fix(cron): parallelize script jobs with workdirs

[soria-clawd-bot]

Summary

• pass no-agent cron workdirs to subprocess cwd instead of process-global os.chdir
• keep agent workdir jobs serialized while allowing script-only watchdogs to run in parallel
• record scheduled_at, dispatch_started_at, and dispatch_lag_seconds in saved run artifacts

Proof

• HERMES_HOME=$(mktemp -d) EMAIL_HOME_ADDRESS= uv run pytest -q tests/cron — 673 passed
• uv run ruff check on touched files — passed
• git diff --check — passed

[alt-glitch]

This was generated by AI during triage.

Related: this lands in a saturated competing cluster on cron/scheduler.py _run_job_script workdir handling — #21397 (earliest-open, general workdir→subprocess-cwd), #42274 / #56005 (no_agent-specific), #60475 (general). All open. This PR uses the same general mechanism (thread workdir into subprocess cwd, drop the inert process-global os.chdir) plus a parallel-pool serialization change and dispatch-lag artifacts. Not marking a duplicate — flagging the cluster so a maintainer can pick one approach.

[teknium1]

Thanks for isolating the no-agent cwd to the subprocess; the premise is present on current main: cron/scheduler.py:2535-2549 mutates the process cwd, and tick() serializes every workdir job at cron/scheduler.py:3640-3647.

Problems

• cron/scheduler.py:3349 in this PR derives scheduled_at from next_run_at. That is not valid for recurring external-scheduler fires: cron/scheduler_provider.py:100-105 claims before reloading the job, while cron/jobs.py:1708-1710 and 1743-1747 advance next_run_at during that claim. The proposed artifact can therefore record the next future occurrence and a zero lag instead of the occurrence that fired.

Suggested changes

• Keep the subprocess-cwd/pool change focused, and split the dispatch metadata; alternatively carry the pre-claim scheduled time explicitly into run_one_job().
• Cover the CronScheduler.fire_due() recurring-job path if retaining the metadata.

Automated hermes-sweeper review.

[teknium1]

next_run_at is not reliably the occurrence being dispatched. CronScheduler.fire_due() claims before it reloads the job, and claim_job_for_fire() advances recurring next_run_at; this will save a future scheduled time and zero lag for external recurring fires. Preserve the pre-claim fire time explicitly, or split this telemetry from the cwd fix.