fix(inflekt): correct -ves singularization (derives -> derife) and audit rules against a 104k-word dictionary

[pyramation]

Summary

singularize('derives') returned derife (and olives -> olife, valves -> valf, solves -> solf, thieves -> thiefe), because inflekt delegated straight to the inflection npm package, whose Rails-derived rule is ([^fo])ves$ -> $1fe. That output is what named Derife in constructive-db's generated SDK/CLI.

The fix is a rule change, not a word list: -ves normally just drops the s, and the f/fe stems are the exception, so the rule is inverted and the exception is enumerated:

// src/rules.ts
GENERIC_SINGULAR_RULES = [..., [/ves$/i, 've']]   // derives -> derive, valves -> valve
F_STEM_PLURALS = [['knife','knives'], ['wolf','wolves'], ['leaf','leaves'], ...]  // both directions

Pluralization is symmetric: only an F_STEM_PLURALS word takes -ves; every other -f/-fe word takes a plain s (cafe -> cafes, roof -> roofs, belief -> beliefs), which inflection also got wrong.

inflekt's own rules now run before the upstream fallback (previously it only patched Latin suffixes and *base/ss afterwards), so upstream's wrong answers never surface. Compound identifiers transform only the final segment (user_derives -> user_derive, UserDerives -> UserDerive) and case is preserved.

Dictionary audit. Rather than fixing one word, singularize/pluralize were swept over the full 104k-word wamerican dictionary using the dictionary itself as the oracle (output must be a real word; plural→singular→plural must round-trip). That surfaced whole broken classes beyond -ves: aches -> ach, pies -> py, shoes -> sho, analyses -> analyasis, police -> polouse (the [m|l]ice character class also matches |), and -is/-us singulars (iris, analysis, bus, status) being stripped as if plural. Rule-level fixes cover the general cases; the residue that is genuinely lexical (cookies -> cookie not cooky, ties -> tie, invariants like jeans, gas, atlas) is a generated table:

• scripts/generate-exceptions.ts reads /usr/share/dict/words, uses the committed __fixtures__/word-frequency-usa.txt to break dictionary ambiguities (cookie vs cooky), and emits a pair only where singularizeByRules(word) !== dictionaryExpected. It imports from src/rules.ts (suffix-only) rather than the public API — comparing against the exception-aware singularize made generation self-referential and idempotent-empty.
• src/exceptions.ts is the output: 747 pairs, byte-identical across consecutive runs.

inflekt was not in the CI test matrix, so its tests never ran; added.

Link to Devin session: https://app.devin.ai/sessions/60b2dc2dcb71403f9388d9785a9a78cf
Requested by: @pyramation

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring