CI workflow: the failure mail links the public run page, by run id (R-1135)

The alarm step built its link from GITHUB_SERVER_URL, which on this runner
is the in-cluster Gitea address, and from the run NUMBER, which Gitea 1.26
reads as a run id first (felhom.eu /actions/runs/1116 opened run id 1116 =
run number 613). Now https://gitea.dooplex.hu + GITHUB_RUN_ID, and the link
is printed into the job log as ALARM-LINK. Same change in all four repos.

.gitea/workflows/gates.yml

@@ -16,7 +16,16 @@
 # A failing run must reach a person — a detector nobody hears is the defect R-29 filed, rebuilt one
 # layer up. That is the last step, and it runs ONLY on failure.
 name: gates
-on: [push]
+# R-1129: `workflow_dispatch` is the manual PROOF that the failure mail can be sent. A push run is
+# unchanged by it: the proof step below is skipped on a push, and `force_alarm` is then empty.
+on:
+  push:
+  workflow_dispatch:
+    inputs:
+      force_alarm:
+        description: "Fail this run on purpose so the alarm step sends its real mail (a proof, not a gate result)"
+        type: boolean
+        default: false
 
 jobs:
   gates:
@@ -57,6 +66,22 @@ jobs:
           GITEA_BASE: http://gitea.gitea-system.svc.cluster.local:3000
         run: cd ws/felhom-agent && python3 scripts/agent_gates.py
 
+      - name: Manual proof - alarm key present, optional forced failure (dispatch only, R-1129)
+        # Runs ONLY on a manual dispatch, never on a push. It prints the key's LENGTH, never the key,
+        # so "CI can see the secret" is an observable rather than an assumption. With
+        # force_alarm=true it then fails on purpose so the alarm step below sends its real mail;
+        # the commit itself stays green, because its push run skips this step.
+        if: github.event_name == 'workflow_dispatch'
+        env:
+          RESEND_API_KEY: ${{ secrets.RESEND_API_KEY }}
+          FORCE_ALARM: ${{ github.event.inputs.force_alarm }}
+        run: |
+          python3 -c 'import os;print("RESEND_API_KEY length: %d" % len(os.environ.get("RESEND_API_KEY","")))'
+          if [ "$FORCE_ALARM" = "true" ]; then
+            echo "force_alarm=true - failing on purpose so the alarm step runs (R-1129 proof)"
+            exit 1
+          fi
+
       - name: Alarm on failure
         # THE POINT OF THE WHOLE THING. Probe P5 measured that a failed run produces NO mail, NO
         # notification row and NO log line from Gitea itself — a red tick in a web UI nobody watches
@@ -70,6 +95,8 @@ jobs:
         if: failure()
         env:
           RESEND_API_KEY: ${{ secrets.RESEND_API_KEY }}
+          # Empty on a push, so a push run's mail is exactly what it was before R-1129.
+          FORCE_ALARM: ${{ github.event.inputs.force_alarm }}
         run: |
           python3 - <<'PY'
           import json, os, sys, urllib.request, urllib.error
@@ -81,13 +108,23 @@ jobs:
 
           repo = os.environ.get("GITHUB_REPOSITORY", "?")
           sha  = os.environ.get("GITHUB_SHA", "?")
-          run  = os.environ.get("GITHUB_RUN_NUMBER", "?")
+          # R-1135: the link is for a PERSON, so it is the public address and the run ID.
-          srv  = os.environ.get("GITHUB_SERVER_URL", "https://gitea.dooplex.hu")
+          # NOT GITHUB_SERVER_URL: the runner registers at the in-cluster address, so that variable
+          # reads http://gitea.gitea-system.svc.cluster.local:3000, which nobody outside the cluster
+          # can open (forced-test mail, job 1997). NOT GITHUB_RUN_NUMBER: Gitea 1.26 reads
+          # /actions/runs/<n> as a run id first, so a run number that is also an older run's id
+          # opens THAT run (measured 2026-10-18: felhom.eu /actions/runs/1116 opened the run whose
+          # id is 1116 = run number 613). The link is printed too, so the job log carries it.
+          run  = os.environ.get("GITHUB_RUN_ID", "?")
+          srv  = "https://gitea.dooplex.hu"
+          print("ALARM-LINK %s/%s/actions/runs/%s" % (srv, repo, run))
 
           body = json.dumps({
               "from": "Felhom CI <monitoring@felhom.eu>",
               "to": ["admin@felhom.eu"],
-              "subject": "[felhom CI] gates FAILED in %s" % repo,
+              "subject": ("[felhom CI] FORCED TEST, not a real failure - gates FAILED in %s"
+                          if os.environ.get("FORCE_ALARM") == "true"
+                          else "[felhom CI] gates FAILED in %s") % repo,
               "text": (
                   "The gate entry point exited non-zero.\n\n"
                   "Repository : %s\n"

CHANGELOG.md

@@ -1,3 +1,179 @@
+## Unreleased (2026-10-18) — CI workflow: the failure mail links the public run page, by run id (R-1135); no code change
+
+`.gitea/workflows/gates.yml`, the alarm step: the mail's run link is now `https://gitea.dooplex.hu/<repo>/actions/runs/<run id>`.
+It used `GITHUB_SERVER_URL`, which on this runner is the in-cluster address (`http://gitea.gitea-system.svc.cluster.local:3000`,
+the address the runner registers at), so nobody outside the cluster could open it. It also used the run NUMBER, and
+Gitea 1.26 reads `/actions/runs/<n>` as a run id first: a number that is also an older run's id opened that other run
+(measured 2026-10-18: `felhom.eu/actions/runs/1116` opened run id 1116, run number 613). It now uses `GITHUB_RUN_ID` and
+prints the link (`ALARM-LINK …`) into the job log. Same change in all four repos' workflows.
+
+## Unreleased — three fixes from the 2026-10-12 security review and the storage build (R-1093, R-978, R-973) (2026-10-18)
+
+Committed, not released — the release and its vouch are the operator's.
+
+- **R-1093 — an EMPTY partition no longer asks the typed „data will be erased" confirmation.** `lsblk` prints the
+  parent disk's table type (PTTYPE) on every partition, so `SudoHostOps.InspectDevice` read a blank partition as
+  „device has a partition table". Now lsblk's PTTYPE is set aside only when `blkid -p` on the partition itself proved
+  it read the device and that it is a partition of a table of exactly that type (`PART_ENTRY_NUMBER` +
+  `PART_ENTRY_SCHEME` equal to lsblk's PTTYPE). The partition's own signatures still come from `blkid -p`: a
+  filesystem or a nested table inside it stays data-bearing; a blkid that printed nothing, a whole disk, or a scheme
+  that differs keeps the old, safe reading. **No new command, no sudoers or config-bundle change** (the two probes are
+  the same allow-listed argv). Test `TestInspect_R1093_PartitionVerdicts` (8 rows), red-proved: with the old
+  `if pttype != ""` both empty-partition rows fail `DataBearing=true reason="device has a partition table"`.
+  **Hardened after the same night's review (fail-closed on every path):** `PART_ENTRY_*` is read from the PARENT's
+  table, so lsblk's PTTYPE is set aside only when blkid also **exited 0** — exit 2, an I/O error or a missing binary
+  keep „device has a partition table"; and an `lsblk` that exits 0 with output that does not parse into a device (or
+  lists none) now leaves the device UNPROBED („could not be reliably inspected") instead of „blank" — that hole
+  predates R-1093. Six more rows in the same test, each red-proved (`blkidErr` check removed → „device is blank";
+  `&& parsed` removed → `Probed=true … „device is blank"`).
+- **R-978 — an abandoned second escrow ceremony no longer makes the customer's recovery code useless.** The root
+  ceremony (the FELHOM_ESCROW argv, unchanged: `--upload --output=json`) used to upload the new package before the new
+  code was claimed, and the hub overwrites the current package when the sealed secrets are unchanged — so a customer
+  who started the ceremony again and walked away (or an agent restart in the 10-minute window) was left with a
+  written-down code that opened nothing. Now json mode hands the package back (`CeremonyOutput.PendingUpload`, output
+  **version 2**; text mode still uploads at once) and the daemon uploads it at `POST /escrow/ceremony/claim`, **before**
+  the code is handed out (`EscrowCeremonyConfig.Upload` → `putEscrowBody`, the same PUT and per-host key). A failed
+  upload answers 502 and keeps the code claimable; the TTL and a new start wait while the claim's upload is in flight;
+  no uploader → 503; an output that uploaded by itself or carries no package fails the ceremony. Status `uploaded`
+  now reads false until the claim. The controller is unchanged (it never read `uploaded`; its claim already assumed
+  the package is on the hub when the claim returns). **No sudoers or config-bundle change.** Tests (localapi)
+  `TestEscrowCeremony_R978_*` (abandon by TTL / by a new start / by a restart → the hub's package still opens with
+  code 1; claim stores before handing out; upload failure; in-flight TTL; no uploader; no package) and (cmd)
+  `TestEscrowOptsFromCLI_JSONModeDefersUpload`, `TestDeliverEscrowPackage_JSONModeSendsNothing`,
+  `TestR978_CeremonyUploadIsWiredIntoTheDaemon`. Red-proved: upload-at-create restored → „code 1 no longer opens it
+  (uploads=2)"; the in-flight TTL guard removed → „the TTL voided the code while its package was uploading";
+  `deferUpload: false` → „json mode must defer the upload"; the `Upload:` line removed → „has no Upload". **Residual:**
+  the customer's typed-back check after the claim is browser-only, so „confirmed" here means „claimed" — a customer
+  who opens the code and closes the page without writing it down has replaced the package.
+- **R-973 — the local-API whole-guest backup route has a cadence cap.** Every successful run prunes its tier, so a
+  compromised controller calling `POST /backup` in a loop pushed out every pre-compromise restore point (local
+  keep-last 3; off-site via ep0's keep-last 2). Now (`internal/localapi/backup_cap.go`): a tier whose cadence is ≤ 24 h
+  allows **3 successful runs per rolling 24 h** (the `07` §6 day: the night run, a „Mentés most" press, and the next
+  night, which since R-899 still runs and can land inside the same 24 h — so a press never costs the night); a longer
+  tier allows **1 per max(cadence − 24 h, cadence/2)** — 6 days for the 7-day off-site tier, which the controller only
+  runs when due and the button never targets. Only successes count (a failed run prunes nothing). A refusal answers
+  **429** with the next allowed time, is logged at WARN, is saved, and for 24 h raises the CRITICAL host-report status
+  `localapi:backup-cadence-cap` — the hub's existing operator-only `agent_capability_degraded` event, **no hub change**.
+  **The cap survives an agent restart:** `BackupSuccessState` (the R-894 file) now also keeps the newest 8 success
+  times and the last refusal per tier and guest (additive JSON fields; a pre-R-973 file counts its one known
+  success). Tests `TestBackupCap_*` (an off-site flood keeps a pre-compromise copy for 5+ days; a local flood runs 3
+  a day; the scheduled day is never refused; restart; failures; the operator signal) + `TestBackupCadenceCap`,
+  `TestBackupSuccessState_R973_*`, `TestBackupCapStatus_*`, `TestR973_CapRefusalIsWiredIntoTheReport`. Red-proved:
+  the cap disabled → „the off-site tier lost every pre-compromise copy: 10 runs"; the disk history ignored → „after a
+  restart the 4th run in a day was allowed"; the wiring line removed → „never binds backupCapRefusal". **Limits, stated:**
+  with keep-last 3 the local tier's three points can still be replaced within one day (3 per day is the schedule's own
+  pace plus one press); the hub's event text reads „agent privileged capability degraded" (the type is generic). No
+  sudoers or bundle change.
+
+## v0.159.2 — the orphan thin check runs only after a failed backup (R-1012 follow-up) (2026-10-18)
+
+**Released 2026-10-18 (committed 2026-10-17): sha256 `4d98868e922d6c10665a6c0da9323e8627ea9c589094e81d27d741e625d9d8f9`, bundle `477704ea2980948433ffc34bb59f008cc11410fd9683d386bfd3e2f085d55bc9`; the vouch is the operator's.** No `configs/` change since v0.159.1. The independent review of 2026-10-18 found nothing in it. v0.159.0/v0.159.1 ran the R-1012 orphan check
+before EVERY snapshot-mode whole-guest backup, and that check takes a short metadata snapshot of the live thin pool.
+Now `BackupRunner` runs it only when the snapshot step has failed (the vzdump start or its task) — the moment the
+answer is needed. The finding is still appended in plain words to the failed backup's error, exactly as before; a
+successful backup never runs the check and keeps a clean record; stop mode and an unwired runner are unchanged. The
+`thin-orphan-check` capability text says „after a failed whole-guest backup". Tests
+`TestBackup_SnapshotFailure_NamesTheOrphan` (now: exactly 1 call, after the failure) and
+`TestBackup_ThinCheck_SilentWhenNotApplicable` (now: 0 calls on a successful backup), red-proved against the v0.159.1
+`runner.go`: failure legs „thin check ran 2 times, want 1", success leg `calls=1`. No root file, sudoers or bundle
+change. (Installer side, felhom.eu, same night: `felhom-host-install.sh --rotate-recovery-only`, R-1141.)
+**Proven live 2026-10-18 (demo-hp):** a green snapshot backup with 0 `felhom-thin-check` calls (the 0.159.1 night backup
+had 1). The failed leg stays covered by the fixture test.
+
+## v0.159.1 — the signed decommission prunes the guest-bind record (found by v0.159.0's own live proof) (2026-10-17)
+
+**v0.159.0 is superseded — do not vouch it.** Everything of v0.159.0 below, plus: `signedjobs.DecommissionExecutor`
+drops the drive from every guest's bind record (`SetGuestBinds`, wired to the agent's `GuestBindStore`), as the
+self-serve decommission always did. Without it a drive decommissioned by an operator-signed job and then forgotten
+(R-1124) left a record the 20-second guest-bind re-assert warned about every tick („reconcile: skipping non-enrolled
+drive") — seen on demo-hp minutes after the R-1124 proof. Test `TestDecommissionExecutor_PrunesTheGuestBindRecord`,
+red-proved. Released: tag `v0.159.1`, sha256 `de8f6ad9…aff9`, bundle `2af3d27d…7cd6` (the same root files as v0.159.0).
+**Not vouched.** Delivery to a box on the v0.158.1 bundle: `agent_update 0.159.1` → `agent_config_update 0.159.0-step1`
+(sha `1d65edb3…d6cf`, published) → `agent_config_update 0.159.1` (the bundle adds `felhom-thin-check`, R16).
+
+## v0.159.0 — forget a decommissioned drive that is gone (R-1124); name an orphan thin device in a backup failure (R-1012); Go 1.27.2 (R-972) (2026-10-17)
+
+**Superseded by v0.159.1 — do not vouch.** Task: „agent release, the catalog allow-list, break-glass rotation, off-site
+history check" (felhom.eu `REPORT-agent-allowlist-rotation-2026-10-17.md`).
+
+- **Go 1.27.2 + x/crypto v0.58.0** (R-972, `37b99d7`): the first release that carries them (v0.158.1 was built with
+  go1.26.0).
+- **R-1124 — a decommissioned drive that is gone is forgotten.** `SudoHostOps.ForgetRetiredDrives` (agent start and the
+  20-second storage tick): for a Felhom drive unit whose intent is `decommissioned`, whose `/dev/disk/by-uuid` link is
+  gone and which is not mounted, `systemctl disable` + the new root verb `felhom-priv-apply unit-forget <unit>` +
+  `daemon-reload`. The unit is gone, so a later boot no longer waits for the drive and the agent stops listing it (no
+  „absent by UUID" line every 20 s). A decommissioned drive still plugged in KEEPS its unit and raw mount, so the
+  one-click re-enroll (H3) still works. `unit-forget` re-checks as root: a felhom by-UUID drive unit (not a network
+  unit), disabled, its `Where` not mounted, its by-UUID link absent — otherwise refused. Sudoers `FELHOM_MOUNT` gains the
+  one verb. Tests `TestForgetRetiredDrives_*` and `configs/test_felhom_priv_apply.py` `UnitForget`, red-proved.
+  **Live on demo-hp:** a loop-backed test drive enrolled through the local API, decommissioned by a signed job (kept
+  its unit while present), unplugged → 8 s later the unit and its wants link were removed and `GET /disks` dropped it.
+- **R-1012 — an orphan thin device is NAMED in the backup failure, never removed.** New READ-ONLY root helper
+  `felhom-thin-check <vg> <pool>` (config bundle): LVM's thin ids before and after, the pool's ids through a metadata
+  snapshot it reserves and releases (refuses — exit 3 — when someone else holds one; releases on every exit; sends no
+  other pool message). `internal/thincheck`: an orphan = in the pool and in neither LVM list, on two reads 3 s apart;
+  fails open. `BackupRunner.SetThinCheck`: before a snapshot-mode backup (a WARN when found) and again after a failure —
+  the finding is appended in plain words to the error the controller mails to the operator, pointing at the new
+  runbook `whole-guest-backup-orphan-thin.md`. Sudoers `FELHOM_THINCHECK`, capability `thin-orphan-check`. Tests
+  `internal/thincheck`, `TestBackup_SnapshotFailure_NamesTheOrphan`, `configs/test_felhom_thin_check.py`, red-proved.
+  **Live on demo-hp:** on a scratch thin pool (a loop file on `/mnt/hdd_1`, never `pve/data`) the helper listed a made
+  orphan (device 7) and refused a snapshot held by someone else; on `pve/data` it read 1, 2 = LVM's and released; a
+  normal whole-guest backup of 9201 ran the check first and finished green.
+- Released: tag `v0.159.0`, sha256 `3cfd0a25…edfed5`, bundle `40663c4f…6d681`; step bundle `0.159.0-step1`.
+
+## CI only — a manual proof that the failure mail can be sent (R-1129, 2026-10-17)
+
+No agent release; no binary changes.
+
+`.gitea/workflows/gates.yml` gains a manual `workflow_dispatch` trigger with one input, `force_alarm` (default false), and
+one step that runs ONLY on a dispatch: it prints the length of `RESEND_API_KEY` (never the key) and, with
+`force_alarm=true`, fails on purpose so the existing alarm step sends its real mail; the mail's subject then starts
+„FORCED TEST, not a real failure". A push run is unchanged: the new step is skipped and the subject is the old one. Same
+change in all four repos' workflows. Purpose: prove the failure mail now goes out after the operator set the key
+(every failed run until today ended „ALARM FAILED: RESEND_API_KEY is empty").
+
+## scripts only — the decoy suite is a gate of the full run (R-1116, 2026-10-17)
+
+No agent release; no binary changes. `scripts/test_gate_decoys.py` is now the `gate-decoys` row of `agent_gates.py`:
+the full run (which CI runs on every push), not `--fast` — it takes ~20 s in the workspace (measured: 30 cases; ~5 s in a
+CI-shaped shallow clone, also green). It was in no runner, the shape that let the catalog's R-1115 fail unseen for a
+night. It plants only in scratch clones and asks a fake Gitea on 127.0.0.1; it does not call the runner (no
+recursion). Declared in its own `COVERS`; `test_agent_gates.py` pins it (registered once, full run only, named by
+`--fast`'s SKIPPED line). Red-proved: a reuse-refs FACT case flipped to expect rc 0 turned the full run red
+(`CONVICTED: gate-decoys`); restored, green. CLAUDE.md's `--fast` sentence names it.
+
+## unreleased — Go 1.27.2 and golang.org/x/crypto v0.58.0 (R-972, 2026-10-16)
+
+**Not released, not deployed** — agent releases wait for the operator's vouch; the next vouched release carries this.
+`go.mod`: `go 1.27.2` (was 1.25.0; the shipped 0.158.1 binary was built with go1.26.0), x/crypto v0.58.0 (was v0.52.0),
+x/sys v0.49.0. `go build`, `go vet`, the full `go test ./...` and `agent_gates.py --fast` green on Go 1.27.2.
+`release-agent.sh` builds with the toolchain `go.mod` names (GOTOOLCHAIN=auto), so no script change.
+
+## scripts only — `release-agent.sh` refuses when the registry cannot say the version is new (R-1098, 2026-10-16)
+
+No agent release; no binary changes. `scripts/release-agent.sh` already refused a version whose tag exists or whose
+package answers 200. It let everything else through — a registry it could not reach (curl prints `000`) or a 500 read
+as „new". Now only a clean `404` proves the version is new; anything else refuses with the code it got. New
+`scripts/test_release_push_once.sh` (the real script, a fake `curl` and `go` on PATH; published / absent / unreachable /
+server error / tag exists), red-proved: with the 404 rule removed, the unreachable and server-error cases fail. Same rule
+as the controller `build.sh` and, since today, felhom.eu `scripts/build-hub.sh`.
+
+## v0.158.1 — security review of v0.158.0: the mkfs wrapper trusts no sibling claim; the internal-place helper fails closed (2026-10-16)
+
+**v0.158.0 is superseded — do not vouch it.** A security review of `912f7c1` found two defects; both fixed here, each test
+seen failing against the 0.158.0 script:
+- **`felhom-mkfs-guarded`**: 0.158.0 let a format through when a SIBLING part was mounted at /mnt/<name> AND under
+  /mnt/felhom-drives („Felhom's own drive"). That corroboration is a mount the agent can make itself, so a compromised agent
+  could unlock a host owner's disk — the exact case this root wrapper exists for. Removed: any mount of any node outside
+  /mnt/felhom-drives refuses again (now read from the kernel's mount table for every node; an unlistable disk refuses).
+  Consequence: the free part next to an in-use Felhom drive cannot be formatted — `CandidatePartition.Formattable` is false
+  for every part of a disk with a mounted node, and the controller does not offer it (a part with a filesystem can still be
+  USED). Pinned by configs/test_felhom_mkfs_guarded.py + partitions_test.go.
+- **`felhom-internal-place`**: a failed `lvs` listing read as „nothing promised" (a too-large place) and a `blkid` error read
+  as „no signature" (the only gate before mkfs). Now: a failed listing, a listing without the pool, or any blkid exit other
+  than 0/2 REFUSES. Three new tests, all failing against 0.158.0.
+- Released: tag `v0.158.1` (code `98f3cdd`), sha256 `20264df6…84b76a1`, bundle `c23c3d35…cb1c41b`. **Not vouched.**
+
 ## v0.158.0 — the internal data place; partitions as places; add_internal_place (2026-10-16)
 
