[jinon86]

Why

The owner approved node-by-node prune-expired cleanup on #1468 (2026-10-23), but ccc-erasure-apply.py refuses every node: a read-only fleet scan (12 nodes, origin/main planner fed over stdin, no remote writes) found 7–110 unclassified files per node, 408 normalized patterns. 88 of them are harness artifacts that appeared after #2066; 21 are sensitive backup variants the group-b patterns miss (so they would be kept forever).

What

• 7 new classes (schemas/memory-artifact-inventory.v1.json, 50 → 59 incl. splits): nunchi.gotgan_store (body-bearing, handled like mempalace.palace), nunchi.jev_review_ledger, nunchi.feed_receipts, nunchi.feed_status, state.skill_listing_policy, bot.jev_skill_advice_config, skill.review_zai_env (live credential config → always retain).
• Existing classes extended: nunchi batch logs / runtime locks / seen-state / runtime state, state runtime locks / rotated logs / installer records / distill runtime / self-update state / skill-autosave state / skill manifests.
• Group-b widened, anchored: .env.bak.* / .env.bak-* / .env.pre-*, sessions.json.bak-*, crontab.bak[.-]* (bot dir, ~/.claude/state, ~/.nunchi), crontab-backup-<digit>*.txt; new state.memory_facts_backups; new state.cert_backups (kept by the key-file rule + explicit retain).
• Deliberately NOT group b: bare sessions.json.bak / tasks.json.bak — they are the bridge's previous-good recovery copies (bridge/session/store.py, bridge/core/task_ledger.py), now registered as live.
• Planner: sqlite_sidecars flag — -wal/-shm/-journal follow their DB (live DB ⇒ sidecars live; in a retention class a sidecar shares the DB's verdict). Prevents a WAL under a live DB from aging out on its own.
• Built-in key-file protection unchanged. Docs + changelog fragment.

Verification (re-run independently)

ccc_erasure_retention_test.py 32/32 (was 22) · ccc-erasure-planner.test.sh 50/50 (was 43) · ccc-erasure-apply.test.sh OK · ccc-erasure-handoff.test.sh 16/16 · ccc-security-audit.test.sh 18/18 · ruff/shellcheck clean · canon-node-name-scan baseline. Two new tests verified to fail when the sessions pattern is made greedy or the sidecar live-claim is removed.

Local plan on one node: unclassified 84 → 48; the only delete targets are group-b backups (.env 3, sessions 4, crontab 3). Nothing deleted.

Still blocking (owner policy pending on #1468)

Operator one-off files (ad-hoc scripts, handover notes, one-time logs), producer-less files, and backups the inventory already marks as owner decisions (wiki-candidates.md.bak*, working-state.md.bak*). Each node's actual run still needs its own approval.

Refs #1468

🤖 Generated with Claude Code

[jinon86]

Added: operator-scratch fallback class (owner decision 2026-10-23, option 1) — commit 4384c279

• state.operator_scratch: fallback, not a pattern — a file directly under the harness state dir that matches no other class becomes scratch instead of a blocker (files matching a real class keep it). Kept 90 days (max(mtime, ctime); env may only lengthen), then prune-expired / node-decommission → delete through the same apply boundary + per-node approval. Not targeted by audience/user/cache requests.
• Scope: state dir only. ~/.nunchi (private memory) and the bot data dir (credentials) get no fallback — unknown files there stay blockers.
• Scheduler guard: a scratch file referenced by the crontab, the agent-cron task store, or systemd unit/drop-in files (abs/real/~/$HOME/%h forms or bare name) is retain (scheduler-referenced). Any unreadable source → every scratch file retain (scheduler references unreadable) + body-free scheduler_guard status in the plan.
• Rule order: live claims → key-file → symlink → scheduler guard → age. Orphan SQLite sidecars stay blockers.
• Extra class state.timed_test_notify_state (live notify state written by timed_test_deadline_scan.py, would otherwise age out as scratch). Inventory text that called *.bak copies "owner-decision blockers" updated.

Verification (re-run independently): retention 43 OK (11 new) · planner 57/0 · apply OK · handoff 16/0 · security audit 18/0 · ruff clean · changelog ok · canon-node-name-scan baseline. New tests fail when the fallback, name matching, or the guard is disabled (19 / 2 / 5 failures).

Local read-only plan (one node): blockers 48 → 2 (one in ~/.nunchi, one in the bot dir); scratch 45 = 4 eligible now · 40 retain-until (earliest 2026-10-27) · 1 scheduler-referenced · 0 key files.

Known limit (safe default): ccc-erasure-apply.py refuses a run whose targets are not owner-only; 24/45 scratch files here are 0644/0755, so an apply would currently be refused. Left as-is (fail-safe); to be decided before the first per-node run.

[seoseo-ai]

Approved after exact-head validation and fresh operator authorization using the relay-held seoseo-ai credential.