feat(meeting): meeting-tuned diarization defaults (fix under-segmentation)

[WingedGuardian]

Why

The meeting_bridge built its Speechmatics session from a bare AmbientConfig(), inheriting the ambient bridge's near-field anti-over-split diarization posture (prefer_current_speaker=True, speaker_sensitivity=None). On the first real multi-speaker capture this under-segmented — the transcript collapsed into long merged mega-turns.

What

Give the meeting bridge its own diarization knobs with meeting-appropriate defaults (segment more), env-tunable per room:

Env	Default	Effect
MEETING_PREFER_CURRENT_SPEAKER	false	Allow re-attributing to a different speaker at a boundary instead of sticking with the current one. Set true if a room over-splits.
MEETING_SPEAKER_SENSITIVITY	0.6	Higher ⇒ splits new speakers more readily (Speechmatics range 0–1, SDK default 0.5). Blank/none ⇒ defer to the SDK.

• config.py: two fields + _env_bool / _env_float_or_none helpers mirroring ambient_bridge/config.py.
• session.py: mapped into the existing active_prefer_current_speaker / active_speaker_sensitivity fields that ambient_bridge/active_session.py:_diar_kwargs already consumes (it omits None, so False passes through explicitly and None sensitivity is dropped).
• Tests: config defaults/overrides/None-sentinels + a factory-wiring test proving both values reach the built session (ActiveSession stubbed; built against the real AmbientConfig, so a wrong field name would TypeError).
• Docs: README config table, deploy/meeting.env.example, CONTRACTS.md §1c.

Meeting-bridge only — does not touch the ambient bridge.

Verification

• ruff (--no-cache) clean, compileall clean.
• 42 meeting-bridge tests pass (config + session + server + vad).
• speaker_sensitivity=0.6 confirmed valid against Speechmatics docs (0–1, and 0.6 is their own docs example).
• Diff: +147, 0 deletions.

Notes

• The defaults are a first calibration from one noisy real capture — direction is strongly indicated (inherited ambient tuning under-splits meetings), but the exact values are meant to be refined on real captures via the env knobs (no redeploy needed).
• Won't fix ASR garbling from room-capture audio quality — that's an input-audio limit, separate from diarization.

🤖 Generated with Claude Code