Avoid unnecessary CTC vocabulary inference

[achembarpu]

Why is this change needed?

Vocabulary rescoring currently computes CTC evidence even when no transcript
word is eligible for correction. This adds hasCTCRescoringCandidates to the
rescorer and prepared session so callers can skip that work. It uses the existing
matching rules, including aliases, compounds, multiword terms and thresholds.

Empty vocabulary or timings return false. Acoustic rescue conservatively returns
true for nonempty inputs. Callers that need standalone keyword detections must
still run rescoring. Term forms are also prepared once, and transcript words are
normalized once per request.

Initial measurements on one Apple Silicon Mac reduced unnecessary separate CTC
work from about 140–153 ms to 0.3 ms, and shared-head work from about 12 ms to
0.3 ms. These are component timings, not whole-dictation speedups.

Validation

• 144 selected checks pass, including parity against real CTC evidence across
  12 candidate cases, threshold handling, empty inputs and acoustic rescue.
• Contribution lint, strict lint for changed Swift files and formatting checks
  pass.
• The default Intel build passes, as does Intel Release compilation for macOS 14
  with the optional text-processing trait disabled.

The streaming vocabulary test class is excluded: its first-word replacement
failure also reproduces on unchanged main. The full suite and hosted CI have not
been verified green.

[Alex-Wengg]

Looks correct — early-stop placement matches all three candidate loops and the cached forms/set are equivalent to the old builders. One nit: hasCTCRescoringCandidates returns true whenever spotterRescueEnabled (default on), but rescue only runs on the term-centric path under largeVocabThreshold; gating it with !useBKTree && terms.count <= largeVocabThreshold would let large vocabularies skip CTC too.

[achembarpu]

Thanks for catching that. Fixed in 6c57258: the rescue shortcut now requires !useBKTree and terms.count <= largeVocabThreshold. Larger vocabularies can skip CTC when no text candidates match, even with rescue enabled. Added regression coverage at and above the threshold, including positive candidates and prepared-session parity, and updated the docs. All 103 selected vocabulary checks and formatting checks pass locally.