[Squeaks72]

Stacked on #84 (chapter names). Review that first; this diff includes its commits until it merges. Only the last two commits belong to this PR.

Summary

Many sources publish no chapter titles at all. Some answer "Chapter 1", "Chapter 2" for the whole of One Piece, while another source has had "Romance Dawn" all along. This PR lets an admin borrow those names from another source, off by default. It can be turned on server-wide (Admin → Settings → Updates & schedules) or per series (a checkbox in the Sources & translations sheet).

The danger is the numbering, not the names. Matching a name to a chapter by number is only sound if both sources number the work the same way. When they don't, every borrowed name is wrong, and wrong in a way that looks right. So bff/src/lib/borrowNames.ts is deliberately conservative:

• Donor choice. A donor must match the series title exactly after case and punctuation are folded (normTitle, from the per-source searchAll map). A fuzzy match would add a second way to be wrong, and the numbering check cannot catch it. A title that folds to nothing (entirely non-Latin script) is never searched, because it would "match" every other such title.
• Numbering check. A donor must pass the same rule the app already uses before following a second source (assess / verdict / followable from lib/fill.ts: at least 90% of the numbers we hold, and the verdict agreeing). A donor that fails is not used at all.
• What gets written. Only chapters whose title is still just their number are written, and that is re-checked in the UPDATE itself. Every borrowed name records its donor in the new lib_books.title_source column. The chapter's own source naming a chapter later (the landed stamp or the listing heal from Show chapter names, and say which chapter Continue opens #84) replaces the borrowed name and clears the mark.
• Switching off. Turning it off for a series removes that series' borrowed names, resetting them to the title the chapter's file gives, which is exactly what the scanner derives. Turning off the server switch does the same for every series that follows it; a series switched on for itself keeps its names.
• Traffic. It runs after the listing is written on each source check. It makes no request once every live chapter has a name. The matched donor is remembered (lib_series.name_donor). A search that found no donor is remembered for a week instead of re-running a cross-source search every few hours.

Schema: server_settings.borrow_names (default false), lib_series.borrow_names (null follows the server), lib_series.name_donor, lib_books.title_source. All are ADD COLUMN IF NOT EXISTS.

API: borrowNames on PATCH /api/admin/settings and PATCH /api/admin/series/:id (nullable there), borrow_names on the settings row, and borrowNames / borrowNamesEffective on the admin view of GET /api/series/:id. The checkbox has to show what actually applies, including the server's setting. All are documented in bff/openapi.yaml. There are no new routes, so the docs/api.md route list is unchanged. It works the same on Desktop: none of these routes are desktop-hidden, and the switch lives next to the existing "Look for failed chapters on other sources" switch.

A known limit, stated plainly: the coverage check proves the donor lists the numbers we hold. It cannot detect a donor that lists the same range with a one-chapter offset, such as a prologue counted as chapter 1. The exact-title requirement, the opt-in default and the recorded title_source (removable in bulk by switching off) are the mitigations. That limit is the main reason this stays off by default.

Type

• Fix
• Feature
• Docs
• Refactor / chore

Testing

Re-run after rebasing onto v0.45.0 (91b41aa): bff 1867 tests, 1854 pass, 9 fail, 4 skipped — the same 9 embedded-Postgres entrypoint failures main has in this environment (main: 1857 / 1844 / 9 / 4), nothing new; web 432/432; both typechecks OK. The detail below is from the first run, on v0.44.0.

I ran the upstream CI setup in node:24-alpine against a real postgres:16-alpine (TEST_DATABASE_URL set), with the same baseline as upstream main (a8bc502).

	upstream main	this branch (with chapter-names beneath it)
bff tsc --noEmit	OK	OK
bff tests	1834 tests: 1821 pass, 9 fail, 4 skip	1844 tests: 1831 pass, 9 fail, 4 skip
web tsc --noEmit	OK	OK
web tests	421 / 421	422 / 422

The same 9 failures appear on both sides. They are the entrypointEmbedded tests, which need an embedded Postgres binary that this sandbox lacks. There are no new failures. The full run first caught predicateHygiene flagging a hand-written deleted_at IS NULL. The series lookup now uses visibleToAll('s'), and I re-ran predicateHygiene, borrowNames, chapterNames, seriesListing, updater and openapiCoverage: all 74 tests passed. openapiCoverage passes with the schema additions.

New: bff/test/borrowNames.int.test.ts, using fake sources for a correct donor, a donor with numbers shifted to 101–110, and a look-alike title with matching numbers. It pins:

• nothing is asked while the feature is off
• only the correct donor is used, and a chapter it does not name either stays unnamed
• a series whose only candidates are the shifted and look-alike sources gets nothing, and its search is not repeated on the next check
• the own source's listing name replaces a borrowed name and clears the mark
• switching off resets to the file's title. Chapter 10 reads "Chapter 10": an earlier trim(trailing '0' ...) reset turned it into "Chapter 1".
• the server switch going off clears only the series that follow it

The new UI strings are translated in all eight locale files. npm run test:i18n is a Puppeteer walk against a running server, so it could not run in this sandbox. The locale-parity unit test passes.

Checklist

• Focused scope (one topic)
• Matches the existing TypeScript style
• Docs updated if user-facing (docs/USAGE.md)
• No new default sources hardcoded to specific sites (the engine system reaches sites by URL)

🤖 Generated with Claude Code

[AngeloSha]

Thanks, this is a clever idea and I do want it. I closed this version though, because a few of the problems need it restructured rather than patched:

• It searches through searchAll, which reports slow and failing sources to source health. So a lookup that's only for names can put someone's main source into a cooldown and stop real downloads. The failure hunt goes out of its way never to do that.
• Donors aren't filtered by adult rating or language, and normTitle strips a title to [a-z0-9], so a Spanish "One Piece" matches and you get Spanish chapter names.
• The numbering check floors both sides and has no minimum, so a donor one chapter off, or a different work with the same title, can pass.
• There's no limit on how much it searches, and the remembered donor ignores a source the admin disabled.

Since #84 keeps names in their own column, I rebuilt this on the existing hunt pipeline (adult rule, own-language order, budgets, no health reporting). It shipped in v0.47.0 as Borrow chapter names from other sources, off by default, with credit to you in the changelog.