night fixes 2026-10-05: R-867/R-95/R-863..R-869 closed, R-870..R-876 opened (R-876 P2: repair misses dpkg's journal); Part E crash record; Part F spike; golden 0.294.0; rulings 100-103, CC decisions 104-108; STATUS

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS

documentation/backlog/CLOSED-ITEMS.md

@@ -26,6 +26,19 @@
 
 ---
 
+## 2026-10-05 (day) — the night's fixes: off-site clean-up guard, first-install image race, R8 download, a killed pass's report (controller v0.294.0, agent v0.144.0 + v0.144.1, golden 0.294.0; rulings 100–103, CC decisions 104–108)
+
+| Row | What | Closed | Evidence |
+|---|---|---|---|
+| **R-867** | **The off-site clean-up guard refused honest 7-day retention and mailed an error every window.** Controller v0.294.0: the guard's "young" line is keep-daily CALENDAR days, built from the same constants as the policy (decision 104); every other refusal kept. Tests run restic 0.14.0's policy itself (proven identical to the binary over 92 snapshots), 5 red-proofs. Live by hand: demo-felhom window 5 16 → 14, demo-hp window 6 145 → 127 — exactly the predicted snapshots; hub rows `pruned`, no event, no mail, key files clean. **Reasoning kept: a line that can sit inside the keep window is a line that refuses the honest case — derive it from the policy, never pick an age.** Full text: `git show 7221ee5c:documentation/backlog/OPEN-ITEMS.md`. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partA/` |
+| **R-95** | **The box could delete its own off-site history.** Append-only key since 2026-10-03 (decisions 68–69); the last open item — a clean-up window that actually removes snapshots — was observed 2026-10-05 on both demo boxes (R-867's live windows, opened by the operator's one-shot grant; the dated check's four conditions all hold). The unattended weekly window is the same code path and is due ~2026-10-11/12; it is not separately re-checked (the DUE-CHECKS entry is removed with this row). Residual: R-822 (an add-only attacker steering older keeps). Full text: `git show 7221ee5c:documentation/backlog/OPEN-ITEMS.md`. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partA/`; `audits/offsite-lock-build-2026-10-03/` |
+| **R-863** | **A new box's first app install could fail: the one-time image clean-up deleted the image compose had just pulled.** Controller v0.294.0: every compose command that can pull holds a shared lock (`dockerexec.BeginImageWork`); a clean-up pass takes it exclusively without waiting and otherwise does not run; the one-time pass is retried every 2 min and writes its marker only after a pass that ran (decision 105). Live on 9202: the clean-up fired at minute 3 (05:31:36 UTC) inside BookStack's 35.5 s install, skipped itself, BookStack installed first time; the retry at 05:33:36 ran and kept everything; teardown through the product. The update path was already guarded (`Updating`); restore/undo were exposed in principle and now hold the same lock. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partB/` |
+| **R-864** | **A failed compose logged the head of stderr (pull progress) and cut the reason.** Controller v0.294.0 `tailStr` (rune-safe), pinned with the night's real first line. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partB/red-proofs.txt` |
+| **R-869** | **The move-aside log line printed an empty destination.** Controller v0.294.0: assigned before the log line; test + red-proof. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partD/r869-red-proof.txt` |
+| **R-865** | **R8 measured every download as 0 B (`--print-uris` with `-s` prints no URIs).** Agent v0.144.0: no `-s`; the fake answers like real apt (verbatim 9202 output), 2 tests, red-proof. Live: the installed wrapper's `download_bytes` on demo-hp read **12 802 456 B** for 13 pending upgrades (0 before), nothing installed. A live R8 REFUSAL line was not produced: it needs < 500 MB free on demo-hp's guest, i.e. 28.5 GB written into a thin pool with 18.7 GB free — it would have stopped every guest. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partC/` |
+| **R-866** | **The debug OS pass could not run with the hub away.** Agent v0.144.0: the daemon saves the hub's block; the selftest falls back to it and says `block=SAVED(<time>; hub unreachable: …)` in its header (decision 107). Live on demo-felhom with the hub blackholed: pass ran (`nothing` ×3); its reports, kept on disk, reached the hub at the next agent start. | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partD/r866-*` |
+| **R-868** | **An OS pass whose agent was killed installed but never reported.** Agent v0.144.0: the wrapper keeps its report beside the plan; the agent sends kept copies and deletes them (decision 106). **Measured live NOT to work** (05:45 UTC, demo-hp): the wrapper died on a broken stderr pipe before saving. Agent v0.144.1 (decision 108): the wrapper survives a dead reader; the agent looks again every 5 min. Live again (06:04 UTC, demo-hp, the A5 shape): wrapper `DONE upgraded=13`, copy kept, the hub got ONE `applied` report (13 packages) at 06:09:05 by the 5-minute look. **Reasoning kept: a root helper whose reader can die must never write unguarded to that reader — the log is not the outcome.** | CLOSED 2026-10-05 — FIXED | `audits/night-fixes-2026-10-05/partD/` |
+
 ## 2026-10-04 (night) — the config bundle, test approvals, the Docker-socket self-heal (agent v0.143.0, hub v0.133.0, controller v0.293.0, installer 1.31.0; rulings 96–99)
 
 | Row | What | Closed | Evidence |