workflow-fix: janitors miss issue-keyed caches outside canonical globs

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised during a chat-mode disk-full diagnosis (2026-07-02, boot disk at 95%; emitting agent: orchestrator).

Goal

Widen disk-janitor coverage to issue-keyed re-downloadable caches outside the canonical data/issue_<N>/{hf_dl,g*_dl} layout (/tmp/i<N>_*, /tmp/issue-<N>*, data/issue<N>_*, data/issue_<N>_dl, VM /workspace/.cache) with terminal-status reap plus active-issue escalation, store/ never touched.

Workflow gap

• Bug observed: Boot disk hit 95% while the guard reported only ~80M reclaimable: ~90G of issue-keyed HF-mirror caches sat in non-canonical locations (/tmp/i779_mirror_dl 59G, data/issue779_hfstage 15G, data/issue_744_dl 11G, /tmp/fact_check_823 6G, /workspace/.cache 22G) invisible to both janitors.
• Why it is a workflow gap: clean_experiment_downloads.py and vm_disk_guard.py tier (b) sweep only the canonical data/issue_<N>/{hf_dl,g*_dl} globs, so any run that stages HF downloads under /tmp, a non-canonical data/ name, or a pod-style /workspace path on the VM escapes both the Step-8 cleanup and the disk-guard backstop — the fleet accumulates until the disk wedges.
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

clean_experiment_downloads.py:
+ CACHE_GLOB_PATTERNS += [
+   "/tmp/i{N}_*", "/tmp/issue-{N}*",          # tmp staging dirs
+   "data/issue{N}_*", "data/issue_{N}_dl",    # non-canonical data/ names
+ ]  # same safety contract: store/ + eval_results/ never touched
vm_disk_guard.py tier (b):
+ include the same issue-keyed patterns; terminal-status -> reap,
+ active-status -> escalate (sidecar + Telegram), never delete active data
+ add a VM /workspace/.cache pass (pod-style HF cache duplicated on the VM)

All local copies verified to be partial downloads of folders that exist on the HF data repo (issue779_monitoring/, issue744_token_continuity/ confirmed via list_repo_tree 2026-07-02) — deletion is pure re-downloadable cache loss.

Scope / surfaces

• Primary target: scripts/clean_experiment_downloads.py, scripts/vm_disk_guard.py
• Grep the workflow surface for the canonical glob pattern before editing (grep -rln 'hf_dl' scripts/ .claude/ CLAUDE.md) and update every janitor touchpoint; list them in the plan.
• Consider also making writers converge: the deeper fix may be a shared "issue cache root" helper so runs stop inventing /tmp/i<N>_* staging paths in the first place — planner's call.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• NEVER delete store/, eval_results/, or any active issue's data (escalate instead).
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; CLAUDE.md § Disk hygiene stays consistent with the implemented behavior.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: scripts/clean_experiment_downloads.py, scripts/vm_disk_guard.py
• fingerprint: 558746fb1c07

Raised as a prose candidate during the 2026-07-02 disk-full diagnosis (boot disk 95%, 458G used; janitor-visible reclaim ~80M vs ~90G actually sitting in non-canonical issue-keyed caches).