fix: remove Wav2Vec2 speech model

[numnx]

What changed

• Removed Wav2Vec2 from the speech catalog, local ONNX runtime, adapter contracts, and model-selection UI fixtures.
• Retained Whisper Base and Whisper Tiny as the supported local transcription models.
• Migrated removed or unknown local STT selections to Whisper Base across system, project, and sprint settings, including persisted raw overrides.
• Preserved external transcription configuration and structured client errors for unknown local STT/TTS model IDs.
• Updated canonical and public speech documentation.

Why

The lightweight CTC checkpoint produced unreliable conversational microphone transcription. Whisper is materially more accurate for the supported local speech workflow, so retaining the weaker runtime path added complexity without a dependable user outcome.

Validation

• pnpm run ci
• Backend: 5,275 passed, 7 skipped; coverage thresholds passed.
• Dashboard: 2,106 passed, 2 skipped.
• Dependency audit: no known vulnerabilities.
• Production build: passed.
• Live /health and /ready: passed; dashboard root loaded on port 4444.
• Repository scan confirms no Wav2Vec2/CTC implementation references remain.

Risk and rollback

Risk is limited to users who had selected the removed local model; their persisted selection is migrated to Whisper Base while external speech settings remain intact. Roll back this PR to restore the old catalog/runtime path.