0021. Metadata lookup dialog: explicit search, more page sources, saved to the .nfo

• Status: Accepted
• Date: 2026-10-19

Context

The lookup dialog (ADR 0005, ADR 0006) searched as soon as it opened, with a query guessed from the file name, and queried every enabled API. A pasted Bédéthèque link was recognized in the same field as the search text, and it was the only site supported. Applying the result only wrote the database, so the .nfo beside the comic, which the scan reads back, stayed stale.

BDfugue (a bookshop) and Nautiljon (a manga database) are the next sites users want to read from a link. Neither has a public API, and this environment's authors could not fetch their pages when this was written.

Decision

• Nothing is searched on opening. The dialog shows its options first: Search in APIs (text, volume, one checkbox per enabled API, sent as MetadataQuery.sources) or From a page link. The search starts when the user presses the button, and the "Arrêter" button abandons it: the renderer ignores the answer (latestSearch), without interrupting the main process.
• More sites by link: BDfugue and Nautiljon join Bédéthèque as link-only sources (pageSourceOf()). Rather than parsing their markup, parseStructuredPage() reads the standard metadata the pages embed for search engines and link previews (schema.org JSON-LD, then Open Graph tags). Their authors get the generic author role, like Google Books'.
• Saving goes through metadata:apply, which updates the entry, rewrites the .nfo (replacing it, unlike the resynchronization's write-if-missing) and makes the cover thumbnail if the book has none. The manual edit form keeps library:update-metadata.

Consequences

• The parsers rely on conventions more stable than the sites' markup, but a site that publishes no JSON-LD or Open Graph title is refused ("cette page ne décrit pas un livre"). The fixtures of structured-page.test.ts were written from the conventions, not from real pages: check them against a real BDfugue and Nautiljon page and extend them if a field is missed.
• Stopping a search doesn't save Comic Vine's quota: its sequential requests continue in the main process until they end.
• The .nfo is overwritten on every applied lookup, so a hand-edited .nfo is replaced by what the user just accepted. The cover step only helps books whose thumbnail is missing; it does not download the source's cover.
• The CSP's img-src gains *.bdfugue.com and *.nautiljon.com for the result thumbnails.

Alternatives considered

• Cancelling in the main process (an id per search and an AbortSignal through the HTTP helpers): rejected for now; it adds a channel and plumbing in every client for a search that lasts a few seconds.
• Scraping each site's markup like Bédéthèque: rejected, as it couldn't be written or tested without real pages and breaks on each redesign.