 Design: felhom.eu `architecture/16-storage-physical-identity.md` §5, §6, §7, §9 (operator rulings D3–D5; the D3 sizes
@@ -17,12 +193,12 @@ ruled 2026-10-16 after the space check, felhom.eu `audits/storage-identity-build
 - **Partitions as places** (`internal/storage/partitions.go`): `GET /disks/candidates` lists each partition of a free
   disk as empty / filesystem / in_use. A data-bearing part's wipe id binds disk id + part number + fs-UUID
   (`byid:<…-partN>+fsuuid:<uuid>`); both must name the same device at format time.
-- **`felhom-mkfs-guarded`**: a SIBLING part that is Felhom's own drive (mounted raw at /mnt/<name> AND under
+- **`felhom-mkfs-guarded`** (REVERTED in v0.158.1): a SIBLING part that is Felhom's own drive no longer blocked formatting
-  /mnt/felhom-drives) no longer blocks formatting the free part next to it; every other mount still refuses; an
+  the free part next to it. First tests of the wrapper: `configs/test_felhom_mkfs_guarded.py`.
-  unlistable disk now refuses (it passed before). First tests of the wrapper: `configs/test_felhom_mkfs_guarded.py`.
 - `configs/` changed (sudoers `FELHOM_INTERNALPLACE`, the helper, the wrapper, the bundle list) — **a config-bundle job
   is needed** with the agent update. **Not vouched** (the operator's act).
-- Released: tag `v0.158.0` (code `912f7c1`), sha256 `a3589d1d…2caba`, bundle `17066273…fb5c76`.
+- Released: tag `v0.158.0` (code `912f7c1`), sha256 `a3589d1d…2caba`, bundle `17066273…fb5c76`. **Superseded by v0.158.1 —
+  do not vouch** (its mkfs wrapper trusted a sibling claim the agent can fake; its helper failed open).
 
 ## v0.157.0 — every storage place knows its physical disk(s); POST /disks/same-disk (2026-10-15)
 

CLAUDE.md

@@ -56,7 +56,8 @@ This is in the core because breaching it is how this component stops being audit
 gate in its `GATES` table (that table is the list); the shared ones — `reuse_refs_check`,
 `instructions_gate`, `observations_gate` — are the copies in `felhom.eu/scripts/`, never copied into
 this repo (a copy recreates the drift they detect; an absent sibling clone FAILS). `--fast` selects the
-gates touching no network and no container runtime, and skips `published` (network), naming it. **A missing gate is a FAILURE, never a skip.**
+gates touching no network and no container runtime, and skips `published` (network) and `gate-decoys` (the decoy
+suite, ~20 s; R-1116), naming them — CI runs the full set. **A missing gate is a FAILURE, never a skip.**
 
 **The pre-push hook** (`.githooks/pre-push`) runs it with `--fast` and refuses a failing push. It is
 **per-clone** — switch it on once with `git config core.hooksPath .githooks`, and a manual run WARNS

CONTEXT.md

@@ -1,5 +1,28 @@
 # CONTEXT — felhom-agent working state
 
+> **2026-10-18 (night) — UNRELEASED on main: R-1093, R-978, R-973** (release + vouch: the operator's). R-1093 an empty
+> partition is blank (lsblk's PTTYPE on a partition is the parent's; set aside only on a clean blkid exit 0 with a
+> matching `PART_ENTRY_SCHEME`; an unparsable lsblk is unprobed). R-978 the escrow package reaches the hub only at the
+> claim (ceremony output v2, `PendingUpload`; `EscrowCeremonyConfig.Upload`). R-973 the `POST /backup` cadence cap.
+> **Decision (decided by CC unattended — operator may reverse): the cap's numbers.** Q: how many whole-guest runs may
+> the guest ask for? Options: (a) 2 per 24 h on the daily tier — keeps one local point older than a day, but refuses
+> the NIGHT after a press (R-899's night can fall inside the same 24 h); (b) 3 per 24 h + 1 per 6 days on the weekly
+> tier — the schedule never refused, a second press in a day refused, the local tier still replaceable in a day;
+> (c) none — status quo. Chose (b): the off-site tier is the one `07` §8 row 10 relies on and it is fully protected
+> for ~6 days; the local half's real fix is a time-bucketed prune (keep-daily), which is a retention change for the
+> operator. No sudoers or bundle change in any of the three.
+
+> **2026-10-17 — v0.159.1 RELEASED (0.159.0 superseded, do not vouch).** Go 1.27.2; R-1124 a decommissioned drive that is
+> gone is forgotten (`ForgetRetiredDrives`, root verb `unit-forget`); R-1012 the read-only `felhom-thin-check` names an
+> orphan thin device in a failed backup's text; the signed decommission prunes the guest-bind record. Bundle adds a path →
+> step bundle `0.159.0-step1`. Proven on demo-hp; delivered by signed jobs on the operator's yes. Vouch: the operator's.
+
+> **2026-10-16 — v0.158.1 RELEASED and VOUCHED (storage identity build B; 0.158.0 superseded, do not vouch).** The internal
+> data place (root helper `felhom-internal-place`, the record in the state dir, mounted at /mnt/belso like a drive), the
+> signed `add_internal_place`, partitions in `/disks/candidates` (`formattable` only while nothing on the disk is mounted),
+> the partition wipe id `byid:<…-partN>+fsuuid:<uuid>`, the strict mkfs wrapper. Delivered through step bundle
+> `0.158.1-step1` (a bundle that adds a path, R16). Next: R-1093 (an empty part's probe reads the parent's partition table).
+
 > **2026-10-15 (night) — v0.157.0 RELEASED, NOT VOUCHED (storage identity build A, felhom.eu `16` §13).** Every storage
 > place carries `kind`, `physical_disks`, `identity` (+ reason) in `GET /disks` and the hub report; `POST /disks/same-disk`
 > answers the controller (v0.317.0, MinAgent 0.157.0). Computed on every read, never cached. Delivered by signed jobs to

REPORT.md

@@ -1,43 +1,12 @@
-# REPORT — v0.157.0: every storage place knows its physical disk(s); POST /disks/same-disk (2026-10-15)
+# REPORT — agent v0.159.2 released (2026-10-18)
 
-Written in ASD-STE100 Simplified Technical English. Design read first: felhom.eu `architecture/16-storage-physical-identity.md`
+**In plain words.** The pool check that names an orphan snapshot now runs only after a whole-box backup has failed — not
-(§3, §4, §8, §10; ACCEPTED 2026-10-15, §13 = what was built), `03-host-agent.md`. Task: storage identity build A (operator
+before every backup. Released today; it runs on demo-hp, Tester 1 and demo-felhom. The vouch is the operator's.
-rulings D1–D6). Controller counterpart: felhom-controller v0.317.0 (MinAgent 0.157.0); hub v0.154.0 shows the fields.
 
-## What changed
+- sha256 `4d98868e922d6c10665a6c0da9323e8627ea9c589094e81d27d741e625d9d8f9`, bundle
-
+  `477704ea2980948433ffc34bb59f008cc11410fd9683d386bfd3e2f085d55bc9`, tag `v0.159.2`; no `configs/` change since 0.159.1.
-- `internal/storage/identity.go`: the physical disk(s) under a place, from `/sys`, `/run/udev/data`, `/proc/mounts` only
+- Live proof on demo-hp: a whole-guest backup of 9201 (snapshot mode, completed) wrote **0** `felhom-thin-check` sudo
-  (no block I/O). `proven`, or `unknown` with `vm` / `network` / `usb_bridge` / `weak_wwn` / `no_serial`.
+  lines; the night's backup on 0.159.1 wrote 1. The failure leg is covered by `TestBackup_SnapshotFailure_NamesTheOrphan`.
-- `GET /disks` rows and the hub report's `storage_targets` gain `kind`, `physical_disks`, `identity`, `identity_reason`.
+- Delivered on the operator's yes by signed jobs (agent_update + agent_config_update): all three hosts `update
-- `POST /disks/same-disk` `{a, b}` → `{different, proven, reason, a, b}`; scoped like every `/disks` call; a path the agent
+  committed`, capability probe 75/75. Evidence: felhom.eu `documentation/audits/day-2026-10-18/P4/`.
-  did not mount itself is `unmapped`.
+- The independent review of 2026-10-18 found nothing in this change.
-- `configs/` unchanged.
-
-## Tests and red-proofs
-
-Fixture `/sys` + udev trees for every spike case and five same-disk handler tests. Seen failing first: the LVM thin
-„diamond" (tmeta/tdata under one pool) read as a failure — fixed by counting a seen node once. Broken on purpose and seen
-failing: the VM rule, the vendor-less WWN rule, the own-mount rule (an absent drive), the `sys_drive` mapping.
-`python3 scripts/agent_gates.py`: all OK.
-
-## Release and delivery
-
-- `scripts/release-agent.sh 0.157.0`: tag `v0.157.0`, sha256 `addd4041…0aef`, bundle `0604bc60…397a`.
-- Delivered by signed `agent_update` jobs to Tester 1, demo-hp and the N100 (the three test boxes). **NOT VOUCHED** — the
-  operator's act. Until the vouch, the global floor cannot declare MinAgent 0.157.0.
-
-## Live proof (method: the agent's local API on each PVE host, with the guest's own token)
-
-| Box | `GET /disks` | same-disk: drive vs system data |
-|---|---|---|
-| demo-hp | `local`/`local-lvm` proven (SATA SSD), NVMe drive proven (`eui:`), PBS `network` | different, proven |
-| N100 | system proven (SATA SSD serial), USB drive proven (the bridge passes the real serial) | different, proven |
-| Tester 1 (VM) | every place `unknown/vm` | not proven (`vm`) |
-
-On all three: system data vs `/var/lib/docker` → same disk, proven; a drive name that is not mounted → `unmapped`. The hub
-stored the same fields (read from its database with the WAL). Evidence: felhom.eu `audits/storage-identity-buildA-2026-10-15/`.
-
-## Observations
-
-1. FILED: R-1083 — scratch box 9202 has no `local_api` in its `bootstrap.json`, so it cannot use any agent call.
-2. NOT-A-FINDING: `usb_bridge` is built from the design's rule and was not seen on a real enclosure; that is R-1077, already open.

REUSE.md

@@ -15,13 +15,14 @@
 | `SudoHostOps.run` | internal/storage/hostops.go | `run(ctx, name, args...) error` | allowlisted exec with stderr-wrapped error | Every arg pre-validated via validate.go before this is called |
 | `Prober.Probe` | internal/capability/probe.go | `Probe(ctx) []Status` | live sudo-policy capability check (`sudo -n -l --`) | Needs a DIRECT runner (never the sudo-prefixing one — double-sudo); never executes probed cmds. v0.86.0: config-gated caps (`Capability.GatedBy` + `Prober.GateActive`) report `inactive`/"disabled by configuration" ONLY when healthy — broken plumbing stays degraded; the pbsdr-* gate answers from `pbsdr.Manager.DRConfigured` (marker-backed across restarts) |
 | ~~`stageTemp`~~ (REMOVED v0.146.0, R-861) | — | — | — | Nothing the agent writes is `install`ed where root reads it any more: use `felhom-priv-apply` (below) or ship a fixed file in the bundle |
-| `felhom-priv-apply` (v0.146.0, R-861) | configs/felhom-priv-apply | `felhom-priv-apply unit <name> \| dnsmasq <tmp> <name> \| wg \| sshd-config \| sshd-key \| controller-image <vmid>` (the last reads the ref on stdin, R-861 (a) A1) | ANY agent-rendered file a root program reads (systemd unit, dnsmasq drop-in, wg-quick conf, OOB sshd) — fixed source + destination, CONTENT checked against the agent's own renderers | A new renderer needs a verb + a contract test (`internal/privapplytest.Check`) feeding its REAL output; never a new `install` sudoers line |
+| `felhom-priv-apply` (v0.146.0, R-861) | configs/felhom-priv-apply | `felhom-priv-apply unit <name> \| unit-forget <name> (v0.159.0, R-1124) \| dnsmasq <tmp> <name> \| wg \| sshd-config \| sshd-key \| controller-image <vmid>` (the last reads the ref on stdin, R-861 (a) A1) | ANY agent-rendered file a root program reads (systemd unit, dnsmasq drop-in, wg-quick conf, OOB sshd) — fixed source + destination, CONTENT checked against the agent's own renderers | A new renderer needs a verb + a contract test (`internal/privapplytest.Check`) feeding its REAL output; never a new `install` sudoers line |
 | `privapplytest.Check` | internal/privapplytest/check.go | `Check(t, verb, name, content) string` | the Go↔root-checker contract: a renderer's real output must read `OK` | Skips without python3; one call per rendered shape + one refused control |
 | `BUNDLE_FILES` + `Bundle` (mode `bundle`, `--install-bundle`; mode `agent_update` v0.146.0, R-861) | configs/felhom-os-apply | the ONE table of root-owned paths + the installer of them | ANY new root-owned file the installer writes (sudoers line, wrapper, unit) — add it to the table, never a new installer fetch (R-840) | The builder (`scripts/build-config-bundle.py`) and the installer read the same table; `test_every_root_file_the_installer_writes_is_in_the_bundle` fails on a path the bundle lacks. Trust files (`/etc/felhom/os-trust.json`, `operator-signers`) are NEVER bundle paths (R17) |
 | `osupdate.ConfigUpdateExecutor` | internal/osupdate/bundle.go | signed op `agent_config_update` {agent_version, bundle_sha256} | delivering the bundle to an installed box | a courier only: the root wrapper re-verifies signature, host, nonce and sha itself |
 | `osupdate.Leg.SendUnsent` / `lockPass` (v0.144.0, R-868) | internal/osupdate/unsent.go | `(ctx) int` | an OS-pass report the agent never sent (killed mid-pass): the wrapper keeps `report-<run>-<layer>-apply.json` beside the plan; the agent deletes it once the hub has it | any new caller that runs an apply pass must hold `lockPass` (flock, across processes) — the sender must never take a running pass's copy |
 | `osupdate.LoadSavedBlock` (v0.144.0, R-866) | internal/osupdate/leg.go | `(planDir) (block, savedAt, ok)` | the hub's newest os_update block as the daemon last received it (`os-update-block.json`) | the debug pass uses it ONLY when the hub cannot be reached, and says so in its header; no saved block → no pass |
 | `dpkg_state()` / `DPKG_STATE_SCRIPT` (v0.145.0, R-876) | configs/felhom-os-apply | `audit, journal = self.dpkg_state()` | dpkg's state in ONE call: `--audit` AND the update journal | never gate a repair on `--audit` alone — a crash leaves only the journal (measured); keep it one call (R-845) |
+| `thincheck.Checker.Check` + `Describe` (v0.159.0, R-1012) | internal/thincheck/thincheck.go | `Check(ctx, vmid) []Finding` | name an ORPHAN thin device (pool holds it, LVM does not) in a whole-guest backup failure — `BackupRunner.SetThinCheck` | READ-ONLY by design: the root helper `felhom-thin-check <vg> <pool>` sends only reserve/release_metadata_snap; an orphan needs two reads; fails open (never stops a backup). The repair stays human |
 | `guesthook.SnippetReady` / `Register` (v0.146.0, R-861) | internal/guesthook/install.go | `SnippetReady(path) error` | is the pre-start hook (a FIXED file from the bundle) in place — register only then | The agent never installs the hook (Proxmox runs it as root); a missing hookscript stops a guest start, so never `Register` without `SnippetReady` |
 
