feat: honour the model's own silence, and spend it on recall

[Endika]

The vocabulary model has carried a trained __NADA__ class — "nobody is signing", cut by
health_dataset.py from the gaps between LSE-Health's translated sentences — since the
co-articulated retrain. Nothing on this side read it: classify treated it as an ordinary
concept and createGloss humanised it to the word __nada__, so the app wrote 131 of 1,477
words as __nada__ on held-out signers, and health_words.py counted every one of those as a
word written. Over half the pause-babble figure this repo published was the model's own silence.

Nothing is retrained here and lse-vocabulary.bin is byte-identical.

What it buys

Four seeds, floor 850, four held-out signers, 1,060 annotated instances and 211 windows landing
in gaps between annotated sentences.

	before	after
signs written correctly, continuous signing	37.3% (sd 0.9)	44.2% (sd 0.7)
words written into pauses	30.0% (sd 2.4)	19.9% (sd 1.9)
__nada__ reaching the transcript	131 of 1,477	0

Both axes improve at once, which the floor/gate grid had never managed. The recall gain holds in
every one of the four seeds, so it is not the best of four.

The two moves are separate and in that order:

1. Honour the abstention. If the model's silence class wins, nothing is written — not even
   the runner-up, which would contradict the answer the model just gave. Measured, this alone
   takes pause babble from 27.5% to 10.9% on shipped weights with recall unmoved at 38.1%. A
   filtering policy that writes the runner-up scored identically to the word, because under an
   abstention no real concept ever cleared the gate; the abstention is what ships and a test
   pins the difference.
2. Spend the freed room on recall. With babble no longer counting the abstention, the window
   gate comes down from 0.60 to 0.45.

Notes for review

• The class is declared in the manifest as abstentionConcept, not hardcoded in TypeScript, and
  check_manifest.py now fails on any __-prefixed concept that goes undeclared. This bug was
  two places having to agree with nothing checking, which is the shape of the handedness and
  normalisation bugs already recorded in tools/train/README.md.
• The engine's own floor drops to 0.30 so it stays below the gate. Otherwise both floors sit at
  0.45, the stabilizer veto becomes unreachable and a real diagnostic dies. The effective
  threshold a word must clear is 0.45 either way, which is exactly what the bench measured.
• abstention is a fourth WindowVeto. The model answering "nobody is signing" is an answer,
  not a threshold, and reporting it as classifier sends anyone reading the panel after a floor
  that was never involved.
• The About panel said 287 signos when one of those 287 is the abstention. Now 286, and
  check_manifest.py subtracts the abstention so it cannot drift back.

What this does not do

No learned abstention for the remaining babble, and tools/train/README.md says why with the
measurement: the class is precise (argmax over a real sign 0.9%, never above 0.60) but of the
pauses where the app was writing a real word, P(__NADA__) > 0.30 in 0.0% of them. Abstention
and remaining babble are disjoint. The negatives that would fix the rest are not labellable in
this corpus — 101 annotated sign types against the 90-150 signs a minute of fluent discourse — and
with 23 of 211 windows left and ±6 points of sampling error, a further 3 or 4 points could not be
demonstrated here even if it worked.

No segmenter changes: floor, grace, window boundaries and DominantHandTracker are untouched.

Verified: npm run lint, npm run typecheck, npm test (205 tests, 11 new), the 33 Python bench
tests, check_manifest.py, and npm run build.