fix(inference): claim word alignment for xai/stt-2

[muntazir868]

Part of #7467 (item 3 of 3 — the word-alignment gap). Items 1 and 2 (keyterm forwarding, smart_turn) need gateway-side xai.ExtraSettings support: the gateway currently rejects those keys with failed to decode extra settings: 'xai.ExtraSettings' has invalid keys, so they can't be fixed from this repo alone.

Summary

• Add xai/stt-2 to XaiModels so it is a first-class inference.STT model in the type union (STTModels).
• Add xai/stt-2 to _WORD_ALIGNED_MODELS. The issue verified 48 timed words come back through Inference for xai/stt-2, same as xai/stt-1, but the capability claim still reported False.

What the claim actually gates

Correcting the issue description: the claim does not control adaptive interruption — interruption detection no longer consults STT alignment (decoupled from word timestamps by #6765). What it does feed:

• AudioRecognition receives it from agent_activity.py:883 / :1319, and when the provider omits speech_end_time, audio_recognition.py:1147-1155 derives it from the word timings. Without the claim, xai/stt-2 fell back to a cruder speech-end estimate despite returning timed words.
• stt.FallbackAdapter reports the combined claim as "word" only if all fallback models claim alignment, so a single unaligned model degrades the whole adapter's claim for downstream consumers.

Testing

• make check (ruff format + lint, mypy strict) — clean
• uv run pytest --unit --audio_eot --ignore=tests/test_room.py — 4012 passed, 3 skipped
• New assertions in tests/test_inference_stt_aligned_transcript_claim.py:
  • xai/stt-1 and xai/stt-2 both report "word" as the primary model
  • new param case aligned-xai-stt2-fallback: fallback="xai/stt-2" keeps the adapter's claim at "word"

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)