 ### Disk / format safety (role gates, durable IDs, format guards)
@@ -56,6 +57,7 @@
 | `SudoHostOps.EnsureMount` | internal/storage/hostops.go | `EnsureMount(ctx, MountSpec) error` | persistent by-UUID systemd .mount | Validate→render→stage→`install`→`enable --now`; idempotent |
 | `SudoHostOps.Unmount` | internal/storage/hostops.go | `Unmount(ctx, where) error` | detach a mount unit | DESTRUCTIVE — caller MUST have gated it; does not self-authorize |
 | `SudoHostOps.ReassertEnrolledMounts` | internal/storage/hostops.go | `ReassertEnrolledMounts(ctx)` | reboot remount (re-resolve by UUID) | Re-asserts unless mounted AND enabled (`shouldReassertMount`); skips absent UUIDs |
+| `SudoHostOps.ForgetRetiredDrives` (v0.159.0, R-1124) | internal/storage/forget_retired.go | `ForgetRetiredDrives(ctx, IntentReader) int` | drop the host mount unit of a DECOMMISSIONED drive that is GONE (start-up + the 20 s tick) | Only decommissioned + absent + unmounted; a present decommissioned drive keeps its unit (H3 re-enroll). The file goes only through `felhom-priv-apply unit-forget`, which re-checks all three as root |
 | `GuestBinder.AttachDrive` / `DetachDrive` | internal/localapi/intermediary.go | `AttachDrive(ctx, vmid, where) (guestPath, err)` | live drive hot-swap under `/mnt/felhom-drives` | Normalizes to EXACTLY ONE bind via `countHostMounts` (converges double-binds); force re-bind when guest can't see it |
 | `GuestBinder.EnsureSharedParent` | internal/localapi/intermediary.go | `EnsureSharedParent(ctx) error` | shared-parent bind + boot unit | make-private+make-shared ONLY on first bind — re-running orphans the guest's slave; F2-a: compares script AND unit for staleness |
 | `StablePathForRaw` / `DriveNameFromRaw` | internal/localapi/intermediary.go | `StablePathForRaw("/mnt/<n>") string` | raw host mount → stable guest path | Single source of truth both repos derive from |
@@ -112,8 +114,8 @@
 | `Engine.RunSignedJob` | internal/reconcile/job.go | `RunSignedJob(ctx, intent, signed, exec) JobResult` | executing a gated destructive job | Idempotency by nonce; journaled |
 | `escrow.Create` | internal/escrow/escrow.go | `Create(ctx, CreateOptions) (CreateResult, R, error)` | PBS-key escrow (zero-knowledge) | Recovery code returned SEPARATELY from the result (anti-log); self-verifies recoverability |
 | `escrow.GenerateRecoveryCode` / `joinSafe` / `RecoveryCodeSep` | internal/escrow/wordlist.go | `GenerateRecoveryCode() (string, error)` | minting the customer recovery code R | Draws from the EFF large list **filtered of every word containing `RecoveryCodeSep`** (4 entries: drop-down, felt-tip, t-shirt, yo-yo) so a code always segments back into exactly 10 words — a hyphenated word made codes ambiguous to transcribe AND flaked the test ~1/5 (v0.93.0). Generation-only: **already-issued codes stay valid**, R is verified as a whole passphrase and never re-split. Never count words by splitting the joined string — count what the generator drew |
-| `escrow.CeremonyBinary` / `CeremonyArgs()` / `CeremonyOutput` | internal/escrow/ceremony.go | the ONE fixed sudo self-invocation argv + the `--output=json` wire object (v1) | controller-driven ceremony (v0.88.0) | SINGLE SOURCE shared by the localapi exec, the capability manifest entry, and (byte-identically) the FELHOM_ESCROW sudoers line — `TestEscrowCeremonyArgvPinned` + `TestManifestCoveredBySudoers` lock all three. Never flag-helpers, never `--`→`-` (spike §2.2) |
+| `escrow.CeremonyBinary` / `CeremonyArgs()` / `CeremonyOutput` | internal/escrow/ceremony.go | the ONE fixed sudo self-invocation argv + the `--output=json` wire object (v2, R-978: json mode returns the package in `PendingUpload` instead of uploading) | controller-driven ceremony (v0.88.0) | SINGLE SOURCE shared by the localapi exec, the capability manifest entry, and (byte-identically) the FELHOM_ESCROW sudoers line — `TestEscrowCeremonyArgvPinned` + `TestManifestCoveredBySudoers` lock all three. Never flag-helpers, never `--`→`-` (spike §2.2) |
-| localapi escrow ceremony job | internal/localapi/escrow_ceremony.go | `POST /escrow/ceremony` + status + ONE-SHOT claim + preflight | the wizard's agent half | R lives ONLY in `Server.escrowR` (NEVER the job struct — snapshots must be structurally R-free); zeroed on claim/supersede/10-min TTL (`unclaimed_void`); in-memory BY DESIGN (restart loses R safely; re-run supersedes); subprocess stdout is SECRET-BEARING → parsed then zeroed, never logged |
+| localapi escrow ceremony job | internal/localapi/escrow_ceremony.go | `POST /escrow/ceremony` + status + ONE-SHOT claim + preflight | the wizard's agent half | R lives ONLY in `Server.escrowR` (NEVER the job struct — snapshots must be structurally R-free); zeroed on claim/supersede/10-min TTL (`unclaimed_void`); in-memory BY DESIGN (restart loses R safely; re-run supersedes); **R-978: the new package reaches the hub only at the claim, before R is handed out (`EscrowCeremonyConfig.Upload`); TTL and a new start wait while that upload is in flight — an abandoned ceremony leaves the previous package current**; subprocess stdout is SECRET-BEARING → parsed then zeroed, never logged |
 | `poke.Listener` + `poke.Port` | internal/poke/poke.go | `NewListener(resolve, trigger, port, logger)`; `poke.Port = 51822` | agent-plane immediate-sync (Direction-2a, v0.89.0) | Binds a contentless UDP socket EXCLUSIVELY to the box's WG /32 (`wgtunnel.LoadAssignedAddr`), fires the hub-loop out-of-band trigger. **Port 51822 is a SHARED cross-repo contract** — the hub poke sender + the ep0 `felhom-poke` forced-command target the SAME number; change one → change all three. Contentless (payload ignored), leading-edge debounced (`DebounceWindow`), WG-confined (kernel EKEYREJECTED refuses non-peer /32s). Wired only when `wg_tunnel.enabled` |
 | `wgtunnel.LoadAssignedAddr` | internal/wgtunnel/manager.go | `LoadAssignedAddr(stateDir) (netip.Addr, bool)` | the box's own WG /32 without a Manager | Reads `registered.json`; ok=false until registered; strips the /32 → bare addr (the poke bind target) |
 | `fasttick.Loop` + `fasttick.SourceFunc` | internal/fasttick/fasttick.go | `New(out chan<- struct{}, interval, logger, sources...)`; `Source.Unconverged() (bool, reason)` | agent-plane immediacy SECONDARY (v0.90.0, R-28): pulse the SAME out-of-band trigger every 30 s while ANY source is unconverged, self-disarm on convergence | STATE-BASED (no timer, nothing to journal). Every source MUST be a CACHED read (no exec/network per tick) — `desiredProvider.Generation()`, `reconcile.Engine.LastResult()` (Planned−Pending>0), `pbsdrLoop.PBSDRStatus().State=="waiting_secret"` ONLY, `wgtunnel.Manager.TunnelConvergence()`. The LOUD pbsdr states + destructive `pending_signature` are DELIBERATELY EXCLUDED (a stuck-loud box must not hammer). Pulses the cap-1 channel non-blocking (coalesces with poke/watchdog); NEVER touch `MinPollSeconds`/`clampInterval`/the ticker |
@@ -172,6 +174,7 @@
 | `capability` store-grant probe (`storeGrantStatuses` / `storeGrantVerdict` / `Client.Permissions`) | cmd/felhom-agent/main.go, internal/proxmox/query.go | *"may the agent READ this backup tier?"*, one `capability.Status` per configured tier | R-185. **Never infer permission from an empty content listing** — `{"data":[]}` is what a FORBIDDEN tier and a NEWBORN tier both return, and that ambiguity hid an unreadable host tier on both demo boxes. Ask `/access/permissions` **as the agent's own token** (root always says yes). **The ungranted answer is not empty and not a 403** — it carries the privileges inherited from the box-wide `/` grant, so test for **`Datastore.AllocateSpace`** specifically; path-presence or `Datastore.Audit` reports a blinded storage healthy. Probed set comes from `BackupTiers()`, never a fixed list. Critical except the `local` fallback. Composes AROUND the sudo prober (the `poolReadStatus` precedent); `Status`'s wire shape is untouched so the hub alert is free. Unreachable PVE ⇒ degraded, never ok. |
 | `backup.RestoreTestState` | internal/backup/restoretest_state.go | `RecordSuccess(target,archive,tier,verified,t)` / `ProvenArchive(target)` / `ProvenRestoreTests(ctx)` / `LastSuccess(target)` / `OldestFirst(targets)` | Per-tier restore-test PROOF state, persisted (atomic tmp+rename) — **which archive** was proven, and when (R-86) | **Credit ONLY on success** — a permanently failing tier must keep sorting first, or it looks freshly proven and stops being retried. Ties break on target id: without it, two tiers proven in the same second rotate by Go's randomised map order. **This one NEEDS persistence unlike R-84** — R-84 had ground truth to consult (the archive is still on the storage); a restore-test destroys its scratch and leaves no artifact. **R-86: the ARCHIVE is the state, the time is metadata** — a time alone cannot answer "have we proven THIS archive", which is the due-check's whole question. A pre-R-86 file (bare RFC3339 per target) keeps its time and yields NO proven archive, so each tier is due once after the upgrade; reading a legacy time as proof of the current archive would invent a guarantee. **R-189: it is also the REPORTABLE half of the restore-test signal.** The in-memory `backup.Store` holds only this process's latest run, and under per-archive due-ness the agent will not re-test a proven archive — so a proof lost to a restart is not repeated for a whole archive generation (observed live: a passing 14.5 GB offsite restore reached no host-report). `ProvenRestoreTests` renders the stored proofs as `hub.RestoreTest` entries and the collector merges them; a record missing the archive or the tier is NOT emitted, because an unproven tier reading as proven is worse than the defect. **Only successes are stored, deliberately:** a success suppresses future work, a failure causes it. |
 | `backup.BackupSuccessState` | internal/backup/backup_state.go | `RecordBackupSuccess(target, b)` / `LastKnownSuccess(target, vmid)` | Newest SUCCESSFUL backup per tier+guest, persisted (atomic tmp+rename) — the due-check's fallback when the tier's storage cannot be read after a restart (R-894) | **Read ONLY when the storage cannot be read** — a storage that answers is the ground truth (R-84), and an archive absent there must make the tier due even when this file remembers one. A saved copy older than the cadence still reads due. Only successes are written. |
+| localapi backup cadence cap (R-973) | internal/localapi/backup_cap.go | `backupCadenceCap(cadence)`, `Server.capRefuses`, `Server.BackupCapRefusal()` | POST /backup refuses (429) a run the `07` §6 schedule would never ask for: ≤24 h tier → 3 successes per rolling 24 h; longer tier → 1 per max(cadence−24 h, cadence/2) | Counts SUCCESSES only, from this process AND `BackupSuccessState.RecentBackupSuccesses` (so a restart does not reset it); refusals saved (`RecordBackupRefusal`) and raised for 24 h as the CRITICAL host-report status `localapi:backup-cadence-cap` (`backupCapStatus` in main) → the hub's existing `agent_capability_degraded`. Tests `TestBackupCap_*`. |
 | `hub.ProvenRestoreTestReporter` + `Collector.SetProvenRestoreTests` | internal/hub/collect.go | the DURABLE restore-test source, merged with the in-memory one | R-189. Merge rule: **one entry per tier, newest by `TestedAt` wins** — a fresh failure beats a stored success (the failure is the news, and it lives nowhere else), a stored success beats a stale in-memory entry after a restart, and a tier never appears twice (the hub would read two tests). An unparseable timestamp counts as OLDER, so a malformed entry cannot displace a good one. **The wiring is pinned by an AST test** — the method this replaced (`RestoreTestState.Snapshot`) carried a doc comment naming a host-report gauge and had no caller for weeks. |
 | `backup.SpecBuilder` / `backup.TierPicker` / `(*BackupRunner).PickSettledRestoreCandidateOn` | internal/backup/schedule.go, runner.go | `func(ctx,archive) RestoreTestSpec`; `func(ctx,target,notAfter) (archive,landed,error)` | The per-run restore-test spec + per-tier **settled** candidate lookup (R-85, widened by R-86) | The spec is built **PER RUN**, never frozen at construction — the pre-R-85 immediately-invoked value made the offsite tier unschedulable AND went stale on any config change. `SourceTier` comes from **the archive**, never the configured target (the v0.100.0 rule). A tier with no archive returns `("", zero, nil)` — **`""` is NOT an error**, or every fresh box looks broken for its first week. **R-86: `notAfter` is the settle cutoff** (zero = no cutoff, which is what keeps `PickRestoreCandidateOn` a one-line call into it), and the picker now skips entries failing `archivePlausiblyComplete` — under per-archive due-ness an incomplete phantom would be picked forever, fail forever, never earn proof, and make the tier due at EVERY evaluation. |
 | `localapi.BackupTier` + `normalizeBackupTiers` / `config.BackupConfig.BackupTiers` | internal/localapi/backup_tiers.go, internal/config/config.go | `normalizeBackupTiers(tiers, legacy, cadence) []BackupTier`; `BackupTiers() ([]BackupTier, []string)` | THE R-82 multi-tier resolution — one runner per tier, primary first | **The untargeted local-API contract is FROZEN**: no `?target=` ⇒ primary tier ⇒ pre-R-82 response BYTES (Target is `omitempty` and stays empty). Never default a missing cadence — reject it and log the warning at ERROR. Never share one retention knob between tiers. Jobs are keyed by (vmid,target). |

cmd/felhom-agent/backup_cap_r973_test.go

@@ -0,0 +1,67 @@
+package main
+
+import (
+	"go/ast"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/capability"
+)
+
+// R-973: a cadence-cap refusal becomes a CRITICAL degraded status — what the hub's existing
+// agent_capability_degraded event (operator-only) keys on — and clears when the refusal ages out.
+func TestBackupCapStatus_RefusalIsCriticalDegraded(t *testing.T) {
+	at := time.Date(2026, 10, 18, 3, 0, 0, 0, time.UTC)
+	s := backupCapStatus(func() (string, int, time.Time, bool) { return "felhom-pbs", 9201, at, true })
+	if !s.Critical || s.Status != capability.StatusDegraded || !strings.Contains(s.Reason, "9201") || !strings.Contains(s.Reason, "felhom-pbs") {
+		t.Fatalf("a refusal must raise a critical degraded status naming the guest and the tier: %+v", s)
+	}
+	if s := backupCapStatus(func() (string, int, time.Time, bool) { return "", 0, time.Time{}, false }); s.Status != capability.StatusOK || !s.Critical {
+		t.Fatalf("no refusal → ok (and still critical, so a later refusal alerts): %+v", s)
+	}
+	if s := backupCapStatus(nil); s.Status != capability.StatusOK {
+		t.Fatalf("no local API → ok: %+v", s)
+	}
+}
+
+// Wiring (the built-but-never-wired class): runDaemon's capability probe appends backupCapStatus, and the
+// late-bound source is set from the local API server.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): delete `backupCapRefusal = localSrv.BackupCapRefusal` →
+// "runDaemon never binds backupCapRefusal to the local API server". Restored.
+func TestR973_CapRefusalIsWiredIntoTheReport(t *testing.T) {
+	_, f := parseMain(t)
+	var appended, bound bool
+	for _, d := range f.Decls {
+		fd, ok := d.(*ast.FuncDecl)
+		if !ok || fd.Name == nil || fd.Name.Name != "runDaemon" || fd.Body == nil {
+			continue
+		}
+		ast.Inspect(fd.Body, func(n ast.Node) bool {
+			switch x := n.(type) {
+			case *ast.CallExpr:
+				if id, ok := x.Fun.(*ast.Ident); ok && id.Name == "backupCapStatus" && len(x.Args) == 1 {
+					if a, ok := x.Args[0].(*ast.Ident); ok && a.Name == "backupCapRefusal" {
+						appended = true
+					}
+				}
+			case *ast.AssignStmt:
+				if len(x.Lhs) == 1 && len(x.Rhs) == 1 {
+					if l, ok := x.Lhs[0].(*ast.Ident); ok && l.Name == "backupCapRefusal" {
+						if sel, ok := x.Rhs[0].(*ast.SelectorExpr); ok && sel.Sel.Name == "BackupCapRefusal" {
+							bound = true
+						}
+					}
+				}
+			}
+			return true
+		})
+	}
+	if !appended {
+		t.Fatal("runDaemon's capability probe does not call backupCapStatus(backupCapRefusal)")
+	}
+	if !bound {
+		t.Fatal("runDaemon never binds backupCapRefusal to the local API server")
+	}
+}

cmd/felhom-agent/escrow_r978_test.go

@@ -0,0 +1,112 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"go/ast"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/escrow"
+)
+
+// R-978, the root half: the controller-driven ceremony (json mode, the FELHOM_ESCROW argv with
+// `--upload`) must hand its package back instead of sending it — the daemon sends it at the claim.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): set `deferUpload: false` in escrowOptsFromCLI → this fails
+// with "json mode must defer the upload (R-978): got deferUpload=false", and the delivery test's json leg
+// then sends the package at create time ("put called 1 times"). Restored.
+func TestEscrowOptsFromCLI_JSONModeDefersUpload(t *testing.T) {
+	// The FELHOM_ESCROW argv: --selftest=escrow-create --upload --output=json.
+	if o := escrowOptsFromCLI("", false, false, true, "", true); !o.deferUpload || !o.upload {
+		t.Fatalf("json mode must defer the upload (R-978): got deferUpload=%v upload=%v", o.deferUpload, o.upload)
+	}
+	// Text mode (an operator at a terminal) keeps the immediate upload; no --upload means no upload at all.
+	if o := escrowOptsFromCLI("", false, false, true, "", false); o.deferUpload {
+		t.Fatal("text mode must upload at once")
+	}
+	if o := escrowOptsFromCLI("", false, false, false, "", true); o.deferUpload {
+		t.Fatal("without --upload there is nothing to defer")
+	}
+}
+
+func TestDeliverEscrowPackage_JSONModeSendsNothing(t *testing.T) {
+	cfg := config.Config{}
+	cfg.Hub.URL, cfg.Hub.HostID, cfg.Hub.APIKey = "https://hub.invalid", "host-1", "k"
+	res := escrow.CreateResult{Blob: []byte("ciphertext"), KeyFingerprint: "f2:87", Posture: escrow.PostureZeroKnowledge, IdentityBlob: []byte("id")}
+
+	for _, jsonMode := range []bool{true, false} {
+		calls := 0
+		var sent []byte
+		put := func(_ context.Context, _ config.Config, body []byte) error { calls++; sent = body; return nil }
+		o := escrowOptsFromCLI("", false, false, true, "", jsonMode)
+		pending, uploaded, err := deliverEscrowPackage(context.Background(), cfg, res, "abc", o.deferUpload, put)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if jsonMode {
+			if calls != 0 || uploaded {
+				t.Fatalf("json mode: put called %d times (uploaded=%v), want 0 — the package must wait for the claim", calls, uploaded)
+			}
+			var m map[string]any
+			if json.Unmarshal(pending, &m) != nil || m["blob_b64"] == "" || m["identity_blob_b64"] == "" || m["restic_pw_sha256"] != "abc" {
+				t.Fatalf("json mode: the pending package is not the hub upload body: %s", pending)
+			}
+		} else if calls != 1 || !uploaded || pending != nil || len(sent) == 0 {
+			t.Fatalf("text mode: put called %d times, uploaded=%v, pending=%q — want one immediate upload", calls, uploaded, pending)
+		}
+	}
+
+	// A hub that is not configured fails in both modes (as the pre-R-978 upload did).
+	if _, _, err := deliverEscrowPackage(context.Background(), config.Config{}, res, "", true, nil); err == nil {
+		t.Fatal("an unconfigured hub must fail the ceremony even when the upload is deferred")
+	}
+}
+
+// R-978 wiring (the built-but-never-wired class): runDaemon's localapi.EscrowCeremonyConfig carries an
+// Upload that calls putEscrowBody. Without it every claim of a new package answers 503 and no customer
+// can finish a ceremony.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): delete the `Upload:` line in runDaemon → this fails with
+// "runDaemon's EscrowCeremonyConfig has no Upload calling putEscrowBody". Restored.
+func TestR978_CeremonyUploadIsWiredIntoTheDaemon(t *testing.T) {
+	_, f := parseMain(t)
+	if !callsWithin(f, "main")["runDaemon"] {
+		t.Fatal("main → runDaemon is broken — the path this test asserts is not the live one")
+	}
+	wired := false
+	for _, d := range f.Decls {
+		fd, ok := d.(*ast.FuncDecl)
+		if !ok || fd.Name == nil || fd.Name.Name != "runDaemon" || fd.Body == nil {
+			continue
+		}
+		ast.Inspect(fd.Body, func(n ast.Node) bool {
+			cl, ok := n.(*ast.CompositeLit)
+			if !ok {
+				return true
+			}
+			sel, ok := cl.Type.(*ast.SelectorExpr)
+			if !ok || sel.Sel.Name != "EscrowCeremonyConfig" {
+				return true
+			}
+			for _, el := range cl.Elts {
+				if kv, ok := el.(*ast.KeyValueExpr); ok {
+					if k, ok := kv.Key.(*ast.Ident); ok && k.Name == "Upload" {
+						ast.Inspect(kv.Value, func(n ast.Node) bool {
+							if ce, ok := n.(*ast.CallExpr); ok {
+								if id, ok := ce.Fun.(*ast.Ident); ok && id.Name == "putEscrowBody" {
+									wired = true
+								}
+							}
+							return true
+						})
+					}
+				}
+			}
+			return true
+		})
+	}
+	if !wired {
+		t.Fatal("runDaemon's EscrowCeremonyConfig has no Upload calling putEscrowBody")
+	}
+}

