[sharpca]

Why are these changes needed?

Application.astream_result() ends a stream in two places: the success path and the exception handler. The exception handler called post_end_stream unconditionally, but pre_start_stream only runs for streaming actions. A non-streaming action that raises therefore fans out post_end_stream for a stream that was never started.

Adapters that keep per-stream state blow up there. LocalTrackingClient.post_end_stream looks the stream up in stream_state -- which only pre_start_stream populates -- so it raised KeyError from inside the handler and replaced the action's own exception with an unrelated one. Because the handler runs while the original exception is in flight, users saw a bare KeyError with no trace of the real failure, even when their application code raised something perfectly readable.

This tracks whether pre_start_stream completed and only fans out post_end_stream when it did. Streams that did start are cleaned up exactly as before, so stream_state bookkeeping is unchanged for the streaming path.

Fixes #999

Related issue number

#999

Checks

• I've successfully run make lint && make test (see notes below on the local equivalent)
• I've added tests where applicable.
• For core features, I have updated the documentation.
• For user-impacting features, I have updated the changelog.

Verification details:

• tests/core/test_application.py -- 143 passed
• tests/tracking/test_local_tracking_client.py -- 17 passed, 1 skipped
• tests/core -- 426 passed; the 9 failures are all in test_graphviz_display.py and come from the optional graphviz package being absent in this environment, unrelated to this change
• black --line-length 100 --check passes

Two regression tests were added:

• test_astream_result_non_streaming_exception_skips_post_end_stream (core): asserts the lifecycle hook is not invoked when the action is not streaming
• test_astream_result_preserves_non_streaming_action_exception (tracking): asserts the action's ValueError reaches the caller with the built-in local tracker attached

[YongGoose]

This makes the same source change as #1000, which was opened on Oct 27 for the same issue (#999): a stream_started flag set after pre_start_stream and checked before post_end_stream.

I applied both on current main (36e3ff3) in a Linux container with Python 3.12. Both fix the repro from #999, and tests/core and tests/tracking pass with either. #1000 also tests that a stream that did start still gets its post_end_stream. It may be worth agreeing on one of the two so review time isn't split.