feat: tts and output text and audio metrics

[YiminW]

No description provided.

[github-actions]

PR Review: TTS Output Text and Audio Metrics

Good overall structure — the change follows the established pattern from cosy_tts_python and the three version bumps are correct.

Bug: rime_tts counts empty/whitespace text in character metrics

In rime_tts/extension.py, the new call is placed without an empty-text guard:

self.sent_tts = True
# Track character metrics  <- added here
self.metrics_add_output_characters(len(t.text))
await self.client.send_text(t)

Unlike the other two extensions, rime_tts.request_tts does not filter empty text before this point. The t.text.strip() == "" check later in the method is a downstream edge-case for a different code path and does not protect the metrics call.

Compare with:

• google_tts_python: call is inside if t.text.strip() != "": (correct)
• qwen3_tts_python: call is after if not t.text or not t.text.strip(): ... return (correct)

Fix: add a guard before the rime_tts call:

self.sent_tts = True
if t.text.strip():
    self.metrics_add_output_characters(len(t.text))
await self.client.send_text(t)

---

Minor: inconsistent None-safety for send_usage_metrics argument

google_tts_python passes a safe fallback:

await self.send_usage_metrics(self.current_request_id or "")

qwen3_tts_python and rime_tts both pass self.current_request_id directly. In qwen3_tts this is safe due to the early if self.current_request_id is None: return guard in _complete_request, but in _handle_tts_audio_end (rime_tts) the guard is only if self.request_start_ts is not None:, which does not formally guarantee current_request_id is non-None. If the base-class send_usage_metrics does not accept None, this could surface as a runtime error on unusual teardown paths.

Suggestion: use self.current_request_id or "" in both places for consistency with google_tts_python.

---

Missing test coverage for the new metrics

The existing test_metrics.py files in both google_tts_python and rime_tts only exercise TTFB. The new functionality — metrics_add_output_characters and send_usage_metrics (usage metrics emission at end-of-request) — is untested. qwen3_tts_python has no test directory at all.

Consider adding:

1. An assertion that output_characters accumulates correctly across multiple text chunks per request.
2. An assertion that a usage metric is emitted exactly once per completed request (and not for empty-text skips).
3. A basic test suite for qwen3_tts_python mirroring the cosy/rime pattern.

---

Everything else looks good

• Placement of metrics_add_recv_audio_chunks (immediately after accumulating bytes, before sending audio data) is consistent with the reference cosy_tts_python pattern in all three extensions.
• send_usage_metrics is correctly placed before finish_request in all three completion paths.
• Character metrics are correctly scoped to the non-empty text path in google_tts_python and qwen3_tts_python.
• All three manifest.json version bumps are appropriate for the behaviour change.