SPIKE R-95: the safety net cannot be seen from the box - and that RAISES the urgency

READ-ONLY STUDY. No code, no version, no image, no golden. No delete verb was issued against any
live store. ep0, DooPlex and Peti's box were not touched at all.

Q1 FIRST, AND IT DID NOT GO THE EXPECTED WAY. The brief supposed a working seven-day snapshot net
might bound the worst case. Measured on BOTH boxes over their own SFTP credential, with a positive
and a negative control on each: NO .snapshots is visible to either sub-account - not in the account
home, not inside the repo - and the account is jailed at /. Either none exist or a sub-account
cannot see them, and the second is not a reprieve: a snapshot the box cannot see is one the box
cannot restore from, so recovery would be an operator act at the Hetzner panel.

The register's claim rests on nothing that was checked. The row has NO R-number, so nothing can cite
it; its "confirm tomorrow" was 2026-07-27, 36 days ago; and the DUE-CHECKS block built for exactly
this (R-341) is EMPTY. R-95's word "ARMED" is withdrawn pending R-429. The confirming field is a
Hetzner API field, so this spike STOPPED at the section 11-D fence and left it for Viktor - ten
minutes in the panel, and it re-ranks everything.

Q2: TEN verbs, not nine. `check` was missing from the brief's list; `dump` is not a verb (it is a
progress phase constant) and was withdrawn. There are TWO `forget --prune` sites - offbox.go:1388
AND offbox.go:1759 - and disarming one without the other reproduces R-191 exactly.

Q3 (documented, from our own API mirror): AccessSettings has five booleans and `readonly` is the
only permission axis. No append-only. So the PBS shape does NOT transfer - PBS is a server that can
refuse; a Storage Box is a filesystem that runs nothing.

Q5 (measured, faithful append-only model, both controls passed first): withdrawing delete does NOT
wedge the store - restic treats a dead owner's lock as stale and proceeds. The constraint everyone
feared is not the blocker. But `unlock --remove-all` printed "successfully removed locks" while the
lock was still there, and resticStep's crash-lock self-heal is built on that call - R-430.

Q6 (measured, with a control): restic 0.14.0 DOES speak rest:. Append-only is a rest-server flag,
not a restic one.

Q7 (measured): detection is nearly free. snapshot_count already reaches the hub and the hub APPENDS
reports, so the history to compare against is already on disk.

RECOMMENDATION: answer Q1 today (Viktor, ten minutes), then build detection, then move retention off
the box. Defer the transport change until Q1 is answered.

Register: OPEN 179 -> 181. Filed R-429, R-430; R-95 updated and kept OPEN. 07 row 10's status is
deliberately UNCHANGED.

