Thinking Path

• Manual WebUI cron runs execute from a multi-threaded server process.
• The previous subprocess isolation fix kept long cron jobs out of the parent cron/profile lock, but still used fork.
• Forking from a threaded server can inherit held import/logging/FD state into the child with no owning thread left to release it.
• The child entry point is already top-level and picklable, so spawn is the narrow safer process boundary.
• The queue drain-before-join behavior from PR fix: shorten cron profile lock for manual runs #1746 must stay intact because large child results can otherwise deadlock the parent.

What Changed

• Switched _run_cron_job_in_profile_subprocess() from multiprocessing.get_context("fork") to multiprocessing.get_context("spawn").
• Added regression coverage that contrasts a held module-level lock under fork vs spawn, demonstrating the ownership model that makes spawn safe.
• Updated cron subprocess tests so spawn children import an isolated fake Hermes agent/cron package instead of relying on parent-process monkeypatched modules.
• Kept the >100 KB queue drain-before-join regression and adjusted manual persistence tests to mock the subprocess boundary directly.

Fixes #1754

Why It Matters

• Avoids probabilistic deadlocks and inherited-state hazards when manual cron runs are launched while sibling WebUI threads hold import/logging locks or other process resources.
• Preserves the prior large-output subprocess fix while moving to a safer child-process startup model.

Verification

• env -u HERMES_CONFIG_PATH -u HERMES_WEBUI_HOST /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_issue1574_cron_profile_lock.py tests/test_scheduled_jobs_profile_isolation.py tests/test_issue617_cron_profile_selector.py tests/test_cron_manual_run_persistence.py tests/test_cron_run_job_import.py -q → 25 passed in 3.63s
• git diff --check → passed
• env -u HERMES_CONFIG_PATH -u HERMES_WEBUI_HOST /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/ -q → 4659 passed, 2 skipped, 3 xpassed, 1 warning, 8 subtests passed in 382.22s

Risks / Follow-ups

• spawn has a small startup overhead because it starts a fresh Python interpreter; manual cron jobs are long-running enough that this should be negligible.
• The tests now use an isolated fake agent package for spawn child imports, which is intentionally closer to the real import boundary than parent-process sys.modules monkeypatching.

Model Used

• OpenAI Codex gpt-5.5 via Hermes Agent CLI.
• Tool use: Kanban task context, shell/git/pytest, file patching, and GitHub CLI.