feat(sync): make the live connection the only one the client opens

[aurpsis-aai]

Stacked on #244 (ccampbell/sync-live-session), which is itself stacked on #241. This PR adds only the consolidation on top.

What

Every entry point — transcribe(), transcribe_async(), transcribe_live(), open_live() — now posts to /v1/transcribe/live. There is one connection in this client.

Audio that is already complete is a stream whose bytes happen to all be ready, so it goes out over that connection as a single chunk rather than through a separate buffered request. What that buys is one request shape to reason about: one framing, one set of errors, one timeout — and a complete clip reaches the service exactly the way a live capture does, with each speech segment transcribed as it lands instead of after the upload finishes.

The endpoint constant becomes ENDPOINT_TRANSCRIBE_LIVE = "/v1/transcribe/live", following the rename in DeepLearning #20044 (merged and live in production).

The public surface is unchanged

api.transcribe() keeps its exact signature and delegates to transcribe_live(), so the legacy assemblyai.sync_api shim still re-exports it and test_old_sync_api_module_surface_is_preserved still passes. ENDPOINT_TRANSCRIBE stays exported for the same reason, with a comment noting nothing here requests it.

I went this way deliberately rather than deleting the buffered function: that test pins the old flat module's surface as a compatibility contract, and breaking it would be a major-version change for no gain — the goal is one connection, not a smaller import surface.

Note this is a client-side consolidation only. Sync still serves both /v1/transcribe and /v1/transcribe/live, and both stay.

One observable change

An empty config used to mean no config part at all. The live endpoint requires one ahead of the audio, so {} now goes out instead. Anyone inspecting the wire will always see a config part. Two tests asserting the old absence are updated to assert the presence and the ordering.

Complete-audio calls also move from sync_http_timeout to sync_live_http_timeout, whose per-operation shape is the right one for a streamed body. sync_http_timeout still backs the warm-up.

Compressed formats are not a concern here: this client has always been WAV/PCM only (_resolve_audio picks between audio/wav and audio/pcm), which is exactly what the live endpoint accepts.

Validation

• 531 unit tests pass (528 before, 3 new), re-run against the current base after each rebase. Verified the baseline first: the base branch is 528/528 on the same interpreter, so every failure I saw along the way was mine to fix.
• New tests pin the invariant: transcribe() and its async twin hit the live URL with no Content-Length and config-before-audio, and one test drives all four entry points and asserts the set of requested URLs is exactly {/v1/transcribe/live} — so a buffered request reappearing anywhere fails.
• ruff check and ruff format --check clean on assemblyai/; mypy reports zero errors in sync/v1 (the errors elsewhere are pre-existing and identical to the base).

No version bump

The stack lands on master as one merge, so the version is set once at the top rather than per PR — this branch leaves __version__ alone. The test matrix and version-bump workflows only trigger for PRs into master, so only lint runs here until the base moves.

🤖 Generated with Claude Code

[aurpsis-aai]

Heads-up @ccampbell-aai — I pushed 55240d7 on top of your floor fix rather than replacing it, because the async-callback form doesn't survive the oldest pytest_httpx.

On pytest-httpx 0.20 (the pin for the py311-httpx0.22 env), _httpx_mock.py:184 does response.is_stream_consumed = False on whatever the callback returned. An async callback returns a coroutine, so it raises AttributeError: 'coroutine' object has no attribute 'is_stream_consumed' before the response is used. Reproduced both locally and in CI on your commit — run 34517828459, 5 async tests plus 2 "mocked but not requested" from the response queue count.

Mine reads the body inside an httpx.MockTransport handler instead, which consumes the request stream the way a real transport does and needs nothing version-specific — the approach test_sync_live.py's module docstring already recommends for body-shape assertions.

Verified 531 pass on the modern stack and 531 on the httpx 0.22 floor via sdk-test-oldhttpx. Your is_reusable diagnosis was right and is kept, just expressed as a loop sized to the four entry points the test drives.

Happy to take a different shape if you prefer — the constraint is just that it can't rely on pytest_httpx consuming an async body.