documentation/backlog/OPEN-ITEMS.md

@@ -224,7 +224,7 @@ unconditional promise, which CAMPAIGN-11 §7 step 7 measured the customer-facing
 | **E-2d** | **Prove E-2 on a fresh VM** — a real `felhom-host-install.sh` 1.22.0 run, Case B naturally, a claimable customer, then add a drive (the offer) and unplug it (`backup_target_absent` end-to-end) | **CLOSED — PARTIALLY PROVEN** (2026-07-29) | — | **C1, C2 proven** (`audits/E2D-fresh-vm-2026-07-29.md`); **C3, C4 proven live** (`audits/SESSION-C-2026-07-29.md`); **C5 FAILED → R-116** — the gate fires and an alarm reaches the hub, but it is the generic event, so the alarm and its recovery cannot be paired. **R-116 is the single named open leg**; per the Session-C runbook §9, decided in advance, a failed claim closes the item as partially proven rather than triggering a re-run. Both audits carry the full record — the `local-lvm` fence, the ISO/PAIRING derivation, the Phase 0 answers, the per-claim observables and the teardown evidence — and are the place to read it, not this cell. **The arc's actual definition of done is R-106 + R-109, R-108 and D5**, none of which this detour touched | CC |
 | **R-121** | **A BOX's installed agent can sit releases behind the vouched one and nothing notices — the R-120 gate does not cover it.** demo-hp ran agent **0.113.0** while the hub vouched **0.116.0**, through the whole R-116/R-117 arc, and no signal existed on any channel | **READY (S) — NEW 2026-07-30** | — | **Fourth instance of the drift family** (R-111 golden's agent 17 releases behind, R-115 built+deployed but never published, R-120 golden a controller behind — and now installed-vs-vouched on a live box). **Confirmed at source that R-120's gate cannot catch it:** `hub/internal/web/configs.go:1165-1169` compares `goldenVer` against `store.NewestReportedControllerVersion()` — it is a **golden-artifact vs fleet-CONTROLLER** check and says nothing about the agent installed on a box. **`MinAgent` does not cover it either:** it is used to HOLD the controller floor for a box whose agent is too old (`hub/internal/api/handler.go:530-538`, `store.go:1857`) — protective, not an alarm — and demo-hp's 0.113.0 **equalled** `min_agent` 0.113.0, so even a floor comparison was satisfied. **The cost, measured:** R-117's whole subject is the R-113 conjunction, which landed in **0.114.0** — so the designated drill host could not exercise the code under investigation at all, and the R-117 spike had to route every predicate result through an out-of-repo probe built from `main` instead of the installed agent (`audits/SPIKE-r117-bind-liveness-2026-07-30.md` §1, §2.3). Discovered because the R-117 task made bringing the box current an explicit prerequisite. **Fix shape (not implemented):** the hub already receives `AgentVersion` on every host report, and already has semver comparison in Go — the missing piece is a checker comparing reported agent vs the **vouched** agent and surfacing it, operator-tier. Note the honest tension: a box legitimately lags between publish and deploy, so this wants a staleness window rather than an instant alarm | CC |
 | **R-118** | **An absent drive's union row advertises the ROOT filesystem's capacity as its own.** In the absent-state payload the registry-union row reports `total_bytes: 49675956224 / used_bytes: 4584579072` — **byte-identical to the `local` row** (`durable_id: path:/var/lib/vz`, i.e. `pve-root`) in the same response. The real drive is **4 GB** | **READY (XS) — NEW 2026-07-30** | — | Cause: `statfsCapacity(d.MountPath)` (`disks.go:335-338`) statfs's `/mnt/cel`, which with the device gone is a **bare directory on the root filesystem**. `observe.go:176-183`'s comment warns about exactly this trap and guards the Observe path ("*an unmounted removable dir-storage's mountpoint reverts to a bare directory on root … catastrophic DR mis-id*"); **the union path has no equivalent guard.** **Not a DR mis-id** — `durable_id` on that row is still the correct `uuid:…`, so re-attach identity is safe. It is a **false capacity** reaching every consumer of `total_bytes`/`used_fraction` (fill monitors, storage cards): a detached 4 GB drive advertises 46 GiB at 9.2 % used. Same class as `role.go:180-181` — an absent drive's fields decaying to the root filesystem's. Evidence: `audits/DIAG-r116-disks-payload-2026-07-30.md` §12 | CC |
-| **R-95** | restic offsite credential **can delete** (`readonly=False`, `forget --prune` runs from the box); SFTP cannot express append-only | **READY** | — | Root exposure still open. Mitigation now ARMED — split prune off-box or move to REST `--append-only` | CC |
