fix(cron): kill process group on script timeout, fix mislabeled failure message (#59549)

[webtecnica]

Summary

Two bugs in cron/scheduler.py's script execution path:

Bug 1: Mislabeled timeout messages

The error classifier in _format_cron_failure matches "timed out" generically before checking the specific source. Script timeouts (which produce "Script timed out after Ns: /path") contain "timed out" and are reported as provider timeout instead of script timeout. Fix: check for "script timed out" before the generic branch.

Bug 2: Orphaned child processes on timeout

subprocess.run(timeout=N) only kills the direct child when the timeout fires; grandchildren spawned by the script survive as orphaned processes. Migrated to subprocess.Popen + preexec_fn=os.setsid (POSIX) to create a fresh process group, and on TimeoutExpired the entire group is killed via os.killpg(pid, 9).

Changes

File	Δ
cron/scheduler.py	+21/-9 lines (Popen migration, process group, error message fix)

Tests

• tests/cron/test_cron_script.py: 38/38 passed ✅

Closes #59549

[teknium1]

Thanks for the focused cron fix. The two reported defects are present on current main: cron/scheduler.py:75-79 classifies the script runner's Script timed out ... result (cron/scheduler.py:2133-2134) as a provider timeout, and the runner uses subprocess.run(..., timeout=...) at cron/scheduler.py:2101-2109.

Problems

• In the PR timeout handler, cron/scheduler.py:2048 limits cleanup to sys.platform != "win32". The Windows branch returns after proc.communicate(timeout=...) without terminating or reaping proc, so this migration drops direct-child timeout cleanup on a supported platform. Windows 10/11 is Tier 1 in website/docs/getting-started/platform-support.md:20.
• The PR adds no regression tests. Current tests/cron/test_cron_script.py:184-196 only checks that a timeout is reported; it does not cover descendant cleanup or the delivery classification.

Suggested changes

• Add a Windows direct-child terminate/kill-and-reap fallback while retaining the POSIX process-group cleanup.
• Add focused cleanup and classifier regression tests.

Automated hermes-sweeper review.

[teknium1]

This branch skips every termination action on Windows and immediately returns the timeout result. Please add a Windows direct-child terminate/kill-and-reap fallback here; otherwise the Popen migration leaves the timed-out script process unmanaged on that platform.

[kshitijk4poor]

Closing as superseded — both bugs this PR targets are now fixed on main: (1) the mislabeled "provider timeout" for script failures was fixed by #85536 (commit d1fc204) — classification is gated on job mode (provider_reachable = not job.get("no_agent")) before any substring match, and script timeouts get their own message; (2) process-group kill on script timeout landed via #93795 (#85125 Phase 4a, merged 2026-08-27), which tree-kills descendants through agent.deadline.kill_process_tree and closed #59549/#71148. Your PR had the right ordering insight (script-timeout check BEFORE the generic matcher) — thanks for both fixes.