feat(soniox): report output tokens in RECOGNITION_USAGE for token-based billing

[Panmax]

What

Soniox bills real-time STT by tokens: input audio tokens (driven by stream duration) plus output text tokens — each element of the tokens array in a WebSocket response is one billable output token (pricing).

The plugin previously reported only audio_duration in RECOGNITION_USAGE events, so RecognitionUsage.output_tokens / STTMetrics.output_tokens (which the framework already supports) were always 0, making it impossible to track token-based costs from metrics.

This PR counts final transcription tokens as they arrive and reports the increment alongside the audio duration in each RECOGNITION_USAGE event.

Counting rules

• Only is_final tokens are counted — non-final tokens are re-sent as final later, so counting them would double-bill.
• <end> / <fin> control markers are excluded (not transcription output).
• In translation mode both original and translated tokens count, as both are billable provider output.
• Reporting is incremental and counters reset on reconnect, mirroring the existing _reported_duration_ms behavior.

Input audio tokens are intentionally not reported: the WebSocket API exposes no input token count, and audio_duration already carries the billing-relevant quantity (Soniox charges input tokens for the full stream duration) without resorting to estimation.

Testing

Added 4 unit tests to tests/test_plugin_soniox_stt.py following the existing fake-WebSocket style:

• final tokens counted, <end> excluded
• non-final tokens excluded (no double-billing)
• incremental reporting across two endpoints
• translation mode counts both sides

pytest tests/test_plugin_soniox_stt.py --plugin soniox: 28 passed (24 existing + 4 new). ruff format --check, ruff check, and strict mypy all pass.

🤖 Generated with Claude Code

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.