+| **R-95** | restic offsite credential **can delete** (`readonly=False`, `forget --prune` runs from the box); SFTP cannot express append-only | **READY** | — | Root exposure still open. Mitigation now ARMED — split prune off-box or move to REST `--append-only` | CC  **SPIKE 2026-09-01 — `audits/SPIKE-r95-offsite-delete-2026-09-01.md`. THE WORD "ARMED" ABOVE IS NOT SUPPORTED AND IS WITHDRAWN PENDING R-429:** no `.snapshots` is visible to either box's sub-account (measured, both machines, with controls), so the seven-day bound is unverified and unverifiable from the product side. **Q3 (documented, `hub/internal/hetznerapi/hetznerapi.go:38-45`): the sub-account API has ONE permission axis, `readonly` — there is no append-only, so the PBS shape does NOT transfer** (PBS is a server that can refuse; a Storage Box is a filesystem that runs nothing). **Q5 (measured): withdrawing delete does NOT wedge the store** — restic treats a dead owner's lock as stale and proceeds — so the constraint everyone feared is not the blocker; but `unlock --remove-all` lies about success (R-430), and the crash-lock window is UNKNOWN. **Q6 (measured): restic 0.14.0 DOES speak `rest:`** (control: `banana:` → invalid backend), and append-only is a rest-server flag, not a restic one — reachable, but it needs a machine in the recovery path and ep0 is protected. **Q7 (measured): detection is nearly free** — `snapshot_count` already reaches the hub and the hub APPENDS reports, so the comparison needs no box change. **RECOMMENDATION: answer R-429 first (Viktor, ten minutes), then build detection, then move retention off the box; defer the transport change.** Two `forget --prune` sites must be disarmed together — `offbox.go:1388` AND `offbox.go:1759` — or R-191 repeats. |
 | **R-191** | **Every weekly offsite backup UPLOADS successfully and then FAILS the job on a prune the box is deliberately not allowed to do — on both demo boxes.** Measured on demo-felhom 2026-08-04 06:49–06:53: the upload completed (223 s, 629 MiB of 1.874 GiB, **67.2 % reused incrementally**), then `ERROR: prune 'ct/9201': proxmox-backup-client failed: Error: permission check failed - missing Datastore.Modify\|Datastore.Prune on /datastore/felhom-offsite/demo-felhom` → `ERROR: Backup of VM 9201 failed - error pruning backups` → `TASK ERROR: job errors`. The hub raised `whole_guest_backup_failed` | **CLOSED — SHIPPED 2026-08-04** (installer **1.25.0**; both live boxes corrected) | — | **This is R-89's rule not reaching the config.** R-89 moved PBS pruning SERVER-SIDE — *"boxes set `keep_last: 0`, ep0 runs prune jobs; box tokens stay write-only, never widen the grant"*. The token behaves exactly as designed: it refuses. But **both** demo boxes still arm the offsite tier with `keep_last=2 prune_pbs_allowed=true` (`backup_targets: [{target_id: felhom-pbs, cadence_seconds: 604800, keep_last: 2}]`), so every run asks for a prune that must fail. **The data is SAFE and that is why this is not a P1:** the snapshot lands before the prune is attempted; what is wrong is the job's VERDICT and the weekly operator e-mail it produces. **But it is corrosive in the specific way this project keeps finding:** a backup that reports FAILED while succeeding trains the operator to discount `whole_guest_backup_failed`, which is the same alert that would carry a real one — and it is exactly the failure the R-100 corollary warns about, an alarm whose text is true and whose trigger is not the thing you would act on. **Fix is one config line per box** (`keep_last: 0` on the PBS tier) plus whatever writes it on a fresh install; **deliberately NOT applied in this session** — the session was a runbook with an explicit "change nothing, and if a change appears necessary, stop and report" rule, and a retention field on a live backup tier is not a change to slip into an observation run. **Check before fixing:** whether ep0's prune jobs actually cover these two namespaces, or the snapshots simply accumulate once the box stops asking **THE GATE WAS RUN FIRST, AND IT MATTERED.** Before disabling anything, ep0 was read (read-only, Tier 2): prune jobs `prune-demo-felhom` and `prune-demo-hp` exist on datastore `felhom-offsite`, one per namespace, `schedule 03:30`, `keep-last 2`, comment *"R-82 retention keep-last=2, server-side (box tokens are write-only)"* — and they have run **every day since 2026-07-27: 18 tasks, all `status=OK`**. The newest task log reads `retention options: --ns demo-felhom --max-depth 0 --keep-last 2` / `keep ct/9201/2026-07-27…` / `keep ct/9201/2026-07-28…` / `TASK OK`. Retention happens, and it happens there. **A METHODOLOGICAL WARNING WORTH MORE THAN THE FIX.