feat(sync): add AsyncSyncTranscriber for asyncio callers

[ccampbell-aai]

Summary

Adds aai.AsyncSyncTranscriber, the asyncio counterpart of SyncTranscriber, closing the gap where the sync product was the only wrapper without an asyncio implementation (prerecorded has Transcriber/AsyncTranscriber, streaming has StreamingClient/AsyncStreamingClient).

• sync/v1/async_api.py — httpx.AsyncClient twin of api.py's transcribe(); reuses the endpoint constants and _error_from_response, so error mapping (problem-details envelope, retry_after, legacy envelopes) stays shared.
• sync/v1/async_client.py — AsyncSyncTranscriber with coroutine transcribe() / warm(). Same input types (path/bytes/file object — no URLs), same SyncTranscriptionConfig, SyncTranscriptResponse, and SyncTranscriptError. Reuses the existing _base.py helpers; path and file-object reads run off the event loop.
• Lifecycle follows AsyncTranscriber: owns an HTTP pool by default (async with / aclose()), or shares a passed-in aai.AsyncClient, which stays the caller's to close. No transcribe_async() — fan-out is plain asyncio.gather.
• Exports wired through sync/v1 and the top-level package; the sync/__init__.py shim stays backwards-compat-only.
• 16 unit tests mirroring test_sync.py plus async-lifecycle cases; README and CLAUDE.md updated (both previously documented this class as a planned follow-up).
• Version bump 0.65.00 → 0.66.00, mirroring the minor bump feat(prerecorded): add AsyncTranscriber for asyncio callers #227 used for AsyncTranscriber.

Test plan

• pytest tests/unit — full suite green (only pre-existing pyaudio-missing extras failures)
• ruff check + ruff format --check clean on changed files with CI-pinned 0.3.5

🤖 Generated with Claude Code

[ccampbell-aai]

Live parity test: threading vs asyncio

Ran a side-by-side comparison of SyncTranscriber (threading) and AsyncSyncTranscriber (asyncio) against a live staging deployment of the sync API — same audio clip and identical configs through both wrappers, comparing the full SyncTranscriptResponse field by field (excluding session_id and request_time_ms, which differ per request by design).

Scenarios (threading side also covers the transcribe_async() Future path; asyncio side fans out with asyncio.gather and both sides exercise warm()):

Scenario	Result
default config, path input	✅ identical
timestamps=True, path input	✅ identical
timestamps=True, bytes input	✅ identical

Transcript text, word texts, and word timestamps were identical between the two wrappers in every run (3 consecutive full passes).

One non-SDK observation surfaced while testing: word/overall confidence values can jitter in the 4th–5th decimal place (observed deltas 4.6e-05 to 7.3e-04) when requests are in flight concurrently. This was isolated to the server, not the wrapper: two identical concurrent requests through the same threading wrapper showed the same drift, while sequential requests were bit-for-bit deterministic every time — consistent with dynamic batching in the inference backend. The comparison treats confidences as equal within a 0.01 tolerance for that reason; text, words, and timestamps compare exactly.

🤖 Generated with Claude Code