cmd/felhom-agent/main.go

@@ -59,6 +59,7 @@ import (
 	"gitea.dooplex.hu/admin/felhom-agent/internal/selfupdate"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/signedjobs"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/storage"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/thincheck"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/vzpartial"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/wgtunnel"
 )
@@ -193,7 +194,7 @@ func main() {
 	flag.StringVar(&pbsStorage, "storage", "", "for --selftest=escrow-create: the pbs storage whose key to escrow (default: escrow.pbs_storage_id)")
 	flag.BoolVar(&paperkey, "paperkey", false, "for --selftest=escrow-create: ALSO emit the raw-key paperkey (opt-in (a); single-factor, unrevocable)")
 	flag.BoolVar(&offline, "offline", false, "for --selftest=escrow-create: ALSO emit the R-wrapped offline copy to print (opt-in (b))")
-	flag.BoolVar(&upload, "upload", false, "for --selftest=escrow-create: upload the opaque blob to the hub")
+	flag.BoolVar(&upload, "upload", false, "for --selftest=escrow-create: upload the opaque blob to the hub (with --output=json: hand the package back in pending_upload — the agent daemon uploads it when the code is claimed, R-978)")
 	flag.StringVar(&blobPath, "blob", "", "for --selftest=escrow-consume: path to the R-wrapped escrow blob file")
 	flag.StringVar(&expectedFP, "fingerprint", "", "for --selftest=escrow-consume: the EXPECTED key fingerprint (the gate target)")
 	flag.StringVar(&keyDest, "keydest", "", "for --selftest=escrow-consume: where to install the recovered key (0600)")
@@ -403,6 +404,31 @@ func (r *gateRemounter) Remount(ctx context.Context, t storage.KnownTarget) {
 // poolReadStatus probes the felhom-pool membership read (GET /pools/felhom) the stale-lock reaper's
 // ownership scoping depends on (v0.62.0, audit A1). Needs `Pool.Audit` at `/pool/felhom` — granted
 // by host-install v1.9.0; on an older ACL this reports degraded and the reaper fail-safes (skips).
+// backupCapStatusName is the host-report status the R-973 cadence cap raises.
+const backupCapStatusName = "localapi:backup-cadence-cap"
+
+// backupCapStatus turns the local-API backup cadence cap's newest refusal (R-973) into a CRITICAL
+// capability status: degraded for 24 h after a refusal, so the hub raises its operator-only
+// agent_capability_degraded event (no hub change). The Reason names the guest, the tier and the time, and
+// says what to do. nil source (local API off) → ok.
+func backupCapStatus(refusal func() (string, int, time.Time, bool)) capability.Status {
+	s := capability.Status{
+		Name:     backupCapStatusName,
+		Feature:  "whole-guest restore points: a guest asked for more whole-guest backups than its schedule allows, and the extra ones were refused (R-973)",
+		Critical: true,
+		Status:   capability.StatusOK,
+	}
+	if refusal == nil {
+		return s
+	}
+	if target, vmid, at, ok := refusal(); ok {
+		s.Status = capability.StatusDegraded
+		s.Reason = fmt.Sprintf("refused a whole-guest backup of guest %d on target %q at %s — more than the schedule allows; the guest's controller may be misbehaving or compromised: check the guest",
+			vmid, target, at.UTC().Format(time.RFC3339))
+	}
+	return s
+}
+
 func poolReadStatus(ctx context.Context, px *proxmox.Client) capability.Status {
 	s := capability.Status{
 		Name:     "pve:pool-read",
@@ -829,8 +855,13 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 		run: (&proxmox.ExecRunner{Mode: proxmox.RunnerMode(cfg.Privileged.Mode)}).Run,
 		log: logger,
 	}
+	// R-973: the local-API backup cadence cap's refusals ride the report as a CRITICAL status, so the hub's
+	// existing agent_capability_degraded event (operator-only) tells the operator. Late-bound like
+	// drConfigured: the local API server is built further down, before the loop starts.
+	var backupCapRefusal func() (string, int, time.Time, bool)
 	probeAll := func(ctx context.Context) []capability.Status {
 		out := append(capProber.Probe(ctx), poolReadStatus(ctx, px))
+		out = append(out, backupCapStatus(backupCapRefusal))
 		return append(out, storeGrantStatuses(ctx, px, cfg, grantRepairer)...)
 	}
 	// (The startup self-check log runs AFTER the pbsdr manager is wired below, so its snapshot
@@ -1091,6 +1122,9 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 		decommIntent = intentStore
 	}
 	decommExec := signedjobs.NewDecommissionExecutor(decommIntent, logger)
+	if guestBindStore != nil { // typed-nil guard, as above
+		decommExec.SetGuestBinds(guestBindStore) // v0.159.1: prune the bind record like the self-serve decommission
+	}
 
 	// Agent self-update (TASK D1): the agent_update executor downloads the operator-signed binary,
 	// verifies it against the SIGNED sha256, and hands it to the root guarded wrapper (A/B flip +
@@ -1193,6 +1227,9 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 			return cfg.Escrow.PBSStorageID
 		},
 		HubConfigured: cfg.Hub.URL != "" && cfg.Hub.HostID != "" && cfg.Hub.APIKey != "",
+		// R-978: the ceremony hands its package back; the claim uploads it with the same PUT the root
+		// subprocess used, under the same per-host key, before the code is handed out.
+		Upload: func(ctx context.Context, pkg []byte) error { return putEscrowBody(ctx, cfg, pkg) },
 		DRConfigured: func() bool {
 			if drConfigured != nil {
 				return drConfigured()
@@ -1202,6 +1239,7 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 	}
 	localSrv := buildLocalAPIServer(cfg, px, backupStore, heavyOps, observer, driveKnown, hostOps, gate, collector, client, intentRec, guestBindStore, formatJobStore, logRing, escrowCeremonyCfg, logger, &localTokens)
 	if localSrv != nil {
+		backupCapRefusal = localSrv.BackupCapRefusal // R-973
 		internalPlaceExec.SetEnsure(localSrv.EnsureInternalPlace)
 		// R-998 (doc 14): POST /host/poweroff — guest shutdown, then host power-off; guards in hostpoweroff.go.
 		pwMode := proxmox.RunnerMode(cfg.Privileged.Mode)
@@ -1425,6 +1463,9 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 		// re-assert below sees its unit.
 		localSrv.EnsureInternalPlace(ctx)
 		if mountReasserter != nil {
+			// R-1124: a decommissioned drive that is gone loses its host mount unit first (start-up and every tick), so
+			// no later boot waits for it and the re-assert below no longer lists it.
+			mountReasserter.ForgetRetiredDrives(ctx, intentReader)
 			mountReasserter.ReassertEnrolledMounts(ctx)
 		}
 		// F9: on startup (the host's bring-up/reconcile trigger), re-assert any enrolled guest data-drive
@@ -1463,6 +1504,7 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 					// unit re-disabled at runtime), then re-assert the guest binds against the now-live mounts.
 					localSrv.EnsureInternalPlace(ctx) // `16` §5: a record written since the last tick (the signed job)
 					if mountReasserter != nil {
+						mountReasserter.ForgetRetiredDrives(ctx, intentReader) // R-1124: decommissioned + gone → unit removed
 						mountReasserter.ReassertEnrolledMounts(ctx)
 					}
 					localSrv.ReassertGuestBinds(ctx)
@@ -2061,8 +2103,12 @@ func buildLocalAPIServer(cfg config.Config, px *proxmox.Client, store *backup.St
 		},
 		Logger: logger,
 	}
+	// R-1012: an orphan thin device (a full host root during a snapshot) is NAMED in the backup failure the controller
+	// mails to the operator, through the read-only root helper. Never removed — the repair stays a human step.
+	thinChecker := &thincheck.Checker{API: px, Runner: &proxmox.ExecRunner{Mode: gaMode, SudoPath: cfg.Privileged.SudoPath}, Logger: logger}
 	for _, r := range tierRunners {
 		r.SetBeforeBackup(func(ctx context.Context, storage string, vmid int) { cleaner.Clean(ctx, storage, vmid) })
+		r.SetThinCheck(func(ctx context.Context, vmid int) string { return thincheck.Describe(thinChecker.Check(ctx, vmid)) })
 	}
 	srv.SetPartialSweep(func(ctx context.Context, vmid int) {
 		for _, t := range tierTargets {
@@ -2814,6 +2860,20 @@ type escrowCeremonyOpts struct {
 	offline            bool
 	upload             bool
 	identityBundlePath string
+	// deferUpload (R-978): with upload, build the hub upload body and hand it back in
+	// CeremonyOutput.PendingUpload instead of sending it. Set for json mode — the controller-driven
+	// ceremony, whose daemon uploads the package only when the customer claims the new code.
+	deferUpload bool
+}
+
+// escrowOptsFromCLI maps the escrow-create flags onto the ceremony options. json mode defers the upload
+// (R-978, pinned by TestEscrowOptsFromCLI_JSONModeDefersUpload); text mode — an operator at a terminal,
+// who sees R at once — keeps the immediate upload.
+func escrowOptsFromCLI(storage string, paperkey, offline, upload bool, identityBundlePath string, jsonMode bool) escrowCeremonyOpts {
+	return escrowCeremonyOpts{
+		storage: storage, paperkey: paperkey, offline: offline, upload: upload,
+		identityBundlePath: identityBundlePath, deferUpload: upload && jsonMode,
+	}
 }
 
 // escrowCeremonyOutcome is the shared core's result. R is the ONLY secret; Sum mirrors the
@@ -2964,16 +3024,36 @@ func escrowCeremony(ctx context.Context, cfg config.Config, logger *slog.Logger,
 		ResticPwSealed:    resticStaged,
 	}
 	if opts.upload {
-		if err := uploadEscrowBlob(ctx, cfg, res, resticPwSHA256); err != nil {
+		pending, uploaded, err := deliverEscrowPackage(ctx, cfg, res, resticPwSHA256, opts.deferUpload, putEscrowBody)
+		if err != nil {
 			// R is minted and the blob self-verified — only the hub leg failed. kind "upload" lets
 			// the text shell keep the pre-extraction order (R surfaced, THEN the failure).
 			return out, &escrowCeremonyErr{kind: "upload", err: err}
 		}
-		out.Sum.Uploaded = true
+		out.Sum.PendingUpload, out.Sum.Uploaded = pending, uploaded
 	}
 	return out, nil
 }
 
+// deliverEscrowPackage builds the hub upload body and either sends it now (text mode) or returns it for
+// the caller to send later (deferUpload — json mode, R-978: the daemon sends it when the customer claims
+// the code). A hub that is not configured fails in both modes, as before: a package that can never be
+// delivered is not a ceremony. put is the sender seam (putEscrowBody in production).
+func deliverEscrowPackage(ctx context.Context, cfg config.Config, res escrow.CreateResult, resticPwSHA256 string, deferUpload bool,
+	put func(ctx context.Context, cfg config.Config, body []byte) error) (pending []byte, uploaded bool, err error) {
+	if cfg.Hub.URL == "" || cfg.Hub.HostID == "" || cfg.Hub.APIKey == "" {
+		return nil, false, fmt.Errorf("hub not configured (url/host_id/api_key)")
+	}
+	body := escrowUploadBody(res, resticPwSHA256)
+	if deferUpload {
+		return body, false, nil
+	}
+	if err := put(ctx, cfg, body); err != nil {
+		return nil, false, err
+	}
+	return nil, true, nil
+}
+
 // printEscrowTextBanner prints the text-mode header line (shared by the success and the
 // post-banner failure paths so the pre-extraction output stays byte-identical — including the
 // posture: the historical banner always printed the DEFAULT posture, not the result's).
@@ -3037,10 +3117,7 @@ func runSelftestEscrowCreate(ctx context.Context, cfg config.Config, logger *slo
 		return 2
 	}
 
-	out, cerr := escrowCeremony(ctx, cfg, logger, escrowCeremonyOpts{
+	out, cerr := escrowCeremony(ctx, cfg, logger, escrowOptsFromCLI(storage, paperkey, offline, upload, identityBundlePath, jsonMode))
-		storage: storage, paperkey: paperkey, offline: offline, upload: upload,
-		identityBundlePath: identityBundlePath,
-	})
 	if cerr != nil {
 		switch cerr.kind {
 		case "usage":
@@ -3078,6 +3155,9 @@ func runSelftestEscrowCreate(ctx context.Context, cfg config.Config, logger *slo
 		if out.Sum.Uploaded {
 			fmt.Fprintln(os.Stderr, "  uploaded the opaque blob(s) to the hub (host record); the hub cannot open them")
 		}
+		if len(out.Sum.PendingUpload) > 0 {
+			fmt.Fprintln(os.Stderr, "  package prepared, NOT uploaded: the agent uploads it when the new code is claimed (R-978)")
+		}
 		wire := out.Sum
 		wire.RecoveryCode = out.R
 		if err := json.NewEncoder(os.Stdout).Encode(wire); err != nil {
@@ -3231,13 +3311,10 @@ type escrowUploadRequest struct {
 	ResticPwSHA256 string `json:"restic_pw_sha256,omitempty"`
 }
 
-// uploadEscrowBlob PUTs the opaque blob (and, for 10D, the identity blob) to
+// escrowUploadBody is the hub upload body for one ceremony's result: the opaque blob (and, for 10D, the
-// the hub, authed with the per-host key. The hub stores ciphertext + non-secret fields; no usable
+// identity blob) plus non-secret fields. Ciphertext only — useless without R; no usable secret leaves
-// secret leaves the agent.
+// the agent. putEscrowBody sends it, authed with the per-host key.
-func uploadEscrowBlob(ctx context.Context, cfg config.Config, res escrow.CreateResult, resticPwSHA256 string) error {
+func escrowUploadBody(res escrow.CreateResult, resticPwSHA256 string) []byte {
-	if cfg.Hub.URL == "" || cfg.Hub.HostID == "" || cfg.Hub.APIKey == "" {
-		return fmt.Errorf("hub not configured (url/host_id/api_key)")
-	}
 	upReq := escrowUploadRequest{
 		BlobB64:        base64.StdEncoding.EncodeToString(res.Blob),
 		KeyFingerprint: res.KeyFingerprint,
@@ -3249,6 +3326,15 @@ func uploadEscrowBlob(ctx context.Context, cfg config.Config, res escrow.CreateR
 		upReq.IdentityBlobB64 = base64.StdEncoding.EncodeToString(res.IdentityBlob)
 	}
 	body, _ := json.Marshal(upReq)
+	return body
+}
+
+// putEscrowBody PUTs an escrow upload body to this host's hub record (the per-host key). Used by the
+// text-mode ceremony at once, and by the daemon when the customer claims the code (R-978).
+func putEscrowBody(ctx context.Context, cfg config.Config, body []byte) error {
+	if cfg.Hub.URL == "" || cfg.Hub.HostID == "" || cfg.Hub.APIKey == "" {
+		return fmt.Errorf("hub not configured (url/host_id/api_key)")
+	}
 	url := strings.TrimRight(cfg.Hub.URL, "/") + "/api/v1/hosts/" + cfg.Hub.HostID + "/escrow"
 	req, err := http.NewRequestWithContext(ctx, http.MethodPut, url, bytes.NewReader(body))
 	if err != nil {

configs/felhom-agent.sudoers

@@ -25,6 +25,7 @@
 
 Cmnd_Alias FELHOM_MOUNT = \
   /usr/local/sbin/felhom-priv-apply ^unit mnt-[A-Za-z0-9_.\\-]+\.(mount|automount)$, \
+  /usr/local/sbin/felhom-priv-apply ^unit-forget mnt-[A-Za-z0-9_.\\-]+\.mount$, \
   /usr/bin/systemctl daemon-reload, \
   /usr/bin/systemctl ^enable --now -- mnt-[A-Za-z0-9_.\\-]+\.mount$, \
   /usr/bin/systemctl ^disable -- mnt-[A-Za-z0-9_.\\-]+\.mount$, \
@@ -101,6 +102,11 @@ Cmnd_Alias FELHOM_VZPARTIAL = \
 # ruled size rule and formats ONLY that volume, ONLY when it carries no signature. The agent calls it from the signed
 # `add_internal_place` job (an old box), always with 0 GiB planned: its guest's volumes already exist. `plan` changes
 # nothing. The installer runs the helper as root itself (a new box).
+# R-1012 (agent v0.159.0): the READ-ONLY orphan-thin check before and after a whole-guest backup — lists the pool's thin
+# device ids (through a metadata snapshot it reserves and releases) next to LVM's. It sends no other pool message; the
+# repair stays a human step. Pinned by configs/test_felhom_thin_check.py.
+Cmnd_Alias FELHOM_THINCHECK = \
+  /usr/local/sbin/felhom-thin-check ^[A-Za-z0-9_+][A-Za-z0-9_.+-]{0\,63} [A-Za-z0-9_+][A-Za-z0-9_.+-]{0\,63}$
 Cmnd_Alias FELHOM_INTERNALPLACE = \
   /usr/local/sbin/felhom-internal-place ^(plan|create) 0$
 
