
`SetCalibreID`, `books.go:1162`	`calibre_id`	no, not in the guarded set
`db.go:647`, `:1064`	`dedup_key`, `sort_key`	no, startup backfills

The main import path is fine: `AddBookFile` goes through `refreshBookStatus` into `Update`, which bumps it.

[magrhino]

Thanks for walking through the UPDATE books writers, that table pinned down the gap.

The SetFilePath fallback from #2793 is fixed in this PR as of 4588a40. It now bumps updated_at alongside status and writes through r.exec. The concurrent-writer table in TestBookRepoUpdateHydratedMetadataLegacyUpdatedAtShapes has a new SetFilePathFallback case. It forces the fallback by storing an unscannable calibre_id so GetByID fails, then checks that a pending hydration write loses. Without the fix it fails on all four stored updated_at shapes.

Verifying that turned up one more pre-existing problem in the same fallback, now open as #2819. The fallback takes the legacy write on any GetByID error and drops that error, never adds a book_files row, and reports success for a missing book. It doesn't affect the hydration guard, so I've kept it out of this PR.

#2759 has merged, so I've rebased onto main. The diff is now just this PR's three commits. I'll mark it ready for review once CI is green.

[vavallee]

Thanks! Heads up that #2824 landed after your fix and removed the SetFilePath fallback entirely: a load error is now returned and a missing book is reported. So the fallback hunk and the SetFilePathFallback subtest can go when you rebase onto main. Once it's rebased I'll take another look.

[magrhino]

Rebased onto main and dropped the fallback hunk and subtest now that #2824 removed it.