fix(cron): run no_agent scripts in workdir without process-global chdir

[xy200303]

Summary

• Re-scoped per the review on the previous revision: per-job cron profile support was deliberately removed in revert(cron): remove per-job profile support (#28124) #43956, so the original profile-home race (cron scheduler: profile-job context bleeds into concurrent non-profile job (script not found) #39886) no longer exists on main — _hermes_home is read-only now and cron runs per-profile inside a profile-scoped gateway.
• This revision fixes the remaining process-global bleed of the same bug class that does exist on current main: the no_agent + workdir path in run_job applied the workdir via a process-global os.chdir() outside _terminal_cwd_lock. A sequential-pool workdir job can overlap parallel-pool jobs, so the chdir bled the workdir into concurrently running workdir-less jobs for the whole script run.
• Additionally, _run_job_script hardcoded subprocess.run(cwd=<script dir>), so the chdir never even reached the script — workdir silently had no effect for no_agent jobs despite the comment claiming it set the subprocess cwd.

Fix

• _run_job_script gains a cwd parameter, passed to subprocess.run(); when omitted the script runs in its own directory, preserving prior behavior for the agent pre-run / wake-gate call sites.
• run_job's no_agent path passes the job's workdir as the script subprocess cwd and drops os.chdir entirely — zero process-global state.
• A workdir that vanished on disk now logs a warning and falls back to the script directory, matching the agent path's behavior.

Tests

• New TestNoAgentWorkdir in tests/cron/test_cron_workdir.py (4 tests):
  • workdir reaches the script subprocess as its cwd;
  • process cwd is not mutated while a workdir job's script runs concurrently (reproduces the bleed — fails on the old code);
  • missing workdir warns and falls back to the script dir;
  • no-workdir default is unchanged.
• Verified the new tests fail against the pre-fix implementation; full tests/cron/ suite passes.

[teknium1]

Thanks for isolating the original module-global race. This patch cannot be applied to current main as a standalone correctness fix.

Problems

• Merged commit 7d8d000b1921cb1c14aa137a41bd96203992e2c0 (revert(cron): remove per-job profile support (#28124) #43956) deliberately removed per-job cron profile support, including _job_profile_context, the scheduler profile override, job-record handling, CLI/tool inputs, and tests/cron/test_cron_profile.py.
• Current main documents the different model in cron/jobs.py:54-63: cron storage is per-profile, and a profile-scoped gateway executes that profile's jobs under its active HERMES_HOME. Current cron/scheduler.py:543-552 has no per-job profile context.
• Consequently, this scheduler-only change would revive behavior for legacy profile fields that current interfaces intentionally ignore; its test target no longer exists on main.

Suggested changes

• If per-job cross-profile execution is still desired, please re-scope it against the current per-profile scheduler architecture and restore the full API/storage/validation contract with end-to-end coverage rather than cherry-picking this old slice.

Automated hermes-sweeper review.

[xy200303]

Thanks for the review — agreed that the previous revision targeted machinery that #43956 deliberately removed. I've re-scoped the PR against the current per-profile scheduler architecture instead of restoring any per-job profile contract.

While auditing cron/scheduler.py on current main, I found the remaining process-global bleed of the same bug class: the no_agent + workdir path in run_job applied the workdir via os.chdir() outside _terminal_cwd_lock. Since sequential-pool workdir jobs can overlap parallel-pool jobs, that process-wide chdir bled the workdir into concurrently running workdir-less jobs — and because _run_job_script hardcoded subprocess.run(cwd=<script dir>), the chdir never even reached the script, so workdir silently had no effect for no_agent jobs.

The revised patch passes the workdir through to _run_job_script(cwd=...) so the script subprocess gets it as its cwd with zero process-global state, and adds regression coverage (including a concurrency test that fails on the old implementation). Happy to adjust if you'd prefer a different shape.