fix(tests): clean tirith background install noise

[gorovmaria626-hub]

Summary

• Suppress unexpected tirith background installer exceptions so shutdown/closed-stream noise does not leak from daemon threads.
• Stabilize local interrupt cleanup coverage by capturing the spawned Popen object instead of scraping process command lines.
• Keep the PR scoped to files still present on current origin/main; legacy RTK integration test was already removed upstream.

Test Plan

• git diff --check origin/main...HEAD
• python -m pytest -q tests/tools/test_tirith_security.py tests/tools/test_local_interrupt_cleanup.py tests/tools/test_interrupt.py -W error::pytest.PytestUnhandledThreadExceptionWarning

Review

• Independent reviewer passed: no security concerns or logic errors.

[teknium1]

Thanks for targeting the installer-thread noise and the flaky process lookup. The current premise remains: ensure_installed() starts _background_install in a daemon thread (tools/tirith_security.py:710-717), and that target directly calls _install_tirith() without an exception boundary (tools/tirith_security.py:621).

Problems

• Catching Exception would also suppress programming errors, despite the module contract that operational failures are handled while programming errors propagate (tools/tirith_security.py:10-11). It would additionally persist an install-failed marker for those errors.
• The proposed regression test invokes _background_install() synchronously (PR diff), so it does not validate the thread-warning behavior described in the PR.

Suggested changes

• Restrict the boundary to the verified operational teardown failure class(es), leaving unexpected implementation errors visible.
• Exercise the mocked failure through a real background thread and join it before asserting state.

Current main has changed all three touched files since the PR base, including later Tirith work in 07cc567dfa208bc7b1adf59dd1f21243930af319; salvage should reapply the narrow fix against that state.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs touch or reference the local interrupt-cleanup area, but they address distinct causes: #16457 combines Tirith installer-thread exception handling with deterministic subprocess capture, #39496 adjusts platform skips and timeout budget, #43157 primarily adds a self-host-kill guard while also retuning interrupt-test timing, and #49460 implements Windows process-tree teardown that has since landed independently on main.

Related pull requests

• #16457 related — (+49/-39) — keep open, revisions required: capturing the spawned Popen object directly addresses the flaky process-command-line scan, but the Tirith change catches every Exception and its synchronous test does not reproduce daemon-thread warning leakage. The keep_open review on #16457 should be resolved by limiting the boundary to verified teardown failures and exercising it through a real joined background thread.
• #39496 related — (+6/-0) — keep open, revisions required: the 75-second timeout addresses the collision between the test's 30-second cleanup wait and the file-level timeout, but the module-wide POSIX skip would also suppress the newer Windows tree-kill regression. Per the keep_open review on #39496, rebase and scope skips only to POSIX-dependent tests while leaving the Windows test active.
• #43157 related — (+398/-13) — do not merge as-is: the interrupt-test budget reduction targets a real timeout race, but most of the diff is a separate self-host-kill feature whose shell-token and shell -c detection currently introduces documented false positives. Despite the keep_open review on #43157, contributor reviews remain blocking until numeric PID matching is retained without treating the terminal shell's $$ as the Python host or scanning data-position shell tokens as executable code.
• #49460 [closed] related — (+82/-1) — remain closed as implemented on main: it correctly replaced wrapper-only Windows termination with whole-tree teardown and added focused regression coverage, but commit e5253d8 independently landed the same behavior through terminate_pid(..., force=True) and taskkill /T /F. It remains relevant as the original focused reference implementation and explains why Windows coverage must not be skipped by #39496.

Suggested consolidation

Do not merge any PR as-is. Use #16457 as the narrow consolidation target after fixing its exception boundary and threaded regression test; optionally fold in only the rebased, per-test POSIX markers and timeout-budget work from #39496. Keep #43157 separate and blocked on its contributor-reviewed guard false positives, and leave #49460 closed because its Windows tree-kill fix is already implemented on main; none of these PRs is a full duplicate that should be closed solely in favor of #16457.

Cross-PR triage: Reviewed 4 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 35 kB of PR diffs, 7 kB of issue/PR text, 16 kB of discussion (13 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.