@@ -336,4 +342,4 @@ Cmnd_Alias FELHOM_GUESTNET = \
   /usr/sbin/pct ^exec [0-9]+ -- pgrep -x dhclient$, \
   /usr/sbin/pct ^exec [0-9]+ -- dhclient -pf /run/dhclient\.eth0\.pid -lf /var/lib/dhcp/dhclient\.eth0\.leases eth0$
 
-felhom-agent ALL=(root) NOPASSWD: FELHOM_MOUNT, FELHOM_DISK, FELHOM_PROVISION, FELHOM_FORMAT, FELHOM_DNSMASQ, FELHOM_GUESTHOOK, FELHOM_INTERMEDIARY, FELHOM_CONTROLLERSWAP, FELHOM_STALELOCK, FELHOM_FSTRIM, FELHOM_NETMOUNT, FELHOM_WG, FELHOM_SELFUPDATE, FELHOM_SSHD, FELHOM_OOB, FELHOM_PBSDR, FELHOM_BACKUPTARGET, FELHOM_SELFHEAL, FELHOM_ESCROW, FELHOM_GUESTNET, FELHOM_SCRATCH_TEARDOWN, FELHOM_OSAPPLY, FELHOM_POWEROFF, FELHOM_VZPARTIAL, FELHOM_INTERNALPLACE
+felhom-agent ALL=(root) NOPASSWD: FELHOM_MOUNT, FELHOM_DISK, FELHOM_PROVISION, FELHOM_FORMAT, FELHOM_DNSMASQ, FELHOM_GUESTHOOK, FELHOM_INTERMEDIARY, FELHOM_CONTROLLERSWAP, FELHOM_STALELOCK, FELHOM_FSTRIM, FELHOM_NETMOUNT, FELHOM_WG, FELHOM_SELFUPDATE, FELHOM_SSHD, FELHOM_OOB, FELHOM_PBSDR, FELHOM_BACKUPTARGET, FELHOM_SELFHEAL, FELHOM_ESCROW, FELHOM_GUESTNET, FELHOM_SCRATCH_TEARDOWN, FELHOM_OSAPPLY, FELHOM_POWEROFF, FELHOM_VZPARTIAL, FELHOM_INTERNALPLACE, FELHOM_THINCHECK

configs/felhom-os-apply

@@ -191,6 +191,8 @@ BUNDLE_FILES = [
     ("/usr/local/sbin/felhom-backup-target-apply", "felhom-backup-target-apply", 0o755, "bash", "replace"),
     # R-1006 (agent v0.156.0): the only delete of backup files — a dead half-written vzdump, re-checked as root.
     ("/usr/local/sbin/felhom-vzdump-partial-clean", "felhom-vzdump-partial-clean", 0o755, "bash", "replace"),
+    # R-1012 (agent v0.159.0): the read-only orphan-thin check around a whole-guest backup (names an orphan, never removes).
+    ("/usr/local/sbin/felhom-thin-check", "felhom-thin-check", 0o755, "bash", "replace"),
     # felhom.eu architecture/16 §5 (agent v0.158.0): the internal data place — the only lvcreate/mkfs of a system-disk volume.
     ("/usr/local/sbin/felhom-internal-place", "felhom-internal-place", 0o755, "bash", "replace"),
     ("/usr/local/sbin/felhom-os-apply", "felhom-os-apply", 0o755, "python", "replace"),

configs/felhom-priv-apply

@@ -13,6 +13,8 @@ longer name the destination: each verb has a fixed source and a fixed (or strict
 
 Verbs (each one sudoers line, exact-match pattern):
   unit <name>              /var/lib/felhom-agent/units/<name> -> /etc/systemd/system/<name>   (.mount | .automount)
+  unit-forget <name>       REMOVES /etc/systemd/system/<name> (R-1124): only a felhom by-UUID drive .mount unit that is
+                           disabled, not mounted, and whose drive is absent (its /dev/disk/by-uuid link is gone)
   dnsmasq <tmp> <name>     /tmp/felhom-resolver-<digits>.conf -> /etc/dnsmasq.d/felhom-<...>.conf
   wg                       /var/lib/felhom-agent/wg/wg-felhom.conf -> /etc/wireguard/wg-felhom.conf (0600)
   sshd-config              /var/lib/felhom-agent/felhom-sshd/sshd_config -> /etc/felhom-sshd/sshd_config
@@ -41,7 +43,7 @@ WG_SRC, WG_DEST = STATE + "/wg/wg-felhom.conf", "/etc/wireguard/wg-felhom.conf"
 SSHD_SRC, SSHD_DEST = STATE + "/felhom-sshd/sshd_config", "/etc/felhom-sshd/sshd_config"
 KEY_SRC, KEY_DEST = STATE + "/felhom-sshd/authorized_keys.felhom-op", "/etc/felhom-sshd/authorized_keys/felhom-op"
 MAX_BYTES = 64 * 1024
-VERBS = ("unit", "dnsmasq", "wg", "sshd-config", "sshd-key", "controller-image")
+VERBS = ("unit", "unit-forget", "dnsmasq", "wg", "sshd-config", "sshd-key", "controller-image")
 # R-861 (a) A1 (`09` §3 decision 165): the SAME pattern as the agent's controllerImageRe (internal/localapi/
 # controllerswap.go) — a compromised agent cannot hand the guest's bootstrap any other image. Pinned by
 # configs/test_felhom_priv_apply.py ControllerImage.
@@ -51,6 +53,11 @@ CONTROLLER_IMAGE_MAX = 256
 VMID_RE = re.compile(r"^[0-9]{1,9}$")
 
 UNIT_NAME_RE = re.compile(r"^mnt-[A-Za-z0-9_.\\-]+\.(mount|automount)$")
+DRIVE_UNIT_NAME_RE = re.compile(r"^mnt-[A-Za-z0-9_.\\-]+\.mount$")
+DRIVE_UNIT_MARKER = "Managed by felhom-agent"
+NET_UNIT_MARKER = "felhom network storage"
+WANTS_DIR = UNIT_DIR + "/multi-user.target.wants"
+BY_UUID_DIR = "/dev/disk/by-uuid"
 DNSMASQ_TMP_RE = re.compile(r"^/tmp/felhom-resolver-[0-9]+\.conf$")
 DNSMASQ_NAME_RE = re.compile(r"^felhom-[a-z0-9][a-z0-9._-]*\.conf$")
 SEG = r"[A-Za-z0-9_-][A-Za-z0-9_.-]*"
@@ -135,6 +142,36 @@ class Host:
     def read_stdin(self, limit):
         return sys.stdin.buffer.read(limit + 1)
 
+    # ---- unit-forget (R-1124) ----
+    def read_installed(self, path):
+        """The installed unit: None when absent; Refused when it is not a regular file (a symlink is never followed)."""
+        try:
+            fd = os.open(path, os.O_RDONLY | os.O_NOFOLLOW | os.O_CLOEXEC)
+        except FileNotFoundError:
+            return None
+        except OSError as e:
+            raise Refused("F1", f"cannot open {path}: {e.strerror}")
+        with os.fdopen(fd, "rb") as f:
+            if not stat.S_ISREG(os.fstat(f.fileno()).st_mode):
+                raise Refused("F1", f"{path} is not a regular file")
+            return f.read(MAX_BYTES + 1)
+
+    def lexists(self, path):
+        return os.path.lexists(path)
+
+    def mountpoints(self):
+        """Every mount point in /proc/self/mountinfo (field 5, octal escapes decoded)."""
+        out = set()
+        with open("/proc/self/mountinfo", "r", encoding="utf-8", errors="replace") as f:
+            for line in f:
+                parts = line.split(" ")
+                if len(parts) > 4:
+                    out.add(re.sub(r"\\([0-7]{3})", lambda m: chr(int(m.group(1), 8)), parts[4]))
+        return out
+
+    def remove(self, path):
+        os.remove(path)
+
     def write_guest_image(self, vmid, data):
         """As root: `pct exec <vmid> -- tee <the fixed file>` with the checked ref on stdin (no shell)."""
         r = subprocess.run(["/usr/sbin/pct", "exec", str(vmid), "--", "tee", CONTROLLER_IMAGE_FILE],
@@ -424,6 +461,45 @@ def controller_image(rest, host):
     return 0
 
 
+def unit_forget(rest, host):
+    """R-1124: remove ONE felhom drive mount unit that the agent has already disabled, whose drive is gone. A drive that
+    is still present or still mounted keeps its unit — this verb can only forget, never unmount or take a drive away."""
+    try:
+        if len(rest) != 1 or not DRIVE_UNIT_NAME_RE.match(rest[0]) or "/" in rest[0]:
+            raise Refused("A1", "usage: felhom-priv-apply unit-forget mnt-<escaped path>.mount")
+        name = rest[0]
+        path = os.path.join(UNIT_DIR, name)
+        data = host.read_installed(path)
+        if data is None:
+            host.log(f"felhom-priv-apply: ABSENT unit-forget {path} (nothing to remove)")
+            return 0
+        text = text_of(data, path)
+        if DRIVE_UNIT_MARKER not in text or NET_UNIT_MARKER in text:
+            raise Refused("F2", f"{name} is not a felhom drive unit")
+        check_unit(name, text)
+        s = parse_ini(text, name)
+        what, where = s["Mount"].get("What", ""), s["Mount"]["Where"]
+        mu = re.match(r"^/dev/disk/by-uuid/(.+)$", what)
+        if not mu:
+            raise Refused("F2", f"{name} is not a by-UUID drive unit")
+        if host.lexists(os.path.join(WANTS_DIR, name)):
+            raise Refused("F3", f"{name} is still enabled (disable it first)")
+        if where in host.mountpoints():
+            raise Refused("F4", f"{where} is mounted")
+        if host.lexists(os.path.join(BY_UUID_DIR, mu.group(1))):
+            raise Refused("F5", f"the drive of {name} is present (filesystem UUID {mu.group(1)})")
+    except Refused as e:
+        host.log(f"felhom-priv-apply: REFUSED [{e.rule}] unit-forget {' '.join(rest)[:120]}: {e.reason}")
+        return 2 if e.rule == "A1" else 3
+    try:
+        host.remove(path)
+    except OSError as e:
+        host.log(f"felhom-priv-apply: FAILED unit-forget {path}: {e}")
+        return 4
+    host.log(f"felhom-priv-apply: REMOVED unit-forget {path} (where={where})")
+    return 0
+
+
 def main(argv, host=None):
     host = host or Host()
     if argv == ["--self-check"]:
@@ -445,6 +521,8 @@ def main(argv, host=None):
         return 0
     if argv and argv[0] == "controller-image":
         return controller_image(argv[1:], host)
+    if argv and argv[0] == "unit-forget":
+        return unit_forget(argv[1:], host)
     try:
         verb, src, dest, mode, checker = plan(argv)
         data = host.read_source(src)

configs/felhom-thin-check

@@ -0,0 +1,74 @@
+#!/bin/bash
+# felhom-thin-check — list the thin device ids a thin pool holds next to the ones LVM knows (R-1012). READ-ONLY: it
+# names an orphan, it never removes one. Installed as /usr/local/sbin/felhom-thin-check (config bundle).
+#
+#   felhom-thin-check <vg> <pool>
+#
+# A full host root while vzdump takes its snapshot can leave a thin snapshot inside the pool that LVM no longer knows
+# (demo-hp 2026-10-14: `thin_ls` device 4, LVM only 1 and 2); from then on every snapshot backup of that guest fails
+# („Creation of new snapshot 4 of device 2 failed") and nothing said why. The repair (`dmsetup message … delete <id>`)
+# stays a HUMAN step, decided by the operator — this helper only reads.
+#
+# Reading the live pool's device list needs a metadata snapshot: the kernel's documented way to read metadata while
+# the pool is in use (`reserve_metadata_snap`, `thin_ls --metadata-snap`, `release_metadata_snap`). It changes no data
+# and no mapping; it is held for the length of one `thin_ls -o DEV` (well under a second on these pools). Rules:
+#   1. the arguments are an LVM name pair; <vg>/<pool> is a thin pool (lv_attr starts with `t`);
+#   2. only one run at a time (a lock file); a metadata snapshot that someone ELSE holds is never touched — the run
+#      refuses (exit 3) and does not release it;
+#   3. LVM's list is read before AND after the pool's, so a volume created in between is not reported as an orphan
+#      (the caller treats an id as an orphan only when it is in the pool and in neither LVM list).
+# Output, one record per line:  lvm-before <id> <lv> | pool <id> | lvm-after <id> <lv> | ok
+# Exit: 0 read, 1 refused (arguments / not a thin pool), 3 a metadata snapshot is already held or cannot be reserved,
+# 4 the pool could not be read. Pinned by configs/test_felhom_thin_check.py.
+set -u
+LVS=/usr/sbin/lvs
+DMSETUP=/usr/sbin/dmsetup
+THIN_LS=/usr/sbin/thin_ls
+LOCK=/run/lock/felhom-thin-check.lock
+NAME_RE='^[A-Za-z0-9_+][A-Za-z0-9_.+-]{0,63}$'
+
+die() { echo "felhom-thin-check: REFUSED: $1" >&2; exit "${2:-1}"; }  # $1 the reason, $2 the exit code
+
+lvm_ids() { # $1 tag → "<tag> <id> <lv>" for every thin volume of this pool
+	"$LVS" --noheadings --separator , -o lv_name,thin_id --select "pool_lv=$POOL" -- "$VG" 2>/dev/null |
+		awk -F, -v tag="$1" '{ gsub(/ /, ""); if ($2 ~ /^[0-9]+$/) print tag, $2, $1 }'
+}
+
+release() { "$DMSETUP" message "$TPOOL" 0 release_metadata_snap >/dev/null 2>&1; }
+
+main() {
+	[[ $# -eq 2 ]] || die "usage: felhom-thin-check <vg> <pool>"
+	VG=$1 POOL=$2
+	local attr held ids
+	[[ "$VG" =~ $NAME_RE && "$POOL" =~ $NAME_RE ]] || die "not an LVM name pair"
+
+	attr=$("$LVS" --noheadings -o lv_attr -- "$VG/$POOL" 2>/dev/null | tr -d ' ')
+	[[ "$attr" == t* ]] || die "$VG/$POOL is not a thin pool"
+
+	DM="${VG//-/--}-${POOL//-/--}"
+	TPOOL="$DM-tpool"
+	TMETA="/dev/mapper/${DM}_tmeta"
+
+	exec 9>"$LOCK" || die "cannot open the lock file" 4
+	flock -w 30 9 || die "another check holds the lock" 3
+
+	# Field 7 of the thin-pool status is the held metadata root: "-" when no metadata snapshot exists.
+	held=$("$DMSETUP" status "$TPOOL" 2>/dev/null | awk '$3 == "thin-pool" { print $7 }')
+	[[ -n "$held" ]] || die "cannot read the status of $TPOOL" 4
+	[[ "$held" == "-" ]] || die "a metadata snapshot of $TPOOL is already held (by someone else) — not touched" 3
+
+	lvm_ids lvm-before
+	"$DMSETUP" message "$TPOOL" 0 reserve_metadata_snap >/dev/null 2>&1 || die "could not reserve a metadata snapshot of $TPOOL" 3
+	trap release EXIT
+	trap 'release; trap - EXIT; exit 4' INT TERM HUP
+	if ! ids=$("$THIN_LS" --metadata-snap --no-headers -o DEV "$TMETA" 2>/dev/null); then
+		die "thin_ls could not read $TMETA" 4
+	fi
+	release
+	trap - EXIT INT TERM HUP
+	awk '$1 ~ /^[0-9]+$/ { print "pool", $1 }' <<<"$ids"
+	lvm_ids lvm-after
+	echo ok
+}
+
+if [[ "${BASH_SOURCE[0]}" == "$0" ]]; then main "$@"; fi

configs/test_felhom_priv_apply.py

@@ -376,5 +376,67 @@ class ControllerImage(unittest.TestCase):
         self.assertIn("controller-image", buf.getvalue())
 
 
+class UnitForget(unittest.TestCase):
+    """R-1124: the agent forgets a decommissioned drive whose disk is gone — the root half removes its unit file, and
+    ONLY a felhom by-UUID drive unit that is disabled, not mounted, and whose /dev/disk/by-uuid link is absent.
+    RED-PROOF: on the pre-R-1124 wrapper `unit-forget` is not a verb (A1 usage, rc 2) — the removal case fails."""
+    NAME = "mnt-hdd_1.mount"
+    PATH = "/etc/systemd/system/mnt-hdd_1.mount"
+    UUID_LINK = "/dev/disk/by-uuid/91d2dc2d-2d28-4929-9bdd-3e11fa2f41ae"
+
+    def host(self, text=LOCAL_UNIT, kind="file", links=(), mounted=()):
+        h = FakeHost()
+        h.installed = {} if text is None else {self.PATH: (kind, text.encode())}
+        h.links, h.mounted, h.removed = set(links), set(mounted), []
+        h.read_installed = lambda p: (None if p not in h.installed else
+                                      (_ for _ in ()).throw(pa.Refused("F1", f"{p} is not a regular file"))
+                                      if h.installed[p][0] != "file" else h.installed[p][1])
+        h.lexists = lambda p: p in h.links
+        h.mountpoints = lambda: set(h.mounted)
+        h.remove = lambda p: h.removed.append(p)
+        return h
+
+    def test_an_absent_disabled_unmounted_drive_unit_is_removed(self):
+        h = self.host()
+        self.assertEqual(run(h, "unit-forget", self.NAME), 0, h.logs)
+        self.assertEqual(h.removed, [self.PATH])
+
+    def test_already_gone_is_fine(self):
+        h = self.host(text=None)
+        self.assertEqual(run(h, "unit-forget", self.NAME), 0, h.logs)
+        self.assertEqual(h.removed, [])
+
+    def refused(self, h, rule, *argv):
+        rc = run(h, *(argv or ("unit-forget", self.NAME)))
+        self.assertIn(rc, (2, 3), h.logs)
+        self.assertEqual(h.removed, [], h.logs)
+        self.assertTrue(any(f"[{rule}]" in l for l in h.logs), h.logs)
+
+    def test_F5_a_present_drive_keeps_its_unit(self):
+        self.refused(self.host(links=[self.UUID_LINK]), "F5")
+
+    def test_F4_a_mounted_drive_keeps_its_unit(self):
+        self.refused(self.host(mounted=["/mnt/hdd_1"]), "F4")
+
+    def test_F3_an_enabled_unit_is_not_removed(self):
+        self.refused(self.host(links=["/etc/systemd/system/multi-user.target.wants/" + self.NAME]), "F3")
+
+    def test_F2_a_network_unit_or_a_foreign_unit_is_not_removed(self):
+        self.refused(self.host(text=LOCAL_UNIT.replace("Managed by felhom-agent", "Written by someone")), "F2")
+        net = LOCAL_UNIT.replace("# Managed by felhom-agent", "# Managed by felhom-agent felhom network storage")
+        self.refused(self.host(text=net), "F2")
+
+    def test_F1_a_symlinked_unit_is_not_followed(self):
+        self.refused(self.host(kind="symlink"), "F1")
+
+    def test_A1_names(self):
+        for argv in (("unit-forget", "mnt-hdd_1.automount"), ("unit-forget", "sshd.service"),
+                     ("unit-forget", "../mnt-hdd_1.mount"), ("unit-forget",), ("unit-forget", self.NAME, "x")):
+            self.refused(self.host(), "A1", *argv)
+
+    def test_U3_a_unit_whose_name_lies_about_where_is_not_removed(self):
+        self.refused(self.host(text=LOCAL_UNIT.replace("Where=/mnt/hdd_1", "Where=/mnt/other")), "U3")
+
+
 if __name__ == "__main__":
     unittest.main(verbosity=2)

