[Proposed] Automatic cleanup agent (scheduled audit + sweep)

From EXPERIMENT_QUEUE.md, added 2026-04-16

Infra task. Currently cleanup (tracking-file audit, orphaned figures, stale draft detection, unuploaded-model sweep, disk-usage alerts, model-weight cleanup after upload) is manual — I run it when asked, or the user notices drift. Want an agent that runs on a schedule without explicit prompting.

Scope per run:

• (a) run python scripts/pod.py health --quick across all pods; flag low-disk or unreachable
• (b) run python scripts/pod.py sync models --sweep to upload any unuploaded checkpoints
• (c) run python scripts/pod.py cleanup --all --dry-run and surface candidates
• (d) AUDIT-mode scan: orphaned eval_results/ dirs not in INDEX.md, drafts > 3 days unreviewed, figures not referenced from RESULTS.md / drafts, stale queue entries
• (e) produce a summary report to research_log/cleanup/YYYY-MM-DD.md

Execution options:

• (i) scheduled cron via the schedule skill (daily, e.g. 09:00 local)
• (ii) /loop in auto mode
• (iii) manual via /cleanup slash command

Start with (i).

Safety: read-only scan + dry-run by default; auto-apply only INDEX.md additions, stale flags, figure moves to figures/unsorted/. Anything destructive (deletes, model cleanup) requires user approval.

Dispatch target: implementer to build the agent + scheduled task config. No gate-keeper — infra.

Compute: 0 GPU. Each run ~2-5 min agent time.

Priority: MEDIUM (nice-to-have; drift doesn't compound dangerously at current pace but is accelerating as the project grows).