Zombie job records accumulate in state.json until SessionEnd #15

Summary

Related to but distinct from the dead-PID detection gap: plugins/opencode/scripts/lib/state.mjs caps at MAX_JOBS = 50, pruning by updatedAt. A crashed job whose status is still "running" has a fresh-looking updatedAt from when it last reported progress, so it survives pruning and occupies a slot. Over a long session this leads to the /opencode:status "running" list filling with zombies while real completed jobs get pruned first.

Unlike codex-plugin-cc, opencode-plugin-cc does not currently block new tasks on the presence of a running job (no resolveLatestTrackedTaskThread guard in handleTask), so the impact is lower than the upstream report — but the state file still drifts.

Suggested fix

1. Pair this with the dead-PID check in issue-to-be-filed (#164-equivalent) so zombies get demoted to failed in enrichJob / buildStatusSnapshot before pruning runs.
2. Consider a separate "zombie sweep" pass inside upsertJob that checks every running row's PID when the state is rewritten.

Upstream reference

Derived from openai/codex-plugin-cc#202.