[Ruby184]

A guide merged from several sources wrote every spelling of a programme's texts: "DJ Mixes premiere" beside "Dj Mixes Premiere", "Film" beside "film", a description cut off with "…" beside the whole of it. homepi carried ~120 lines of transform to undo it, and measured thousands of such programmes a day. One commit, breaking (guides change on upgrade).

Breaking

• A programme's texts that say the same thing are written once. A multi-source guide writes fewer titles, categories and descriptions than before; none is rewritten, only left out.
  Migration: merge: { union: 'exact' } writes every text as the sources gave them, as before.
• ResolvedMatch.normalizeTitle is required. resolveMatch always set it; only a hand-built ResolvedMatch needs the field.

merge.union

Per field, which texts are one — within a language, with the higher-priority source's kept:

field	default	the same when
title, subTitle	'match'	the match's normalizeTitle makes them the same (case, accents and spacing by default), so a title rule that loosens what is one broadcast loosens what is one title too
desc	'contained'	one contains the other as whole words, but for case, spacing and a trailing …; the fuller one takes the shorter one's place. Krimi beside Krimiseriál o … stays
category, keyword, country	'caseless'	the same but for case, spacing, hyphens and underscores, which is how a profile finds a genre (Sci-Fi and Sci Fi are one)

A field also takes 'exact', { contained: fn } (the source's own suffix taken off first, then compared as 'contained' does), or a key function of its own. union: 'exact' sets every field exact. Languages are never compared with each other; a missing tag and an empty one are the same, and tags are compared regardless of case.

merge: {
  union: {
    desc: { contained: (value) => value.replace(/\s*\(skryté titulky\)$/, '') },
    title: (value) => value.replace(/\s+[IVX]+\.?$/, '').toLowerCase(),
  },
}

Applied as each programme is written

Not as two programmes merge. Matching decides what is one broadcast by its titles, so it has to see every one: a spelling dropped when the first two sources merged would leave a third source's shifted programme with nothing to match, and it would come out as a duplicate. So:

• mergeProgrammes, mergeInto and backfillInto keep every text, as before.
• The guide applies the union as it writes each programme, and hands a transform the next programme with the union applied too.
• mergeProgrammeLists(lists, strategy, match?, fillStopMs?, union?) applies it to what it returns.
• resolveUnion(union, resolveMatch(match).normalizeTitle).apply(programme) applies it to one programme.

Settled at load

• A misspelt field or mode, or a union that isn't a plain object, is refused as the config loads, before anything is grabbed.
• The loaded config carries the union settled, as a ResolvedUnion whose own fields are the union it was settled from:
  • a spread or a clone of it is that plain union again;
  • serve's ETag fingerprint still sees what it says;
  • merge.union's type stays what a config writes.
• A settled union is recognized across two installs of the package by a registered symbol, and is settled again when the match's normalizeTitle has changed.

Smaller changes

• genreKey folds Unicode forms together, so a category written decomposed finds its genre in a profile too.
• The text helpers moved to a leaf module, src/merge/text.ts, which removes an import cycle.
• resolveMatch accepts null, and merge: null in a config loads as it did before.

Testing

• New tests for every mode, priority order, languages, the three-source case, backfill and the day boundary, the transform's next, a loaded config's settled union, copies and another install's union, a stale match, serve's ETag, and every refusal.
• Each behaviour was broken in turn to confirm a test fails.
• 1,808 tests, passing on Node 22.18, 24.15 and 26.12.
• The docs example is in test/readme-examples.ts and typechecks.