fix: retain WordNet adjective satellites and key synsets by offset+POS

[aaronstevenwhite]

Closes #11.

The reported bug

The WordNet converter discarded every adjective satellite. parse_data_file filtered on synset.ss_type == pos, and the adjective pass runs with pos="a", so all 10,717 ss_type="s" synsets in data.adj, roughly 59% of WordNet's adjectives, were dropped while the pointers referencing them were still emitted. Any consumer following adjective relations hit ~10,720 dangling targets.

Why the one-line fix wasn't enough

Two things I verified against raw WordNet 3.1 shaped the design:

1. Pointer POS is never "s". Scanning every pointer in all four data files: all 10,717 & (similar-to) pointers and all 7,856 + pointers that target satellites carry pos == "a". "s" appears only in ss_type. So a naive (offset, ss_type) key would leave every pointer-to-satellite dangling. The key must normalize s to a, which is exactly the consistency the issue's "Note for consumers" asks for.

2. Offsets don't identify synsets. They're byte offsets into a per-POS data file, so they collide across files: 117,791 synsets share only 117,458 distinct offsets, giving 333 collisions. glazing keyed synsets by bare offset throughout, so WordNetLoader silently overwrote real synsets, WordNetSearch.add_synset raised ValueError on them, and every pointer resolved POS-blind.

New Synset.key / Pointer.target_key pair the offset with a normalized POS ("00014377a"), and the loader, search index, and relation traverser key on that. Satellites index under both "s" and "a", so an adjective query returns heads and satellites together while ss_type still reports "s".

Satellite sense enrichment needed the same treatment or the 10,717 new satellites would have arrived with no sense_number/tag_count: enrichment built a head-less sense key (lemma%5:LL:II::), but real satellite keys carry a head_word:head_id suffix data.adj doesn't record, and the head-less prefix is ambiguous for 2,966 of 20,384 satellite keys. Lookups now key on the synset offset from index.sense, which is exact.

Also fixed, found while verifying

• Seven more silently-dropped synsets. LEMMA_PATTERN required an alphanumeric first character, so .22_caliber, 'hood, 'tween, 's_Gravenhage and friends failed lemma validation. A failed lemma discards the entire synset.
• glazing search relations --dataset wordnet was broken outright, failing with Unknown format code 'd' for object of type 'str' (synset IDs built via f"{synset.offset:08d}" against a string offset). The CLI swallowed it into "Relation search failed". Confirmed broken on main, working here.
• Stale caches never re-converted. The .initialized marker was contentless, so upgraded users kept outdated data forever, which for this release means a WordNet cache missing all 10,717 satellites. It now records a schema version.

Verification against WordNet 3.1

	before	after	raw
adjective synsets	7,465	18,185	18,185
satellites	0	10,717	10,717
total synsets	n/a	117,791	117,791
distinct keys	n/a (333 collisions)	117,791, 0 collisions
dangling pointer targets	~10,720	0
satellite words enriched	0	20,384 / 20,384

Every raw synset in all four data files now converts. 1424 passed; ruff, mypy (checked against 1.18.2 and CI's latest 2.3.0), and all pre-commit hooks clean. 29 new tests in tests/test_wordnet/test_satellites.py cover the key primitives, satellite emission, cross-POS collisions, satellite pointer round-trips, reachability from pos="a", and the enrichment case prefix-matching would get wrong.

Breaking change

WordNetLoader.synsets, lemma_index, and the four relation indices are keyed by SynsetKey ("00001740n") rather than bare offset. This is unavoidable, since a bare-offset mapping cannot be correct given the collisions. get_synset still accepts a bare offset and returns None when one is ambiguous; pass pos or a canonical key to disambiguate. Version bumped 0.2.3 to 0.3.0 accordingly.

Upgrading: existing converted data is missing all satellites and can't be repaired in place. Run glazing init to re-convert from already-downloaded sources; no re-download needed.