Background job state machine retains zombie entries; `--fresh` does not clear them

Summary

The background-job state machine (used by /codex:rescue, /codex:status,
/codex:cancel) appears to retain completed/failed job entries
indefinitely. Re-invoking with --background --fresh does not clear or
recycle these zombie entries, so subsequent runs can see and act on stale
state.

Reproduction

Environment:

• Windows 10 / 11
• @openai/codex-plugin-cc 1.0.5

Steps:

1. Run a Codex rescue task to completion via
   codex:rescue --model gpt-5.5 --effort high --background.
2. Wait until it shows status: completed.
3. Start a new task with --background --fresh.
4. Run /codex:status — the previously-completed job is still listed,
   and in some cases its job id is reused or its state bleeds into the
   new run's status reporting.

Suggested fix

• Auto-GC completed / failed job entries after N minutes (e.g. 30 min).
• Have --fresh actively wipe terminal-state job entries before starting
  the new run, not just decline to resume.

Happy to clarify with logs if useful — the state files live under
~/.claude/plugins/data/codex-openai-codex/state/....

Inspect the state files under ~/.claude/plugins/data/codex-openai-codex/state/... and trace how --background, --fresh, /codex:status, and /codex:cancel handle terminal states. Reproduce the rescue sequence from the issue, then verify that fresh runs do not retain completed or failed entries and that status no longer reports them after the cleanup period.