feat(soniox): populate SpeechData.words with per-word timing and confidence

[FurlanLuka]

What

The Soniox websocket API returns per-token confidence, start_ms and end_ms, but the plugin currently keeps only the per-transcript mean confidence — the per-token metadata is dropped. This PR groups consecutive sub-word tokens ("walk" + "ing") into words and exposes them as TimedString entries on SpeechData.words, following the Deepgram plugin's precedent (offset applied to times, start_time_offset attribute set).

Word confidence = mean of its tokens' confidences; word timing spans first token start_ms → last token end_ms.

Why

Word-level ASR confidence enables downstream consumers to detect likely mistranscriptions. Our production use case (language-learning voice agent): the agent issues grammar corrections against phrases the learner never said — ~16% of corrections traced to ASR artifacts. A plausible-in-context mishear ("walking" → "working") is invisible to text-only downstream logic but shows up as a per-word confidence dip. The scalar per-transcript mean (added recently) is too coarse: one garbled word barely moves the mean of a long utterance.

Word-boundary rules

• Words split at whitespace (Soniox marks word starts with a leading space in token text) — correct for 57 of the 60 documented supported languages.
• The three unspaced scripts in Soniox's catalog — Han (zh), Kana (ja), Thai (th) — never produce that whitespace, so token boundaries are used as word boundaries there (Soniox emits roughly word/character-granularity tokens for these scripts).
• Punctuation tokens attach to the preceding word, matching Deepgram's behavior.

Open question for Soniox folks: is a leading space in token["text"] a guaranteed word-start marker for spaced languages? If the server could emit an explicit boundary flag, the script-range heuristic here could be deleted.

Notes

• Purely additive — text, scalar confidence and timing fields are untouched; words was previously always None for this plugin.
• In translation mode, words follows the same side as text/confidence (the target side).
• On INTERIM_TRANSCRIPT/PREFLIGHT_TRANSCRIPT (merged_speech_data), a word straddling the final/non-final seam may appear split until the endpoint — same instability interim text already has.
• speaker_id per word is first-token-wins, consistent with the transcript-level field.

Testing

8 new tests in tests/test_plugin_soniox_stt.py (existing fake-WebSocket harness): sub-word grouping + confidence mean, start_time_offset application, Han/Thai token-boundary words, NOT_GIVEN fallbacks when metadata absent, reset() clearing, merged final+non-final words, and end-to-end FINAL_TRANSCRIPT assertions. ruff check, ruff format and scripts/check_types.py (mypy) all pass.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}