configs/test_felhom_thin_check.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+"""Tests for configs/felhom-thin-check (R-1012): it lists the thin pool's device ids next to LVM's and NEVER removes
+anything. Fake `lvs`, `dmsetup` and `thin_ls` record every call; nothing real is touched. Each refusal has a positive
+control (the healthy read), so a test cannot pass because the helper refuses everything.
+Run: python3 configs/test_felhom_thin_check.py   (also run by internal/thincheck's Go test)
+"""
+import os, pathlib, re, subprocess, tempfile, unittest
+
+HELPER = pathlib.Path(__file__).resolve().parent / "felhom-thin-check"
+
+
+class Box:
+    def __init__(self, pool_ids=(1, 2), lvm=(("vm-9201-disk-0", 1), ("vm-9201-disk-1", 2)), attr="twi-aotz--",
+                 held="-", reserve_rc=0, thin_ls_rc=0):
+        self.t = tempfile.TemporaryDirectory()
+        r = pathlib.Path(self.t.name)
+        self.log = r / "calls"
+        self.log.write_text("")
+        lvm_lines = "\\n".join(f"  {n},{i}" for n, i in lvm)
+        pool_lines = "\\n".join(str(i) for i in pool_ids)
+        self.bins = {}
+        for name, body in {
+            "lvs": f'''case "$*" in *lv_attr*) echo "  {attr}";; *thin_id*) printf "{lvm_lines}\\n";; esac''',
+            "dmsetup": f'''case "$1" in
+  status) echo "0 112975872 thin-pool 585 6860/262144 517080/882624 {held} rw discard_passdown queue_if_no_space - 1024";;
+  message) [ "$4" = reserve_metadata_snap ] && exit {reserve_rc}; exit 0;;
+esac''',
+            "thin_ls": f'''[ {thin_ls_rc} -ne 0 ] && exit {thin_ls_rc}; printf "{pool_lines}\\n"''',
+        }.items():
+            p = r / name
+            p.write_text(f'#!/bin/bash\necho "{name} $*" >> "{self.log}"\n{body}\n')
+            p.chmod(0o755)
+            self.bins[name] = p
+        self.lock = r / "lock"
+
+    def run(self, *args):
+        script = (f'source "{HELPER}"; LVS="{self.bins["lvs"]}"; DMSETUP="{self.bins["dmsetup"]}"; '
+                  f'THIN_LS="{self.bins["thin_ls"]}"; LOCK="{self.lock}"; main "$@"')
+        r = subprocess.run(["bash", "-c", script, "x", *args], capture_output=True, text=True)
+        return r.returncode, r.stdout, r.stderr
+
+    def calls(self):
+        return self.log.read_text().splitlines()
+
+
+def messages(calls):
+    return [c.split()[-1] for c in calls if c.startswith("dmsetup message")]
+
+
+class Reads(unittest.TestCase):
+    def test_healthy_pool_lists_both_sides_and_releases(self):
+        b = Box()
+        rc, out, err = b.run("pve", "data")
+        self.assertEqual(rc, 0, err)
+        self.assertEqual(out.splitlines(), ["lvm-before 1 vm-9201-disk-0", "lvm-before 2 vm-9201-disk-1",
+                                            "pool 1", "pool 2",
+                                            "lvm-after 1 vm-9201-disk-0", "lvm-after 2 vm-9201-disk-1", "ok"])
+        self.assertEqual(messages(b.calls()), ["reserve_metadata_snap", "release_metadata_snap"])
+        self.assertIn("thin_ls --metadata-snap --no-headers -o DEV /dev/mapper/pve-data_tmeta", b.calls())
+
+    def test_the_2026_10_14_orphan_is_listed(self):
+        rc, out, _ = Box(pool_ids=(1, 2, 4)).run("pve", "data")
+        self.assertEqual(rc, 0)
+        self.assertIn("pool 4", out.splitlines())
+
+    def test_dashed_names_map_to_the_device_mapper_name(self):
+        b = Box()
+        rc, _, err = b.run("felhom-scratch", "thin-pool")
+        self.assertEqual(rc, 0, err)
+        self.assertIn("dmsetup status felhom--scratch-thin--pool-tpool", b.calls())
+
+
+class Refuses(unittest.TestCase):
+    def refused(self, box, rc_want, *args):
+        rc, out, err = box.run(*args)
+        self.assertEqual(rc, rc_want, (out, err))
+        self.assertNotIn("ok", out.splitlines())
+        return box.calls()
+
+    def test_arguments(self):
+        for args in (("pve",), ("pve", "data", "x"), ("pve;id", "data"), ("-pve", "data"), ("pve", "../data"),
+                     ("pve", "da ta")):
+            calls = self.refused(Box(), 1, *args)
+            self.assertEqual(calls, [], args)
+
+    def test_not_a_thin_pool(self):
+        calls = self.refused(Box(attr="Vwi-aotz--"), 1, "pve", "vm-9201-disk-1")
+        self.assertEqual(messages(calls), [])
+
+    def test_a_snapshot_someone_else_holds_is_never_touched(self):
+        b = Box(held="1234")
+        rc, _, err = b.run("pve", "data")
+        self.assertEqual(rc, 3)
+        self.assertEqual(messages(b.calls()), [])  # neither reserve nor release
+        # the reason line ends with the reason, not the exit code (found live on demo-hp 2026-10-17)
+        self.assertTrue(err.strip().endswith("not touched"), err)
+
+    def test_reserve_failure_does_not_release(self):
+        calls = self.refused(Box(reserve_rc=1), 3, "pve", "data")
+        self.assertEqual(messages(calls), ["reserve_metadata_snap"])
+
+    def test_thin_ls_failure_still_releases(self):
+        calls = self.refused(Box(thin_ls_rc=1), 4, "pve", "data")
+        self.assertEqual(messages(calls), ["reserve_metadata_snap", "release_metadata_snap"])
+
+
+class NeverDeletes(unittest.TestCase):
+    def test_the_only_pool_messages_are_reserve_and_release(self):
+        src = HELPER.read_text()
+        sent = re.findall(r'"\$DMSETUP" message "\$TPOOL" 0 (\S+)', src)
+        self.assertEqual(sorted(set(sent)), ["release_metadata_snap", "reserve_metadata_snap"])
+        for word in ("delete", "lvremove", "thin_repair", "create_thin", "create_snap", " rm "):
+            code = "\n".join(l for l in src.splitlines() if not l.lstrip().startswith("#"))
+            self.assertNotIn(word, code, word)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

go.mod

@@ -1,7 +1,8 @@
 module gitea.dooplex.hu/admin/felhom-agent
 
-go 1.25.0
+go 1.27.2
 
-require golang.org/x/crypto v0.52.0
+require (
-
+	golang.org/x/crypto v0.58.0
-require golang.org/x/sys v0.45.0 // indirect
+	golang.org/x/sys v0.49.0
+)

go.sum

@@ -1,6 +1,6 @@
-golang.org/x/crypto v0.52.0 h1:RMs7fP2rXdep0CftQlK8Uf+kibLm7qkCcradZWYz988=
+golang.org/x/crypto v0.58.0 h1:COkYLr4k7nDI4r6QtMML2AfzeIswWHAJthLS6K/M/54=
-golang.org/x/crypto v0.52.0/go.mod h1:1QgfPxDqh0T2M/elOJtp9RvuR95kVjir0e6/BvEmGbc=
+golang.org/x/crypto v0.58.0/go.mod h1:Xh+kl5A+M0gMiWDytH2Hhr5Z7WbNCpv/eB/EFgfSk+w=
-golang.org/x/sys v0.45.0 h1:dO4czNzziLiiXplLQgBCEpCvXQ3dnkn0SdaZSYdQ+FY=
+golang.org/x/sys v0.49.0 h1:XbzkgYJHdqh/8m2Uu0W/dQv8nktxx4BFHp1M0gROTXA=
-golang.org/x/sys v0.45.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.49.0/go.mod h1:hNLxWAXmnKAxqDtdwIYC4bM9oQPEecfsnNMuSxOs3og=
-golang.org/x/term v0.43.0 h1:S4RLU2sB31O/NCl+zFN9Aru9A/Cq2aqKpTZJ6B+DwT4=
+golang.org/x/term v0.47.0 h1:cpyOwQkr2Nzxe2p6KnrTG7Ei2wtzjsuSjhpXXepZ/BM=
-golang.org/x/term v0.43.0/go.mod h1:lrhlHNdQJHO+1qVYiHfFKVuVioJIheAc3fBSMFYEIsk=
+golang.org/x/term v0.47.0/go.mod h1:efAfAuG6RKABvg1ArGebRkvX6LXul14SM67o2e4mUCY=

internal/backup/backup_state.go

@@ -39,13 +39,25 @@ type savedSuccess struct {
 	target string
 	vmid   int
 	at     time.Time
+	// recent (R-973) holds the start times of the newest successes (at most maxRecentSuccesses, newest
+	// last) — what the local-API backup cadence cap counts, so an agent restart does not reset the cap.
+	recent []time.Time
+	// refusedAt (R-973) is the newest time the cap refused a backup of this guest on this tier.
+	refusedAt time.Time
 }
 
-// backupSuccessJSON is one entry on disk.
+// maxRecentSuccesses bounds the per-tier success history kept for the cadence cap (R-973). The cap
+// allows at most 3 runs in a window, so 8 is ample.
+const maxRecentSuccesses = 8
+
+// backupSuccessJSON is one entry on disk. Recent and LastRefusedAt are additive (R-973): a file written
+// before them reads as "no history", and an older agent ignores them.
 type backupSuccessJSON struct {
-	Target    string `json:"target"`
+	Target        string   `json:"target"`
-	VMID      int    `json:"vmid"`
+	VMID          int      `json:"vmid"`
-	StartedAt string `json:"started_at"`
+	StartedAt     string   `json:"started_at"`
+	Recent        []string `json:"recent,omitempty"`
+	LastRefusedAt string   `json:"last_refused_at,omitempty"`
 }
 
 func backupStateKey(target string, vmid int) string { return target + "/" + strconv.Itoa(vmid) }
@@ -67,7 +79,19 @@ func NewBackupSuccessState(path string) *BackupSuccessState {
 		if perr != nil {
 			continue // one unreadable entry must not lose the others
 		}
-		s.last[backupStateKey(e.Target, e.VMID)] = savedSuccess{target: e.Target, vmid: e.VMID, at: t.UTC()}
+		ss := savedSuccess{target: e.Target, vmid: e.VMID, at: t.UTC()}
+		for _, r := range e.Recent {
+			if rt, rerr := time.Parse(time.RFC3339, r); rerr == nil {
+				ss.recent = append(ss.recent, rt.UTC())
+			}
+		}
+		if len(ss.recent) == 0 {
+			ss.recent = []time.Time{ss.at} // a pre-R-973 file: its one known success counts
+		}
+		if rt, rerr := time.Parse(time.RFC3339, e.LastRefusedAt); rerr == nil {
+			ss.refusedAt = rt.UTC()
+		}
+		s.last[backupStateKey(e.Target, e.VMID)] = ss
 	}
 	return s
 }
@@ -85,13 +109,72 @@ func (s *BackupSuccessState) RecordBackupSuccess(target string, b hub.Backup) er
 	s.mu.Lock()
 	defer s.mu.Unlock()
 	k := backupStateKey(target, b.VMID)
-	if old, ok := s.last[k]; ok && !t.After(old.at) {
+	old, had := s.last[k]
+	if had && !t.After(old.at) {
 		return nil
 	}
-	s.last[k] = savedSuccess{target: target, vmid: b.VMID, at: t.UTC()}
+	ss := savedSuccess{target: target, vmid: b.VMID, at: t.UTC(), refusedAt: old.refusedAt}
+	ss.recent = append(append([]time.Time(nil), old.recent...), t.UTC())
+	if len(ss.recent) > maxRecentSuccesses {
+		ss.recent = ss.recent[len(ss.recent)-maxRecentSuccesses:]
+	}
+	s.last[k] = ss
 	return s.saveLocked()
 }
 
+// RecentBackupSuccesses returns the saved success start times for this tier and guest that are not
+// older than since (R-973 — the cadence cap's restart-proof count).
+func (s *BackupSuccessState) RecentBackupSuccesses(target string, vmid int, since time.Time) []time.Time {
+	if s == nil {
+		return nil
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	var out []time.Time
+	for _, t := range s.last[backupStateKey(target, vmid)].recent {
+		if !t.Before(since) {
+			out = append(out, t)
+		}
+	}
+	return out
+}
+
+// RecordBackupRefusal saves that the cadence cap refused a backup of this guest on this tier at at
+// (R-973), so the operator signal it raises survives an agent restart.
+func (s *BackupSuccessState) RecordBackupRefusal(target string, vmid int, at time.Time) error {
+	if s == nil {
+		return nil
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	k := backupStateKey(target, vmid)
+	ss, ok := s.last[k]
+	if !ok {
+		return nil // nothing to cap against without a success; the cap never refuses then
+	}
+	if !at.After(ss.refusedAt) {
+		return nil
+	}
+	ss.refusedAt = at.UTC()
+	s.last[k] = ss
+	return s.saveLocked()
+}
+
+// LastBackupRefusal returns the newest saved cadence-cap refusal across every tier and guest.
+func (s *BackupSuccessState) LastBackupRefusal() (target string, vmid int, at time.Time, ok bool) {
+	if s == nil {
+		return "", 0, time.Time{}, false
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, e := range s.last {
+		if !e.refusedAt.IsZero() && e.refusedAt.After(at) {
+			target, vmid, at, ok = e.target, e.vmid, e.refusedAt, true
+		}
+	}
+	return target, vmid, at, ok
+}
+
 // LastKnownSuccess returns the newest saved success for this tier and guest (ok=false = none on file).
 func (s *BackupSuccessState) LastKnownSuccess(target string, vmid int) (time.Time, bool) {
 	if s == nil {
@@ -129,7 +212,14 @@ func (s *BackupSuccessState) KnownBackupSuccesses() []hub.Backup {
 func (s *BackupSuccessState) saveLocked() error {
 	entries := make([]backupSuccessJSON, 0, len(s.last))
 	for _, e := range s.last {
-		entries = append(entries, backupSuccessJSON{Target: e.target, VMID: e.vmid, StartedAt: e.at.Format(time.RFC3339)})
+		j := backupSuccessJSON{Target: e.target, VMID: e.vmid, StartedAt: e.at.Format(time.RFC3339)}
+		for _, r := range e.recent {
+			j.Recent = append(j.Recent, r.Format(time.RFC3339))
+		}
+		if !e.refusedAt.IsZero() {
+			j.LastRefusedAt = e.refusedAt.Format(time.RFC3339)
+		}
+		entries = append(entries, j)
 	}
 	// Deterministic file content (Go's map order is random).
 	sort.Slice(entries, func(i, j int) bool {

internal/backup/backup_state_test.go

@@ -74,3 +74,40 @@ func TestBackupSuccessState_KnownBackupSuccessesSurviveAReopen(t *testing.T) {
 		t.Fatal("uncovered_volumes must marshal as [], not null")
 	}
 }
+
+// R-973: the cadence cap's history and refusals survive a reopen; the history is bounded; a file
+// written before R-973 counts its one known success (never "no history" for a tier that has a copy).
+func TestBackupSuccessState_R973_RecentHistoryAndRefusalSurviveAReopen(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	s := NewBackupSuccessState(path)
+	base := time.Date(2026, 10, 18, 2, 30, 0, 0, time.UTC)
+	for i := 0; i < 10; i++ {
+		at := base.Add(time.Duration(i) * time.Hour)
+		if err := s.RecordBackupSuccess("local", hub.Backup{VMID: 9201, Success: true, StartedAt: at.Format(time.RFC3339)}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	refused := base.Add(11 * time.Hour)
+	if err := s.RecordBackupRefusal("local", 9201, refused); err != nil {
+		t.Fatal(err)
+	}
+	r := NewBackupSuccessState(path)
+	all := r.RecentBackupSuccesses("local", 9201, time.Time{})
+	if len(all) != maxRecentSuccesses || !all[len(all)-1].Equal(base.Add(9*time.Hour)) {
+		t.Fatalf("recent history after a reopen: %v (want the newest %d, ending at +9h)", all, maxRecentSuccesses)
+	}
+	if got := r.RecentBackupSuccesses("local", 9201, base.Add(7*time.Hour)); len(got) != 3 {
+		t.Fatalf("successes since +7h: %d, want 3", len(got))
+	}
+	if tg, vm, at, ok := r.LastBackupRefusal(); !ok || tg != "local" || vm != 9201 || !at.Equal(refused) {
+		t.Fatalf("refusal after a reopen: %q %d %v %v", tg, vm, at, ok)
+	}
+
+	old := filepath.Join(t.TempDir(), "old.json")
+	if err := os.WriteFile(old, []byte(`[{"target":"felhom-pbs","vmid":9201,"started_at":"2026-10-17T03:00:00Z"}]`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if got := NewBackupSuccessState(old).RecentBackupSuccesses("felhom-pbs", 9201, time.Time{}); len(got) != 1 {
+		t.Fatalf("a pre-R-973 file must count its one success: %v", got)
+	}
+}