** Three separate queries said the OPPOSITE — *no prune jobs have ever run* — and **all three were broken instruments**: `worker-type` where the field is `worker_type`; the value `prune` where the worker type is `prunejob`; and `journalctl -u proxmox-backup` where the unit is `proxmox-backup-proxy`. A fourth reading (3 snapshots under keep-last 2) was mis-framed by CC and self-corrected — the third snapshot had landed AFTER that day's 03:30 window. Acting on any of them would have disabled the only pruning ATTEMPT while reporting that nothing prunes: a weekly false alarm traded for unbounded growth on the protected endpoint, invisible for months. **The gate is what caught it, and only because it demanded evidence rather than a verdict.** **Shipped:** installer **1.25.0** writes `keep_last: 0` on the offsite tier (the agent's existing guard `allowPBSPrune = !primary && keep_last > 0` already reads that as *never prune from the box* — no agent change), the justifying paragraph is rewritten to say where retention lives and cite R-89, and `hostinstall_gates.py` asserts it (red-proved: pinning `keep_last: 2` back fails the gate). **Both live boxes corrected in their own config** — `backup tier armed target=felhom-pbs … keep_last=0 … prune_pbs_allowed=false` on demo-felhom and demo-hp, with the local tier untouched at `keep_last=3`. Served over HTTPS at `1.25.0` with `"keep_last":0` in the served bytes. **STILL TO OBSERVE:** the next weekly offsite run completing OK end-to-end. The change removes the failing step; the *schedule* proving it is next week's event, and this row should carry that line when it happens. | CC |
 | **R-194** | **PVE's permission cache delays every grant-state verdict by an unknown amount, so "the agent can read it" and "the ACL exists" are not the same measurement.** Observed twice while validating R-190's self-repair on demo-felhom 2026-08-04: both ACL rows for `/storage/felhom-backup` were deleted, and `GET /access/permissions` continued to report `Datastore.AllocateSpace` present — for **~40 s** in one run and **~16 minutes** in another. During that window the capability probe reads healthy and the self-repair does not fire | **OPEN** | — | **Why it matters beyond the delay:** it puts a floor under how fast a lost grant can be noticed, it makes any single permission read a lagging indicator, and — the interesting part — **it is a candidate contributor to R-190's own timeline**: a grant removed at an unknown moment could keep working until a cache expiry, which is exactly the shape of *worked at 04:44, refused at 09:24*. That does not explain what removed it, but it may explain when the refusal SURFACED, and the two have been treated as the same instant. **Not a defect in our code** — it is PVE behaviour, and the mitigation already tolerates it (the repair fires on the next probe after the cache clears). **What is worth deciding:** whether the store-grant probe should ALSO consult the storage content listing as a second signal, since that appeared to reflect the loss immediately (`{"data":[]}` while the permission read still said present) — two signals disagreeing is itself information, and today only one of them is read | CC |
 | **R-200** | **The DR password-injection seam has a handler, a route and tests — and no form.** `POST /backup/offbox/inject-password` is routed (`controller/internal/web/server.go:510`) to `offboxInjectPasswordHandler` (`offbox_handlers.go:174-196`) → `InjectOffboxPassword` (`backup/offbox.go:541`). **No template in the repository contains that path or any form posting to it** (grep over `internal/web/templates/`: one unrelated hit, an XSS comment) | **PLUMBING COMPLETE** (controller **v0.196.0**); **the FORM is not built — still open** | — | **The tenth instance of this project's built-but-never-wired class, and the exact shape `CLAUDE.md` and `felhom.eu/CLAUDE.md`'s seam-wiring rule were written for:** handler tests that POST directly (`offbox_escrow_test.go:167,180`) prove nothing about reachability. To use the only implemented recovery seam today, a person must hand-craft an authenticated POST with a session cookie and CSRF token. **Note the layering while fixing it:** this form takes a **64-hex repo password**, not a recovery code (`offboxRepoPwPattern`, `offbox.go:543`) — they are different secrets at different layers, and the operator's 2026-08-04 ruling asks for a form that takes **R**. Build the R form and treat this one as the operator/DR fallback it was written as, but **ship it with a render test per branch of whatever gate it sits behind**. Source: `audits/RECON-offsite-dr-chain-2026-08-04.md` §3 link 9 **THE DIAGNOSTIC HALF IS DONE AND IT ANSWERED THE QUESTION.