fix(ai): preserve Gemini 3.8 TTS WAV output

[opencode-agent]

Why

Google's September 22 Gemini 3.8 TTS release changes the unary default from headerless PCM to WAV (audio/wav) with a RIFF header. The Google Speech route already reads the response MIME type, but always attaches info.format: "pcm" and pcm_s16le encoding even when the bytes are WAV. That can make callers wrap a WAV file as if it were raw samples.

What changed

• Classify Google speech output from the returned MIME type: WAV gets WAV container metadata, while existing PCM responses keep their raw-sample metadata.
• Reject explicit raw-PCM requests for Gemini 3.8 unary generation before sending, since this route does not select a non-default output format. Do not silently return WAV as requested PCM.
• Add deterministic fixture-driven tests for WAV classification and both Gemini 3.8 TTS model variants. Existing recorded Google PCM generation/streaming tests remain unchanged.

Checks

• bun test test/speech.test.ts test/provider/google-speech.recorded.test.ts --timeout 30000 (7 passed)
• bun typecheck in packages/ai
• Scoped Prettier and oxlint checks

No provider credentials or live inference were used. This PR does not add a Voices API or change model-catalog metadata.

Requested by: @nexxeln (nexxel via Slack)