internal/backup/r1012_thincheck_test.go

@@ -0,0 +1,67 @@
+package backup
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// R-1012: a snapshot backup that fails carries the orphan finding in its error — the text the controller mails to
+// the operator (`whole_guest_backup_failed`). The check runs ONCE, after the failure — never before the run.
+// RED-PROOF (P1/redproof-thincheck.txt): with explainFailure's append removed this test fails. RED-PROOF (2026-10-17
+// night, felhom-agent REPORT.md): with the v0.159.0 pre-run check restored, calls = 2 and this test fails.
+func TestBackup_SnapshotFailure_NamesTheOrphan(t *testing.T) {
+	const why = "probable cause: thin pool pve/data holds device 4 that LVM does not know"
+	for name, api := range map[string]*fakeBackupAPI{
+		"task fails":  {vzdumpUPID: "UPID:x", waitErr: errors.New("task failed: job errors")},
+		"start fails": {vzdumpErr: errors.New("vzdump: 500")},
+	} {
+		t.Run(name, func(t *testing.T) {
+			r := NewBackupRunner(api, "local", "", "", "", quiet())
+			calls := 0
+			r.SetThinCheck(func(context.Context, int) string { calls++; return why })
+			rec, err := r.Backup(context.Background(), 9201)
+			if err == nil || rec.Success {
+				t.Fatalf("want a failure, got %+v", rec)
+			}
+			if !strings.Contains(rec.Error, why) {
+				t.Fatalf("the failure does not name the orphan: %q", rec.Error)
+			}
+			if calls != 1 {
+				t.Fatalf("thin check ran %d times, want 1 (after the failure only)", calls)
+			}
+		})
+	}
+}
+
+// A successful backup never runs the check (it takes a metadata snapshot of the live pool, and nobody needs the answer)
+// and keeps a clean record; a stop-mode backup takes no snapshot, so the check never runs; an unwired runner behaves
+// exactly as before. RED-PROOF (2026-10-17 night): with the v0.159.0 pre-run check restored, the success leg fails
+// with calls=1.
+func TestBackup_ThinCheck_SilentWhenNotApplicable(t *testing.T) {
+	ok := &fakeBackupAPI{vzdumpUPID: "UPID:x",
+		content: []proxmox.StorageContent{{VolID: "local:backup/vzdump-lxc-9201-x", Content: "backup", VMID: 9201, Size: 100, CTime: 1}}}
+	r := NewBackupRunner(ok, "local", "", "", "", quiet())
+	calls := 0
+	r.SetThinCheck(func(context.Context, int) string { calls++; return "probable cause: x" })
+	rec, err := r.Backup(context.Background(), 9201)
+	if err != nil || !rec.Success || rec.Error != "" || calls != 0 {
+		t.Fatalf("success: err=%v rec=%+v calls=%d (want no check on a successful backup, clean record)", err, rec, calls)
+	}
+
+	stop := NewBackupRunner(&fakeBackupAPI{vzdumpUPID: "UPID:x", waitErr: errors.New("boom")}, "local", proxmox.ModeStop, "", "", quiet())
+	calls = 0
+	stop.SetThinCheck(func(context.Context, int) string { calls++; return "probable cause: x" })
+	rec, _ = stop.Backup(context.Background(), 9201)
+	if calls != 0 || strings.Contains(rec.Error, "probable cause") {
+		t.Fatalf("stop mode: calls=%d error=%q", calls, rec.Error)
+	}
+
+	plain := NewBackupRunner(&fakeBackupAPI{vzdumpUPID: "UPID:x", waitErr: errors.New("boom")}, "local", "", "", "", quiet())
+	if rec, _ := plain.Backup(context.Background(), 9201); rec.Error != "boom" {
+		t.Fatalf("unwired: error = %q, want the task error unchanged", rec.Error)
+	}
+}

internal/backup/runner.go

@@ -61,8 +61,13 @@ type BackupRunner struct {
 	// beforeBackup runs just before the space preflight (R-1006): it removes a dead half-written vzdump of this guest on
 	// this target, so a leftover from a power cut cannot hold the space this backup needs. nil = nothing.
 	beforeBackup func(ctx context.Context, storage string, vmid int)
-	logger       *slog.Logger
+	// thinCheck (R-1012) names an orphan thin device in the guest's pools in plain words ("" = none / not checked). It
-	now          func() time.Time
+	// runs ONLY when a snapshot-mode vzdump has failed (start or task), so the failure the controller mails to the
+	// operator says why; a successful backup never runs it (it takes a metadata snapshot of the live pool). It never
+	// removes anything. nil = not wired.
+	thinCheck func(ctx context.Context, vmid int) string
+	logger    *slog.Logger
+	now       func() time.Time
 	// rejected remembers the volids already announced by warnRejectedArchiveOnce, so an incomplete
 	// archive is reported ONCE rather than on every 5-minute due-check. Bounded in practice: one
 	// entry per aborted upload, and a process restart clears it. Guarded by rejectedMu because the
@@ -110,6 +115,22 @@ func (r *BackupRunner) SetBeforeBackup(fn func(ctx context.Context, storage stri
 	r.beforeBackup = fn
 }
 
+// SetThinCheck wires the R-1012 orphan check (see BackupRunner.thinCheck).
+func (r *BackupRunner) SetThinCheck(fn func(ctx context.Context, vmid int) string) { r.thinCheck = fn }
+
+// explainFailure appends the orphan finding to a failed snapshot-mode backup's error — the only place the check runs, so
+// it sees the pool as the failure left it. Bounded, so a stuck helper cannot hold the job.
+func (r *BackupRunner) explainFailure(ctx context.Context, vmid int, rec *hub.Backup) {
+	if r.thinCheck == nil || r.mode != proxmox.ModeSnapshot {
+		return
+	}
+	cctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 2*time.Minute)
+	defer cancel()
+	if why := r.thinCheck(cctx, vmid); why != "" {
+		rec.Error = rec.Error + " — " + why
+	}
+}
+
 // localPruneSpec returns the `--prune-backups` spec to apply to THIS backup, or "" to skip pruning. It
 // applies the configured retention ONLY when the target is a non-PBS storage — PBS offsite retention is a
 // separate lifecycle and must never be pruned by the per-run flag (§9). Fail-safe: if the target's type
@@ -201,12 +222,15 @@ func (r *BackupRunner) backup(ctx context.Context, vmid int, onSnapshot func())
 		return rec, fmt.Errorf("backup: %s", rec.Error)
 	}
 
+	// R-1012: no orphan check here. It takes a metadata snapshot of the live pool, and its answer is only needed when the
+	// snapshot step has failed — explainFailure runs it then (pinned by TestBackup_ThinCheck_SilentWhenNotApplicable).
 	upid, err := r.api.Vzdump(ctx, proxmox.VzdumpOptions{
 		VMID: vmid, Storage: r.target, Mode: r.mode, Notes: r.notes,
 		PruneBackups: r.localPruneSpec(ctx), // local target → keep-last=N; PBS/unknown → "" (no prune)
 	})
 	if err != nil {
 		rec.Error = err.Error()
+		r.explainFailure(ctx, vmid, &rec)
 		rec.DurationSeconds = time.Since(start).Seconds()
 		return rec, fmt.Errorf("backup: vzdump vmid %d: %w", vmid, err)
 	}
@@ -220,6 +244,7 @@ func (r *BackupRunner) backup(ctx context.Context, vmid int, onSnapshot func())
 		}
 		if _, err := r.api.WaitTask(ctx, upid, proxmox.WaitOptions{Timeout: r.waitTimeout}); err != nil {
 			rec.Error = err.Error()
+			r.explainFailure(ctx, vmid, &rec)
 			rec.DurationSeconds = time.Since(start).Seconds()
 			return rec, fmt.Errorf("backup: vzdump task vmid %d: %w", vmid, err)
 		}

internal/capability/manifest.go

@@ -91,9 +91,11 @@ var manifest = []Capability{
 	{"disk-mkfs-xfs", "guarded format (xfs)", "/usr/local/sbin/felhom-mkfs-guarded", []string{"/dev/sda", "xfs"}, false, ""},
 	{"disk-smart", "disk SMART health read", "/usr/sbin/smartctl", []string{"-a", "-j", "/dev/sda"}, false, ""},
 	{"disk-lvs", "thin-pool usage read", "/usr/sbin/lvs", []string{"--reportformat", "json", "--units", "b", "-o", "lv_name,data_percent,metadata_percent", "--", "pve/data"}, false, ""},
+	{"thin-orphan-check", "orphan thin device check after a failed whole-guest backup (read-only, R-1012)", "/usr/local/sbin/felhom-thin-check", []string{"pve", "data"}, false, ""},
 
 	// ---- Storage mount units (watchdog re-mount) ----
 	{"mount-unit-install", "fs-UUID mount unit install (root content check, R-861)", "/usr/local/sbin/felhom-priv-apply", []string{"unit", "mnt-felhom\\x2dx.mount"}, false, ""},
+	{"mount-unit-forget", "decommissioned-drive unit removal (root check, R-1124)", "/usr/local/sbin/felhom-priv-apply", []string{"unit-forget", "mnt-felhom\\x2dx.mount"}, false, ""},
 	{"mount-daemon-reload", "systemd reload after unit write", "/usr/bin/systemctl", []string{"daemon-reload"}, false, ""},
 	{"mount-unit-enable", "mount unit enable", "/usr/bin/systemctl", []string{"enable", "--now", "--", "mnt-felhom\\x2dx.mount"}, false, ""},
 	{"mount-unit-disable", "mount unit disable", "/usr/bin/systemctl", []string{"disable", "--", "mnt-felhom\\x2dx.mount"}, false, ""},

internal/escrow/ceremony.go

@@ -1,5 +1,7 @@
 package escrow
 
+import "encoding/json"
+
 // Controller-driven ceremony contract (v0.88.0, TASK 2026-07-13; mechanics validated by
 // felhom.eu/documentation/audits/SPIKE-controller-escrow-2026-07-13.md). The agent's local API
 // re-invokes the agent binary as root via `sudo -n` with ONE fixed argument vector; sudoers
@@ -43,7 +45,15 @@ type CeremonyOutput struct {
 	IdentityBlobBytes int     `json:"identity_blob_bytes"`
 	ResticPwSealed    bool    `json:"restic_pw_sealed"`
 	Uploaded          bool    `json:"uploaded"`
+	// PendingUpload (version 2, R-978) is the hub upload body — the R-wrapped blobs (ciphertext), the key
+	// fingerprint and the sealed password's hash — that json mode hands back INSTEAD of uploading. The agent
+	// daemon uploads it only when the customer claims the new recovery code, so a ceremony that is started
+	// and abandoned never replaces the package the customer's previous code opens. Not a secret without R,
+	// but never logged. Uploaded is false whenever this is set.
+	PendingUpload json.RawMessage `json:"pending_upload,omitempty"`
 }
 
-// CeremonyOutputVersion is the current CeremonyOutput.Version value.
+// CeremonyOutputVersion is the current CeremonyOutput.Version value. 2 (R-978): json mode no longer
-const CeremonyOutputVersion = 1
+// uploads; it returns PendingUpload. The daemon refuses any other version (a binary/daemon mismatch
+// fails the ceremony instead of losing a package).
+const CeremonyOutputVersion = 2

internal/localapi/backup_cap.go

@@ -0,0 +1,146 @@
+package localapi
+
+import (
+	"fmt"
+	"net/http"
+	"sort"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// The whole-guest backup cadence cap (R-973, security review 2026-10-12).
+//
+// POST /backup is reachable from inside the guest, and every successful run prunes the tier
+// (`--prune-backups keep-last=N` on a local tier; ep0's server-side keep-last 2 on the off-site tier). So
+// a compromised controller calling it in a loop pushed every pre-compromise whole-guest restore point out
+// within minutes — the threat `07` §8 row 10 counts the off-site tier as the answer to. The agent now
+// refuses a run that the schedule in `07` §6 would never ask for:
+//
+//   - a tier whose cadence is at most 24 h (the local tier): at most 3 successful runs in any rolling
+//     24 h. The schedule is one night run; a „Mentés most" press makes a LOCAL copy and, since R-899, the
+//     night still takes its own after it, which can land inside the same 24 h as the previous night (the
+//     window opens anywhere in [W+2h, W+6h)). So a normal day needs at most 3: two nights plus one press.
+//     A second press inside 24 h is refused — the margin is spent on the night, which must never be refused
+//     because of the household's presses.
+//   - a longer tier (the 7-day off-site tier): at most 1 successful run per max(cadence − 24 h, cadence/2)
+//     — 6 days for 7 days. The controller asks /backup/due first and runs only when the newest copy is older
+//     than the cadence, and the button never targets this tier (`07` §6, controller v0.301.0), so a
+//     legitimate run never comes sooner; 24 h of slack absorbs clock and archive-time differences.
+//
+// Only SUCCESSES count — a failed run prunes nothing and must stay retryable. The count survives an agent
+// restart: it reads the success history saved on disk (LastKnownBackupStore, the R-894 file) as well as
+// this process's own record. A refusal answers 429, is logged at WARN, is saved, and raises the operator
+// signal BackupCapRefusal feeds (a critical capability in the host report → the hub's existing
+// `agent_capability_degraded` event, operator-only). Pinned by TestBackupCap_* (backup_cap_test.go).
+
+// backupCapRefusalHold is how long one refusal keeps the operator signal raised.
+const backupCapRefusalHold = 24 * time.Hour
+
+// backupCadenceCap returns the rolling window and the most successful runs allowed in it for a tier.
+func backupCadenceCap(cadence time.Duration) (window time.Duration, max int) {
+	if cadence <= 0 {
+		cadence = defaultBackupCadence
+	}
+	if cadence <= 24*time.Hour {
+		return 24 * time.Hour, 3
+	}
+	w := cadence - 24*time.Hour
+	if w < cadence/2 {
+		w = cadence / 2
+	}
+	return w, 1
+}
+
+// recentCapRuns returns the start times of the successful runs on this tier for this guest inside the
+// window ending now — this process's record and the saved one, each counted once. Caller holds no lock.
+func (s *Server) recentCapRuns(key backupJobKey, since time.Time) []time.Time {
+	seen := map[int64]bool{}
+	var out []time.Time
+	add := func(t time.Time) {
+		if t.Before(since) || seen[t.Unix()] {
+			return
+		}
+		seen[t.Unix()] = true
+		out = append(out, t)
+	}
+	s.capMu.Lock()
+	for _, t := range s.capRuns[key] {
+		add(t)
+	}
+	s.capMu.Unlock()
+	if s.lastKnown != nil {
+		for _, t := range s.lastKnown.RecentBackupSuccesses(key.target, key.vmid, since) {
+			add(t)
+		}
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Before(out[j]) })
+	return out
+}
+
+// recordCapRun notes a successful run for the cap (the saved half is written by RecordBackupSuccess).
+func (s *Server) recordCapRun(key backupJobKey, b hub.Backup) {
+	t, err := time.Parse(time.RFC3339, b.StartedAt)
+	if err != nil {
+		t = s.now()
+	}
+	s.capMu.Lock()
+	defer s.capMu.Unlock()
+	if s.capRuns == nil {
+		s.capRuns = map[backupJobKey][]time.Time{}
+	}
+	runs := append(s.capRuns[key], t.UTC())
+	if len(runs) > 8 {
+		runs = runs[len(runs)-8:]
+	}
+	s.capRuns[key] = runs
+}
+
+// capRefuses applies the cap to a new run request. It writes the 429 and returns true when refused.
+func (s *Server) capRefuses(w http.ResponseWriter, key backupJobKey, tier BackupTier) bool {
+	window, max := backupCadenceCap(tier.Cadence)
+	now := s.now()
+	runs := s.recentCapRuns(key, now.Add(-window))
+	if len(runs) < max {
+		return false
+	}
+	next := runs[len(runs)-max].Add(window)
+	s.capMu.Lock()
+	s.capRefusal = capRefusal{target: key.target, vmid: key.vmid, at: now.UTC()}
+	s.capMu.Unlock()
+	if s.lastKnown != nil {
+		if err := s.lastKnown.RecordBackupRefusal(key.target, key.vmid, now); err != nil {
+			s.logger.Warn("local-api: could not save the backup refusal on disk (R-973)", "vmid", key.vmid, "target", key.target, "err", err)
+		}
+	}
+	s.logger.Warn("local-api: backup REFUSED — more whole-guest backups than the schedule allows (R-973 cadence cap; check the guest)",
+		"vmid", key.vmid, "target", key.target, "runs_in_window", len(runs), "allowed", max,
+		"window_h", int(window.Hours()), "next_allowed", next.UTC().Format(time.RFC3339))
+	writeStatus(w, http.StatusTooManyRequests, false, nil, fmt.Sprintf(
+		"too many whole-guest backups on target %q: %d successful in the last %d h, the schedule allows %d — refused (the next one is allowed after %s)",
+		key.target, len(runs), int(window.Hours()), max, next.UTC().Format(time.RFC3339)))
+	return true
+}
+
+type capRefusal struct {
+	target string
+	vmid   int
+	at     time.Time
+}
+
+// BackupCapRefusal reports the newest cadence-cap refusal still inside the 24 h hold (this process's or
+// the saved one) — the source of the operator signal in the host report (R-973). ok=false = none.
+func (s *Server) BackupCapRefusal() (target string, vmid int, at time.Time, ok bool) {
+	s.capMu.Lock()
+	r := s.capRefusal
+	s.capMu.Unlock()
+	if s.lastKnown != nil {
+		if t, v, a, found := s.lastKnown.LastBackupRefusal(); found && a.After(r.at) {
+			r = capRefusal{target: t, vmid: v, at: a}
+		}
+	}
+	if r.at.IsZero() || s.now().Sub(r.at) > backupCapRefusalHold {
+		return "", 0, time.Time{}, false
+	}
+	return r.target, r.vmid, r.at, true
+}

