feat(morphology): substantivized noun tags, with a typed MISC column

[noomorph]

Three items from a bug report against @interslavic/morphology, plus what follows from the third.

Singular accusative of substantivized adjectives

declineAdjective(...).singular.acc holds three genders: masculine (an animate/inanimate pair), neuter, feminine. Every other singular case merges masculine with neuter and leaves the feminine second. declensionSubstAdj read the accusative as if it were merged too, so:

• neuter took the masculine column, giving srědnji for srědnje (-ego) and šampansky for šampansko;
• feminine took the neuter column, giving prěměnno for prěměnna (-oj) and krivo for kriva.

Now the neuter accusative repeats the nominative and the feminine takes index 2, so those forms come out srědnje, prěměnnų, krivų. Nine lexicon entries move.

Vocative of num.subst. and num.fract.

A numeral that delegates to declineNoun built its cases without a voc key, and CASES had no name for it either, so inflectNumeral('četverka', 'num.subst.') never emitted četverko. Both tags now carry the slot.

Cardinals do not get one. A vocative is the case of address, and šesť, though it declines like the i-stem kosť in the oblique cases, is a quantifier. četverka and četvŕtina are ordinary nouns and take the case as such.

*.subst. noun tags

The lexicon annotates substantivized adjectives as ordinary nouns plus a bracketed genitive hint, učeny (-ogo). The tags m.anim.subst., m.subst., f.subst., n.subst. now work on their own, in both numbers: parsePos marks the noun substantivized and the declension is read off the lemma.

Hardness is what the lemma has to yield. Masculine and neuter endings carry it (pųtujųći and srědnje against dežurny and dobro); the feminine -a and both plural endings (dobre beside srědnje) do not, so for those it is read off the last consonant of the stem. That step is exact, not a guess: y cannot follow ž š č c in the orthography, so an adjective with a soft stem is spelled soft and declines soft.

A lemma that ends as no adjective ends gets no hint, whatever the tag claims. dom under m.subst. would otherwise be handed the genitive *doogo and a whole paradigm built on it; instead the gender half of the tag is honoured and subst. is dropped.

*.pl.subst. yields the same paradigm as the bracketed m.pl. (-yh), which two lexicon entries already use. A test asserts the two routes agree on nine lemmas, and ten paradigms are snapshotted, veliki among them: its hint comes out -ego, and the adjective tables still return velikogo, because establish_root drops the soft marker after k g h.

Declension=Adj in MISC

These entries stay NOUN. The UD guideline that keeps the ADJ tag covers adjectives that exceptionally head a nominal phrase, which is a statement about tokens in running text; here the lexicon has already made the call, giving prěměnna (35526) its own id and citation form beside the adjective prěměnny (35525). Tagging them ADJ would put two ADJ lemmas on one paradigm and promise a three-gender declension we do not emit, and it would drop the gender the dictionary supplies.

That leaves nothing in UPOS or FEATS to record that the forms come from the adjective tables. UD has no universal feature for substantivization, so the tokens carry Declension=Adj in MISC. AdapterResult gains an optional misc, and inflectWord merges it under whatever the caller passes.

The marker is computed rather than carried: NounParadigm holds cases and nothing else, so the bracketed genitive is put to one shared predicate twice, by declineNoun, which routes on it, and by inflectNoun, which reports it. The two plural entries in the lexicon, drobne and prisęžni, pick the marker up for the first time.

A typed MISC column

The key above started out as Derivation=Adj, which was already taken: derive writes the code of the rule that produced a token there, v_prap or adj_adv, and website/src/main.ts reads it back as a label. Record<string, string> accepted both meanings on the one key.

Misc now lists the attributes the column carries, with no index signature: SpaceAfter from the CoNLL-U spec, and LID, SID, Derivation and Declension from this project. An unknown key is a type error, and a misspelled one gets the intended name suggested. Spec attributes nothing in the repository sets are left out until something sets them.

DerivationCode moves from derive/virtual-id.ts into @interslavic/conllu, beside the key it fills, and loses its trailing | string. That member had been collapsing the whole union to string, so Derivation was untyped in effect as well as in name. virtual-id.ts re-exports it.

The same change fixes a latent bug: formatMisc filtered empty values but not undefined, which optional properties make reachable, so an attribute spread in from another bundle would have printed as Key=undefined. formatFeats already guarded this.

Also

• The numeral adapter kept its own copy of the case map. It imports the noun one.
• docs/conllu-conventions.md listed one MISC key out of five, and had never listed Derivation. It lists all of them now, and records that the golden dump covers sid, lemma, form and FEATS but not MISC, which is why a MISC change moves snapshots and nothing else.

yarn test, yarn lint and yarn build pass; snapshots and the golden lexicon dump are updated.