** `--recover-offsite-check` is a `docker exec` escape hatch in the shape of `--print-reset-code`: R on **STDIN** (never argv, never `ps`, never shell history, never a transcript), fetch+unseal via the agent, and a verdict of **two sha256 hashes**. **It compares and never installs** — the recovered password is not written to `offbox/repo_password`; a test asserts the data dir is byte-unchanged and its red-proof (adding the install call) fails it. Confirmed live: `repo_password` mtime still `2026-08-03 07:18:02` after the successful check at `2026-08-04 11:49`. **Exit codes are load-bearing** — `0` match, `2` a clean MISMATCH, `1` a step failed; "it failed" and "it worked and disagreed" must never share a status because only one is a finding about the system. A box with no local password reports distinctly (the rebuilt-box shape, where the next step is to INSTALL rather than compare). **WHAT IS NOT BUILT, deliberately:** no card, no form, no preview, no customer-facing text — building an interface on top of a chain nobody had walked is how the preceding three weeks went wrong. **What remains for this row:** link 9 (the recovered password placed so `WriteOffboxSecrets` keeps it) and the customer-facing shape the operator ruled on 2026-08-04 (yell → R form → preview → proceed), which is now priced against a chain that exists rather than one that is assumed **PART 0 SHIPPED 2026-08-04 (v0.196.0):** `--recover-offsite-install` is the sibling of the check — same fetch/unseal path, same STDIN discipline for R — and it **places** the recovered password via `InjectOffboxPassword`. The confirmation is a SECOND invocation (`--confirm-install`): without it, both hashes print and nothing is written, so the operator sees the comparison before a write is possible. Three outcomes named distinctly: **installed** (no local password — the rebuilt-box shape), **unchanged** (identical key, nothing written), **refused** (a DIFFERENT key present — installing would clobber the key the current repository is encrypted under; exit 2, no force offered). It re-reads the file after writing rather than trusting the call. **Red-proof observed:** removing the confirmation gate makes the dry run write the password. The R-persistence test carries a **positive control** (a planted copy found, then removed and not found). **NOT YET EXERCISED AGAINST A LIVE RECOVERY** — the R-201 drill halted before step 9, so this is unit-proven only. **What remains for this row:** the customer-facing shape the operator ruled on 2026-08-04 (yell → recovery-code form → preview → proceed) | CC |
@@ -602,6 +602,8 @@ class (an image `VOLUME` at an unmounted path) is still live — `immich-server`
 | **R-426** | **The decoy-coverage exemption list — 20 registered gates that ship WITHOUT a decoy test, each named.** `scripts/decoy_coverage_gate.py`'s `EXEMPT` map is debt, and this row owns it so it lives in the register and not only in a Python literal. **Four kinds:** (a) genuinely covered in the 2026-09-01 sweep but not yet moved into a suite — `hub-copy`, `instructions`, `docker-v`, `image-pins`; (b) blocked by an open hole and therefore un-assertable as rejecting — `site` (R-423), `one-register` (R-424), `offbox-rename` (R-425); (c) shared scripts whose decoy lives in `felhom.eu` and is counted there — `reuse-refs`, `instructions`, `observations` in the controller and agent runners; (d) **no plausible decoy constructed yet** — `hostinstall`, `wire-contract`, `due-checks`, `published`, `image-resolvable`, `volume-persistence`. Group (d) is the honest unknown: six gates whose soundness is UNTESTED, not established. **The list is green today and shrinks; a NEW gate with no decoy fails immediately.** | **OPEN — 20 names; group (d) is six untested gates** |
 | **R-427** | **`closed_register_gate.py` checks ONE direction only: an open word in a CLOSED row. The mirror — a CLOSED verdict on a row still sitting in `OPEN-ITEMS.md` — is unchecked, and there are TWELVE.** MEASURED 2026-09-01 during the decoy sweep, by reading the leading verdict of every open row with the gate's own predicate: **R-385, R-387, R-341, R-378, R-405, R-88a, R-88b** read unambiguously closed; **R-123, R-190, R-352** read `PARTLY CLOSED` / `MITIGATION SHIPPED` and almost certainly belong where they are. **The rows were NOT moved by this session** — telling a finished row from a partly-finished one is a judgement, and R-378 is itself the record of what happens when a machine makes that judgement on a substring (six still-open rows moved out of the register). **This is R-405's finding mirrored:** that row exists because R-87 sat in the CLOSED file while its state read READY, and the gate written for it looks only the way it was bitten. Fix: the same leading-verdict predicate applied to `OPEN-ITEMS.md`, reporting rather than convicting until the twelve are adjudicated by a person — a gate registered while twelve rows fail it would refuse every push. | **OPEN — 12 rows named; the adjudication is Viktor's, the gate is mine** |
