fix: dictionaries plugin — inflected forms (plurals, tenses) resolve to their lemma definition

[symbiodev]

Bug

Looking up a plural noun (or other inflected form) in a built dictionary only ever shows a stub like "plural of cat" — never the actual definition of cat.

Root cause

The kaikki/Wiktionary source tags inflected-form entries (cats, ran, found, …) with form_of/alt_of and a gloss that is just the stub text ("plural of cat", "simple past of find"). build-dictionaries.mjs wrote that stub as the entry's only definition. Since the reader's dictionary engine resolves a word via a direct .idx hit before it ever falls back to .syn synonyms or stemming (see docs/dictionary.md in crosspoint-reader — "1. Direct match" wins over "3. Stemming"), the stub permanently shadowed the real definition. Stemming, which would have found the lemma, never got a chance to run.

Fix

writeStardict() now resolves form_of/alt_of pointers at build time:

• A headword whose senses are purely references (no own sense) becomes an alias: its .idx row is folded onto the lemma's .idx row — same (offset, size), zero extra bytes in the .dict.dz.
• A mixed headword (own sense and a form-of sense, e.g. "found" = "to establish" + "simple past of find") keeps its own text and appends the lemma's definition after it.
• Alias chains ("runnin'" → "running" → "run") resolve to the material end; a chain that loops keeps its stub rather than resolving to nothing.
• A form-of entry whose lemma is missing from the source (dead end) is left as-is — the stub is still better than nothing.

Rebuilding english with this fix: 632k of 1.39M headwords (45%) turned from a bare stub into a real, readable definition, at the same total dictionary size (the aliasing adds index rows, not dictionary bytes).

Testing

• test/dictionaries.test.mjs (new): runs the full build pipeline against a synthetic kaikki source (--source id=file://..., new flag) and asserts on the resulting .idx/.dict.dz — pure aliasing, mixed headwords, dead-end stubs, alias chains, and that aliasing still respects the existing case-fold merge (marched → march/March group).
• npm test: 20/20 passing (existing suite unaffected).
• Manually verified against a full rebuild of english (1,385,197 words, 59.3 MB, same as before the fix) loaded onto a device: cats/ran/went/found/children all now show the lemma's definition; mice (whose form_of target isn't in the source under that exact key) still shows its stub as a fallback.

Also touches: --source/--catalog-dir CLI flags (test-only, let the suite build into a scratch dir against a local fixture instead of the network + committed catalog), and a fileURLToPath fix for CATALOG_DIR/the new test's script path so both survive a working directory containing a space.

Related — firmware feature requests

While tracking this down I also confirmed two things worth a firmware-side feature request on crosspoint-reader (not part of this PR, filing separately there):

1. Select more than one word at once in DictionaryWordSelectActivity — useful for phrasal verbs and set phrases that aren't in the dictionary as a single headword.
2. Look up a word from inside the definition screen (DictionaryDefinitionActivity) — chaining a lookup when the definition text itself uses an unfamiliar word.

Neither is reachable from the plugin side; both need firmware changes. Filing as issues on crosspoint-reader and will link them here once open.

[symbiodev]

Firmware feature requests filed, as mentioned above:

• Select more than one word at once: Dictionary: allow selecting more than one word at once (phrasal verbs, set phrases) crosspoint-reader/crosspoint-reader#3268
• Look up a word from inside the definition screen (chained lookup): Dictionary: allow looking up a word from inside the definition screen (chained lookup) crosspoint-reader/crosspoint-reader#3269

[itsthisjustin]

I think this is an issue for crosspoint not for the dictionary downloader right?