preserve-numbers: dropped year masked by an unrelated same-value digit

[s-annam]

Problem

checkNumbersPreserved's drop check (src/lib/webllm/preserve-numbers.ts) has a known masking gap for year-claimed atoms specifically. Every 4-digit number in 1900–2099 auto-classifies as a "year" claim in bareIntegerClaim, with no surrounding-context gate at all (unlike headcount, which needs a management verb or a people noun). That means a coincidental, unrelated 4-digit number in that range — a suite number, a model number, a version — is itself always "claimed," so it can mask a genuinely dropped year.

Repro (from src/lib/webllm/preserve-numbers.test.ts, "does NOT extend to year" test, added in #874's revision round):

checkNumbersPreserved(
  ["Founded the program in 1900.", "Operated out of suite 1900."],
  ["Founded the program.", "Operated out of suite 1900."],
)
// -> { ok: true, dropped: [], added: [] }

The founding year 1900 is genuinely dropped, but checkNumbersPreserved scores it clean because the unrelated, unchanged "suite 1900" shares the same num:1900 key and (per the current classifier) is always claimed.

Why this wasn't fixed in #874's revision round

The equivalent bug for headcount claims was fixed (count-aware presence check — a NEW unclaimed occurrence of a key still counts as a legitimate reword, but a pre-existing, unrelated unclaimed occurrence doesn't mask a drop). That mechanism doesn't transfer to year: because every year-shaped digit is unconditionally claimed, there's no "unclaimed" bucket for the count-aware guard to compare against — the fix is a no-op for this claim kind.

What a real fix looks like

Give year classification a context gate, mirroring headcount's PEOPLE_VERB_PREFIX / PEOPLE_NOUN_FOLLOW: require a temporal cue near the digit (e.g. "in", "since", "during", "by", "from", a preceding/following year-range dash) before classifying a bare 4-digit number as a year, falling back to unclaimed otherwise. This needs its own false-positive audit — a badly-scoped gate would create the missed-legitimate-year-reword failure mode that headcount's design docblock (preserve-numbers.ts module docblock, rule 1) explicitly warns about.

Pointers

• src/lib/webllm/preserve-numbers.ts — bareIntegerClaim, checkNumbersPreserved, module docblock (rules 1 & 3)
• src/lib/webllm/preserve-numbers.test.ts — describe block "a dropped headcount is not excused by an unrelated pre-existing unclaimed digit (#874 review)", in particular the "does NOT extend to year (documented residual)" test
• Originally flagged in PR fix(rewrite): reject/revert rewrites that drop or invent numbers (#778) #874's review (fix(rewrite): reject/revert rewrites that drop or invent numbers (#778) #874 (comment)), fixed for headcount only in bfc8cf7