feat(devx): three surveyed VOCABULARY rows for the teaching-corpus claim ratchet

#13582 shipped `scripts/check-corpus-claim-drift.mjs` with exactly one row and a
self-test assertion pinning that, because a row is only as good as the survey of
legitimate usages someone did before adding it. This is the follow-up card that
pays three such surveys, one candidate at a time and none of them batched.

- `exists-portability` — the `(NoSQL)` gloss. A DIFFERENT claim from key
  presence (portability, not presence), hence a row of its own. Survey: 8 NoSQL
  hits over 7 corpus files; two of them sit within 4 lines of `$exists` and BOTH
  are legitimate — the repaired "portable, not a NoSQL-only operator" sentence
  and a skills column header — so the claim is the parenthesised gloss and an
  unnegated "NoSQL-only" / "MongoDB-only", never the word. 0 baseline entries.
- `regex-retired` — the retired `$regex` spelling. Survey: 5 hits over 3 corpus
  files, ALL legitimate retirement notices, so the claim is the live-operator
  phrasing and never the mention. 0 baseline entries.
- `section-visiblewhen-unbound` — #13532's section-binding claims. Survey: 88
  `visibleWhen` hits over 21 corpus files, 0 firing at any window from 4 to 40.
  Window 9, measured on the pre-repair tree: it is the smallest width reaching
  the binding-root table row, and 9..40 yield an identical ledger. 0 baseline
  entries.

The one-row cap moves to an EXACT ENUMERATION of the shipped row ids in order,
never a floor, with a discrimination leg proving a smuggled row still fails it.
Every row ships its pre-repair fixtures, its repaired/legitimate green legs and
a discrimination leg. The baseline file is untouched and byte-identical.

Part of #13745

Co-authored-by: Claude <noreply@anthropic.com>

Author: claude