Every Crawler designation is a globally sequential registry number, and the sequence is the point: it is a record of order of entry, so it has to be unique, gapless in intent, and impossible to hand out twice.

That rules out generating one client-side, or picking one and hoping. Numbers are issued by the database itself, under conditions that make two people registering in the same second a non-event, and certain numbers and ranges can be reserved or withheld rather than handed out. The interesting work was getting that to hold under concurrency; the exact mechanism is the sort of thing a colophon is worse for describing than a code review is.

The scaffold shipped collections as a single flat table, one row per thing you own. I kept it for about two months before admitting it was wrong. A book is not one object. A book is a work, which has editions (hardcover, paperback, audio, a signed printing), and every edition has its own images: cover, back, spine, interior, poster. None of that fits in one row, and no amount of columns was going to save it.

So the table got dropped and the feature rebuilt as a normalized schema across works, editions, images, authors, series, genres, and tags. The migration that deletes the generated version is still in the repo, sitting in the sequence where it happened. It is probably the most honest file in the project: the point where what got handed to me stopped being good enough and I had to know why.

Two third parties get contacted on this site: Vercel Analytics and Google Fonts. Without consent, the analytics component is never mounted at all, so nothing is requested.

The fonts were the harder half. They were originally pulled in with an @import in the stylesheet, which fires while the CSS is being parsed, long before React can mount a banner and ask anyone anything. A consent gate in front of that is decorative. So the stylesheet link is now injected only after consent is given, and until then headings fall back to the self-hosted faces.