Stop telemetry test servers when a test body fails

[MaxGhenis]

When a telemetry emitter test fails, pytest prints its summary and then hangs until someone interrupts it. This fixes that. It changes one test file; no production code changes.

Cause. Two tests in packages/microcosm-build/tests/engine_free/shared/test_telemetry_emitter.py start a server on a non-daemon thread and stop it only at the end of the happy path:

• test_subprocess_exchanges_ambient_token_and_delivers_events runs ThreadingHTTPServer.serve_forever and calls shutdown() only after assert emitter.available and _process.wait(timeout=10). This is the case the NZ hub's review of microcosm#1162 hit (finding 3).
• test_local_socket_acknowledges_after_durable_queue runs EmitterService.run, which serves until it receives a close message. Its _FakeSampler.parent_alive() always returns True, so nothing else stops it.

If an assertion fails first, the thread keeps running. Interpreter exit waits for every non-daemon thread (threading._shutdown), so the pytest process never exits.

Fix. All three tests that start a server thread now use one context manager, _serving_thread(target, stop). The third, test_token_bearing_http_post_does_not_follow_redirects, already had try/finally but used a non-daemon thread. The helper:

• runs target on a daemon thread;
• calls stop() and joins the thread (up to 30 s) in finally.

The HTTP servers are also entered as context managers, so server_close() always runs.

The ambient-token test also kills its emitter service subprocess in a finally if the service is still running after a failure. This is defensive: in the forced-failure runs below the service had already exited on its own both before and after the change.

Invariant. For every test in this file that starts a server thread, on every exit path of the test body (pass or fail):

1. the thread is a daemon; and
2. stop() is called and the thread is joined.

Either one alone keeps pytest from hanging. Two new tests force an AssertionError inside _serving_thread, one for the HTTP server and one for EmitterService. Each then asserts that the thread is a daemon and no longer alive; the emitter test also asserts the socket was removed. Three helper mutants (non-daemon thread, no finally, no stop() call) each fail both new tests (6 of 6).

Before and after. Each run used a watchdog that sends SIGINT if pytest is still alive 30 s after printing its summary (60 s for the full-file runs). The forced-failure copies differ from their source by one injected line.

Run	main	this branch
Whole file (failed on its own under load, see below)	1 failed, 32 passed, hung; interrupted 60.9 s after the summary	1 failed, 34 passed, exited 1.0 s after
Ambient-token test, forced startup_timeout_seconds=0 (emitter.available false)	hung; interrupted 30.7 s after	exited 0.7 s after
Ambient-token test, service started, forced wait(timeout=0.001) (TimeoutExpired)	hung; interrupted 30.4 s after	exited 0.7 s after
Socket test, forced wrong rss_bytes expectation before the close message	hung; interrupted 30.4 s after	exited 0.5 s after
Ambient-token test with startup_timeout_seconds=60 (success path)	n/a	1 passed

Not fixed here. On this machine (load average 35–325) the ambient-token test still fails: "the local telemetry emitter service did not become ready". The service subprocess imports the microcosm.build package, which pulls in torch and pandas. That took 7–8 s against the 3 s DEFAULT_STARTUP_TIMEOUT_SECONDS. CI passes on main, presumably because its runners import faster. That is a separate, partly production issue: busy build hosts can silently lose hosted telemetry. It is queued as a follow-up. After this PR, that failure is reported and pytest exits.

Tests. Only this file: 1 failed, 34 passed (the load-caused failure above), and pytest exits 3.5 s after its summary. ruff check and ruff format --check pass. No changelog fragment, matching microcosm#1162 (test-only).

🤖 Generated with Claude Code