 | **R-428** | **The decoy-coverage gate — written to catch instruments that match a NAME instead of a fact — identified a repository by its DIRECTORY NAME.** MEASURED on its own first CI run (felhom.eu job 490, 2026-09-01): `os.path.basename(root)` looked up in a `RUNNERS` map, and Gitea's act-runner checks the repo out into a directory called `hostexecutor`, so the gate reported *"unknown repo 'hostexecutor'"* and went INCONCLUSIVE. **The gate that hunts label-matching was matching a label, in the first ten lines of its own main loop, and it shipped that way.** FIXED the same day: it now identifies a repo by which registered runner FILE exists under the root, which is a fact. **Recorded rather than quietly patched because it is the strongest evidence in the sweep that this class is not a matter of carelessness** — it was written by a session that had spent the morning reading 29 gates for exactly this, with the four shapes on screen. Verified under a renamed directory before and after. | **CLOSED 2026-09-01 — fixed, and kept as the class's best example** |
+| **R-429** | **The Storage Box snapshot mitigation that R-95 calls "ARMED" has never been confirmed, cannot be seen from either box, and its row has no id.** MEASURED 2026-09-01 over the SFTP credential each box already holds, read-only, with positive and negative controls on both machines: **no `.snapshots` directory is visible to either sub-account** — not in the account home, not inside the repo path — while the account is jailed (`/` returns `Permission denied`) and a bogus path errors correctly. Two readings the box cannot distinguish: no snapshots exist, or they exist and a sub-account cannot see them. **Both are bad, and the second is not a reprieve — a snapshot the box cannot see is one the box cannot restore from, so recovery would be an operator act through the Hetzner panel, not a product capability.** The claim lives at `OPEN-ITEMS.md:233` as a row whose ID cell is a bare em-dash instead of an R-number, so no gate and no grep-the-register rule can cite it; its "confirm tomorrow" was **2026-07-27** (last touched in `72692e1`), **36 days** unconfirmed; and the `DUE-CHECKS` block built for exactly this (R-341) is **EMPTY**. **The confirming field, `size_snapshots`, is a Hetzner API field and §11-D fences it — so this spike stopped and left it for Viktor.** Fix: Viktor reads the Storage Box panel once (ten minutes) and the answer goes in this row; give the WATCHING row an id or delete it; and R-95's word "ARMED" must not stand until then. Evidence: `audits/SPIKE-r95-offsite-delete-2026-09-01.md` §Q1. | **OPEN — Q1 is Viktor's, and it re-ranks R-95** |
+| **R-430** | **`restic unlock --remove-all` printed `successfully removed locks` while the lock was still there.** MEASURED 2026-09-01 in a throwaway local repo (no live store touched), under a faithful append-only model — a sticky locks directory owned by root holding a root-owned lock, restic run as `nobody`; both controls passed first (create allowed, delete refused). The command reported success, returned, and `ls` showed the lock present. **`resticStep`'s crash-lock self-heal is built directly on this call** (`felhom-controller/controller/internal/backup/offbox.go:~768`), and its licence to escalate rests on the escalation actually working. **A self-heal that cannot fail is a self-heal that cannot be trusted** — this is this project's *"exit codes that lie"* class, in the one path that runs unattended against the customer's off-site history. It is harmless TODAY because the credential can delete and the removal really happens; it becomes load-bearing the moment delete is withdrawn, which is what R-95 is about. **Not yet established:** whether restic reports success because it removed zero locks by design, or because it did not check. Settling it: read restic 0.14.0's unlock source, or re-run with `--verbose`. | **OPEN — precondition on any R-95 build** |
 
 <!-- DUE-CHECKS-BEGIN — machine-readable. Parsed by scripts/due_checks_gate.py.
      One row per dated check. The R-number must have a row above. Dates are UTC.