fix(cli): keep after_run callbacks when /run client disconnects

[CodeAlex52]

Fixes #7394.

Problem

When a client disconnects from /run, the disconnect monitor cancels the worker task with a bare Task.cancel(). The runner treats a bare cancellation as an external cancel, which deliberately skips after_run_callback (see test_run_async_cancellation_does_not_execute_after_run_plugin), so a plugin that finalizes a run — writing a terminal event to the session, telemetry, cleanup — gets no callback at all for a disconnected run.

That path already exists: cancelling with the caller-closed-early marker makes the runner treat the run as "the caller stopped consuming events", which does run the after-run callbacks. /run_sse gets this for free because it closes the generator; the non-streaming /run endpoint is the only path that cancels without the marker.

Fix

One line in cli/api_server.py, plus the test that pins it:

worker_task.cancel(_CALLER_CLOSED_EARLY_MSG)

I deliberately did not change the bare-cancellation semantics in runners.py / workflow/_node_runner_utils.py. Flipping those would contradict the existing intentional behavior and its test, and it would change what happens for genuine external cancellations (shutdown, a caller cancelling its own task). The disconnect case is a caller going away, which is exactly what the marker means, so the fix belongs at the call site.

The error path is unaffected: the generator still re-raises CancelledError, the worker task still ends cancelled, and /run still answers 499.

Tests

Added test_agent_run_disconnect_marks_cancellation_as_caller_closed_early next to the existing disconnect test in tests/unittests/cli/test_fast_api.py, reusing the same harness and asserting the CancelledError that reaches run_async carries the marker.

Validation:

• The new test fails against the current code (worker_task.cancel() reaches run_async with e.args == ()) and passes with the change.
• The four pre-existing disconnect tests pass unchanged both with and without the change.
• tests/unittests/test_runners.py, tests/unittests/plugins, tests/unittests/workflow: 1589 passed, 1 skipped, 5 xfailed.
• tests/unittests/cli/test_fast_api.py: 176 passed with the change vs 175 on the unmodified tree; the 6 failures and 5 errors in that file are pre-existing in my environment (metrics/agent-identity tests and optional-dependency collection errors) and are identical before and after.
• pyink --check clean on both files.

Note for discussion

@Li-john1021 commented on #7394 proposing to restore after_run_callback for any external cancel of Runner.run_async. That is a different, larger behavior change and conflicts with the deliberate semantics noted above — happy to discuss, but I think the disconnect path is the actual gap worth fixing first.