internal/localapi/backup_cap_test.go

@@ -0,0 +1,293 @@
+package localapi
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+	"path/filepath"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// R-973 — the whole-guest backup cadence cap. A compromised controller calling POST /backup in a loop
+// pushed out every pre-compromise restore point: each success prunes the tier. These tests model the tier
+// the way Proxmox and ep0 keep it (keep-last N, pruned after each success) and ask what is LEFT.
+
+// keepLastTier is a whole-guest tier: every successful run adds an archive stamped with the clock and
+// prunes to keepLast, as vzdump's --prune-backups keep-last (local) and ep0's keep-last 2 (off-site) do.
+type keepLastTier struct {
+	mu       sync.Mutex
+	clock    *testClock
+	keepLast int
+	archives []time.Time
+	fail     bool
+	runs     int
+}
+
+func (k *keepLastTier) BackupWithSnapshotHook(_ context.Context, vmid int, _ func()) (hub.Backup, error) {
+	k.mu.Lock()
+	defer k.mu.Unlock()
+	k.runs++
+	now := k.clock.Now()
+	if k.fail {
+		return hub.Backup{VMID: vmid, StartedAt: now.Format(time.RFC3339)}, fmt.Errorf("vzdump failed")
+	}
+	k.archives = append(k.archives, now)
+	if len(k.archives) > k.keepLast {
+		k.archives = k.archives[len(k.archives)-k.keepLast:]
+	}
+	return hub.Backup{VMID: vmid, Success: true, Archive: "x", StartedAt: now.Format(time.RFC3339)}, nil
+}
+
+// oldestBefore reports whether the tier still holds an archive made before t (a pre-compromise point).
+func (k *keepLastTier) oldestBefore(t time.Time) bool {
+	k.mu.Lock()
+	defer k.mu.Unlock()
+	return len(k.archives) > 0 && k.archives[0].Before(t)
+}
+
+func (k *keepLastTier) count() (runs, kept int) {
+	k.mu.Lock()
+	defer k.mu.Unlock()
+	return k.runs, len(k.archives)
+}
+
+type testClock struct {
+	mu sync.Mutex
+	t  time.Time
+}
+
+func (c *testClock) Now() time.Time          { c.mu.Lock(); defer c.mu.Unlock(); return c.t }
+func (c *testClock) Advance(d time.Duration) { c.mu.Lock(); c.t = c.t.Add(d); c.mu.Unlock() }
+
+// capServer builds the two-tier box of `07` §6: local 24 h keep-last 3, off-site 7 d keep-last 2, the
+// success history on disk at path (a second capServer on the same path is an agent restart).
+func capServer(t *testing.T, path string, clock *testClock, local, offsite *keepLastTier) *Server {
+	t.Helper()
+	srv, err := NewServer(Options{
+		ListenAddr: "127.0.0.1:0", Guests: &fakeGuests{}, Backups: &fakeBackups{}, Store: &fakeStore{},
+		Storage: fakeStorage{targets: []hub.StorageTarget{{Name: "local"}, {Name: "felhom-pbs"}}},
+		Tokens:  staticTokens{"A": 8200},
+		BackupTiers: []BackupTier{
+			{TargetID: "local", Cadence: 24 * time.Hour, Primary: true, Service: local},
+			{TargetID: "felhom-pbs", Cadence: 7 * 24 * time.Hour, Service: offsite},
+		},
+		LastKnownBackups: backup.NewBackupSuccessState(path),
+		Logger:           slog.New(slog.NewTextHandler(io.Discard, nil)),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	srv.baseCtx = context.Background()
+	srv.now = clock.Now
+	return srv
+}
+
+// runBackup POSTs one backup and waits for the job to end. Returns the HTTP status of the POST.
+func runBackup(t *testing.T, srv *Server, target string) int {
+	t.Helper()
+	rr := do(t, srv.Handler(), "POST", "/backup?target="+target, "A", "")
+	if rr.Code == http.StatusAccepted {
+		waitFor(t, func() bool {
+			srv.jobsMu.Lock()
+			defer srv.jobsMu.Unlock()
+			j := srv.jobs[backupJobKey{vmid: 8200, target: target}]
+			return j != nil && !backupInFlight(j.Phase)
+		})
+	}
+	return rr.Code
+}
+
+// The off-site tier — the one `07` §8 row 10 counts on against ransomware. Before the cap ten calls in a
+// row left ep0 holding only the attacker's two copies. Now ONE runs (the tier was due), the other nine
+// are refused, and the newest pre-compromise copy survives; five days of flooding later it still does.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): make capRefuses always return false → this fails with
+// "the off-site tier lost every pre-compromise copy: 10 runs" . Restored.
+func TestBackupCap_OffsiteFloodKeepsAPreCompromiseCopy(t *testing.T) {
+	clock := &testClock{t: testNow}
+	local := &keepLastTier{clock: clock, keepLast: 3}
+	off := &keepLastTier{clock: clock, keepLast: 2}
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	srv := capServer(t, path, clock, local, off)
+
+	// The honest history: two weekly copies, the newest 7 days and 1 hour old (so the tier is due).
+	for _, age := range []time.Duration{14*24*time.Hour + time.Hour, 7*24*time.Hour + time.Hour} {
+		clock.t = testNow.Add(-age)
+		if code := runBackup(t, srv, "felhom-pbs"); code != http.StatusAccepted {
+			t.Fatalf("seeding the honest weekly copy: %d", code)
+		}
+	}
+	clock.t = testNow
+	compromise := testNow
+
+	refused := 0
+	for i := 0; i < 10; i++ {
+		if runBackup(t, srv, "felhom-pbs") == http.StatusTooManyRequests {
+			refused++
+		}
+		clock.Advance(time.Minute)
+	}
+	runs, _ := off.count()
+	if !off.oldestBefore(compromise) {
+		t.Fatalf("the off-site tier lost every pre-compromise copy: %d runs", runs-2)
+	}
+	if runs-2 != 1 || refused != 9 {
+		t.Fatalf("flood: %d runs, %d refused — want 1 run (the tier was due) and 9 refused", runs-2, refused)
+	}
+	for d := 0; d < 5; d++ { // five more days of flooding, every hour
+		for h := 0; h < 24; h++ {
+			clock.Advance(time.Hour)
+			if code := runBackup(t, srv, "felhom-pbs"); code != http.StatusTooManyRequests {
+				t.Fatalf("day %d hour %d: an off-site run was allowed %v after the last one", d, h, clock.Now().Sub(compromise))
+			}
+		}
+	}
+	if !off.oldestBefore(compromise) {
+		t.Fatal("five days of flooding pushed out the pre-compromise off-site copy")
+	}
+}
+
+// The local tier: a flood runs exactly 3 in a day (the schedule's own day: two nights + one press) and
+// every further request is refused. Honest about the limit: with keep-last 3 those 3 replace the 3 kept
+// points — the cap slows a flush from minutes to a day; it cannot keep a point a 3-per-day pace prunes.
+func TestBackupCap_LocalFloodRunsAtMostThreeADay(t *testing.T) {
+	clock := &testClock{t: testNow}
+	local := &keepLastTier{clock: clock, keepLast: 3}
+	srv := capServer(t, filepath.Join(t.TempDir(), "s.json"), clock, local, &keepLastTier{clock: clock, keepLast: 2})
+	refused := 0
+	for i := 0; i < 20; i++ {
+		if runBackup(t, srv, "local") == http.StatusTooManyRequests {
+			refused++
+		}
+		clock.Advance(10 * time.Minute)
+	}
+	if runs, _ := local.count(); runs != 3 || refused != 17 {
+		t.Fatalf("local flood over ~3 h: %d runs, %d refused — want 3 and 17", runs, refused)
+	}
+	clock.Advance(24 * time.Hour) // a day later the window has moved on
+	if code := runBackup(t, srv, "local"); code != http.StatusAccepted {
+		t.Fatalf("a run a day later must be allowed: %d", code)
+	}
+}
+
+// The schedule itself is never refused: a night at W+6h, a press at 10:00, the next night at W+2h (all
+// three inside one rolling 24 h) run; a second press that day is refused; nights keep running after it.
+func TestBackupCap_TheScheduledDayIsNeverRefused(t *testing.T) {
+	day := time.Date(2026, 6, 10, 0, 0, 0, 0, time.UTC)
+	clock := &testClock{t: day.Add(8*time.Hour + 29*time.Minute)} // W=02:30 → W+6h ≈ 08:29
+	local := &keepLastTier{clock: clock, keepLast: 3}
+	srv := capServer(t, filepath.Join(t.TempDir(), "s.json"), clock, local, &keepLastTier{clock: clock, keepLast: 2})
+	steps := []struct {
+		at   time.Time
+		want int
+		what string
+	}{
+		{day.Add(8*time.Hour + 29*time.Minute), http.StatusAccepted, "night 1 at W+6h"},
+		{day.Add(10 * time.Hour), http.StatusAccepted, "a press at 10:00"},
+		{day.Add(24*time.Hour + 4*time.Hour + 30*time.Minute), http.StatusAccepted, "night 2 at W+2h (20 h later)"},
+		{day.Add(24*time.Hour + 7*time.Hour), http.StatusTooManyRequests, "a second press inside the same 24 h"},
+		{day.Add(48*time.Hour + 4*time.Hour + 30*time.Minute), http.StatusAccepted, "night 3"},
+	}
+	for _, s := range steps {
+		clock.t = s.at
+		if got := runBackup(t, srv, "local"); got != s.want {
+			t.Fatalf("%s: got %d, want %d", s.what, got, s.want)
+		}
+	}
+}
+
+// A restart must not reset the cap: the history is read back from disk.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): make recentCapRuns ignore s.lastKnown → this fails with
+// "after a restart the 4th run in a day was allowed (got 202)". Restored.
+func TestBackupCap_SurvivesARestart(t *testing.T) {
+	clock := &testClock{t: testNow}
+	path := filepath.Join(t.TempDir(), "s.json")
+	local := &keepLastTier{clock: clock, keepLast: 3}
+	srv := capServer(t, path, clock, local, &keepLastTier{clock: clock, keepLast: 2})
+	for i := 0; i < 3; i++ {
+		if code := runBackup(t, srv, "local"); code != http.StatusAccepted {
+			t.Fatalf("run %d: %d", i+1, code)
+		}
+		clock.Advance(time.Minute)
+	}
+	srv2 := capServer(t, path, clock, local, &keepLastTier{clock: clock, keepLast: 2}) // the restart
+	if code := runBackup(t, srv2, "local"); code != http.StatusTooManyRequests {
+		t.Fatalf("after a restart the 4th run in a day was allowed (got %d)", code)
+	}
+}
+
+// Failed runs prune nothing and must stay retryable: they do not count.
+func TestBackupCap_FailuresDoNotCount(t *testing.T) {
+	clock := &testClock{t: testNow}
+	local := &keepLastTier{clock: clock, keepLast: 3, fail: true}
+	srv := capServer(t, filepath.Join(t.TempDir(), "s.json"), clock, local, &keepLastTier{clock: clock, keepLast: 2})
+	for i := 0; i < 6; i++ {
+		if code := runBackup(t, srv, "local"); code != http.StatusAccepted {
+			t.Fatalf("failed attempt %d was refused: %d", i+1, code)
+		}
+		clock.Advance(15 * time.Minute)
+	}
+	local.mu.Lock()
+	local.fail = false
+	local.mu.Unlock()
+	if code := runBackup(t, srv, "local"); code != http.StatusAccepted {
+		t.Fatalf("a success after six failures was refused: %d", code)
+	}
+}
+
+// A refusal raises the operator signal for 24 h, and the signal survives a restart.
+func TestBackupCap_RefusalRaisesTheOperatorSignal(t *testing.T) {
+	clock := &testClock{t: testNow}
+	path := filepath.Join(t.TempDir(), "s.json")
+	off := &keepLastTier{clock: clock, keepLast: 2}
+	srv := capServer(t, path, clock, &keepLastTier{clock: clock, keepLast: 3}, off)
+	if _, _, _, ok := srv.BackupCapRefusal(); ok {
+		t.Fatal("a fresh box must not raise the signal")
+	}
+	runBackup(t, srv, "felhom-pbs")
+	clock.Advance(time.Hour)
+	if code := runBackup(t, srv, "felhom-pbs"); code != http.StatusTooManyRequests {
+		t.Fatalf("second off-site run an hour later: %d, want 429", code)
+	}
+	target, vmid, at, ok := srv.BackupCapRefusal()
+	if !ok || target != "felhom-pbs" || vmid != 8200 || !at.Equal(clock.Now()) {
+		t.Fatalf("signal after a refusal: ok=%v target=%q vmid=%d at=%v", ok, target, vmid, at)
+	}
+	clock.Advance(23 * time.Hour)
+	srv2 := capServer(t, path, clock, &keepLastTier{clock: clock, keepLast: 3}, off) // a restart
+	if _, _, _, ok := srv2.BackupCapRefusal(); !ok {
+		t.Fatal("the signal did not survive a restart")
+	}
+	clock.Advance(2 * time.Hour)
+	if _, _, _, ok := srv2.BackupCapRefusal(); ok {
+		t.Fatal("the signal must clear 24 h after the last refusal")
+	}
+}
+
+func TestBackupCadenceCap(t *testing.T) {
+	cases := []struct {
+		cadence time.Duration
+		window  time.Duration
+		max     int
+	}{
+		{0, 24 * time.Hour, 3},
+		{24 * time.Hour, 24 * time.Hour, 3},
+		{12 * time.Hour, 24 * time.Hour, 3},
+		{7 * 24 * time.Hour, 6 * 24 * time.Hour, 1},
+		{36 * time.Hour, 18 * time.Hour, 1},
+		{30 * 24 * time.Hour, 29 * 24 * time.Hour, 1},
+	}
+	for _, c := range cases {
+		if w, m := backupCadenceCap(c.cadence); w != c.window || m != c.max {
+			t.Errorf("cadence %v: got %v/%d, want %v/%d", c.cadence, w, m, c.window, c.max)
+		}
+	}
+}

internal/localapi/escrow_ceremony.go

@@ -24,9 +24,16 @@ import (
 //     be structurally incapable of carrying R).
 //   - One claim, then the holder is zeroed. Unclaimed past the TTL → zeroed + phase
 //     unclaimed_void ("R unclaimed → ceremony void → a re-run supersedes").
-//   - Crash-safety is IN-MEMORY BY DESIGN: an agent restart loses R, which is SAFE (the blob is
+//   - Crash-safety is IN-MEMORY BY DESIGN: an agent restart loses R AND the package it opens,
-//     on the hub; a re-run supersedes it). Status reporting "none" after a restart is the honest
+//     which is SAFE (R-978: the package was never uploaded, so the hub still holds the previous one
-//     answer. No journal, deliberately — persistence is the one property R must never have.
+//     and the customer's previous code still opens it; a re-run starts over). Status reporting "none"
+//     after a restart is the honest answer. No journal, deliberately — persistence is the one
+//     property R must never have.
+//   - R-978: the NEW package reaches the hub only at the claim, and the code is handed out only after
+//     the hub took it. Until then the previous package stays current — so a ceremony that is started
+//     and abandoned (never claimed, superseded, TTL, restart) leaves the customer's written-down code
+//     working. The TTL and a new start both wait while a claim's upload is in flight, so a package is
+//     never stored without its code being handed out by that same claim.
 //   - Nothing from stdout is ever logged; stderr (log-clean, spike §2.1) is tail-captured for
 //     failure diagnostics only.
 
@@ -50,8 +57,16 @@ type EscrowCeremonyConfig struct {
 	// DRConfigured answers "is the DR tier applied on this box?" (pbsdr.Manager.DRConfigured,
 	// late-bound). nil → reported not-applied.
 	DRConfigured func() bool
+	// Upload sends a ceremony's package (CeremonyOutput.PendingUpload, the hub upload body) to this
+	// host's hub record — called by the claim, BEFORE the code is handed out (R-978). nil → a claim of
+	// a pending package is refused (503): a code must never be handed out for a package the hub lacks.
+	Upload func(ctx context.Context, pkg []byte) error
 }
 
+// escrowUploadTimeout bounds the claim-time package upload (one small PUT; the subprocess's whole
+// ceremony incl. its upload was measured at ~2.4 s).
+const escrowUploadTimeout = 30 * time.Second
+
 // ceremonyRunner executes the fixed-argv sudo self-invocation. stdout is SECRET-BEARING until
 // parsed (it carries R inside the JSON object); the caller must zero it. Tests inject canned
 // spike-shaped output; production is runCeremonySubprocess.
@@ -129,6 +144,9 @@ func (s *Server) tryStartEscrowCeremony(job *escrowCeremonyJob) bool {
 	if s.escrowJob != nil && s.escrowJob.Phase == escrowPhaseRunning {
 		return false
 	}
+	if s.escrowClaiming {
+		return false // R-978: a claim is uploading this job's package — its code must reach the claimer
+	}
 	s.wipeEscrowRLocked()
 	cp := *job
 	s.escrowJob = &cp
@@ -137,7 +155,7 @@ func (s *Server) tryStartEscrowCeremony(job *escrowCeremonyJob) bool {
 
 // finishEscrowCeremony records the terminal phase and (on success) arms the one-shot R holder.
 // R is stored as a byte slice so the claim/TTL paths can zero it in place.
-func (s *Server) finishEscrowCeremony(job *escrowCeremonyJob, r string) {
+func (s *Server) finishEscrowCeremony(job *escrowCeremonyJob, r string, pending []byte) {
 	s.escrowMu.Lock()
 	defer s.escrowMu.Unlock()
 	job.UpdatedAt = s.nowFn().UTC().Format(time.RFC3339)
@@ -145,6 +163,7 @@ func (s *Server) finishEscrowCeremony(job *escrowCeremonyJob, r string) {
 	s.escrowJob = &cp
 	if job.Phase == escrowPhaseDone && r != "" {
 		s.escrowR = []byte(r)
+		s.escrowPending = append([]byte(nil), pending...)
 		s.escrowRClaimed = false
 		s.escrowRExpiry = s.nowFn().Add(escrowClaimTTL)
 		// Active TTL belt: zero the holder even if nobody ever polls again. The lazy check in
@@ -164,6 +183,7 @@ func (s *Server) finishEscrowCeremony(job *escrowCeremonyJob, r string) {
 func (s *Server) wipeEscrowRLocked() {
 	zeroBytes(s.escrowR)
 	s.escrowR = nil
+	s.escrowPending = nil // R-978: the package goes with its code — never uploaded once the code is gone
 }
 
 // expireEscrowRLocked applies the TTL outcome: an armed, unclaimed holder is zeroed and the job
@@ -172,6 +192,9 @@ func (s *Server) expireEscrowRLocked() {
 	if s.escrowJob == nil || s.escrowJob.Phase != escrowPhaseDone || s.escrowRClaimed {
 		return
 	}
+	if s.escrowClaiming {
+		return // R-978: the claim is uploading the package; voiding now would store a package nobody gets the code to
+	}
 	if len(s.escrowR) == 0 {
 		return
 	}
@@ -217,7 +240,7 @@ func (s *Server) handleEscrowCeremonyStart(w http.ResponseWriter, r *http.Reques
 		UpdatedAt: s.nowFn().UTC().Format(time.RFC3339),
 	}
 	if !s.tryStartEscrowCeremony(job) {
-		writeErr(w, http.StatusConflict, "an escrow ceremony is already running")
+		writeErr(w, http.StatusConflict, "an escrow ceremony is already running, or its code is being claimed")
 		return
 	}
 	s.logger.Info("local-api: escrow ceremony started (controller-driven)", "vmid", vmid, "job_id", job.JobID)
@@ -247,7 +270,7 @@ func (s *Server) handleEscrowCeremonyStart(w http.ResponseWriter, r *http.Reques
 				detail += " | stderr: " + tail
 			}
 			job.Phase, job.Detail = escrowPhaseFailed, detail
-			s.finishEscrowCeremony(job, "")
+			s.finishEscrowCeremony(job, "", nil)
 			s.logger.Warn("local-api: escrow ceremony failed", "job_id", job.JobID,
 				"exit", exit, "duration_ms", time.Since(start).Milliseconds())
 			return
@@ -255,24 +278,34 @@ func (s *Server) handleEscrowCeremonyStart(w http.ResponseWriter, r *http.Reques
 		var out escrow.CeremonyOutput
 		if err := json.Unmarshal(stdout, &out); err != nil {
 			job.Phase, job.Detail = escrowPhaseFailed, "ceremony output is not the expected JSON object"
-			s.finishEscrowCeremony(job, "")
+			s.finishEscrowCeremony(job, "", nil)
 			s.logger.Warn("local-api: escrow ceremony output unparseable (never logged)", "job_id", job.JobID)
 			return
 		}
 		if out.Version != escrow.CeremonyOutputVersion || out.RecoveryCode == "" {
 			out.RecoveryCode = ""
 			job.Phase, job.Detail = escrowPhaseFailed, fmt.Sprintf("unexpected ceremony output (version %d)", out.Version)
-			s.finishEscrowCeremony(job, "")
+			s.finishEscrowCeremony(job, "", nil)
+			return
+		}
+		// R-978: the package must come back un-uploaded. An output that uploaded by itself (or carries no
+		// package) would mean the previous package is already gone, or the new one can never be stored.
+		if out.Uploaded || len(out.PendingUpload) == 0 {
+			out.RecoveryCode = ""
+			job.Phase, job.Detail = escrowPhaseFailed, "ceremony output carries no pending package"
+			s.finishEscrowCeremony(job, "", nil)
+			s.logger.Warn("local-api: escrow ceremony output has no pending package — failed (R-978)", "job_id", job.JobID,
+				"uploaded", out.Uploaded)
 			return
 		}
 		job.Phase = escrowPhaseDone
 		job.KeyFingerprint = out.KeyFingerprint
 		job.EntropyBits = out.EntropyBits
 		job.ResticPwSealed = out.ResticPwSealed
-		job.Uploaded = out.Uploaded
+		job.Uploaded = false // R-978: uploaded by the claim
-		s.finishEscrowCeremony(job, out.RecoveryCode)
+		s.finishEscrowCeremony(job, out.RecoveryCode, out.PendingUpload)
 		out.RecoveryCode = "" // drop the parsed reference promptly (GC caveat: best-effort)
-		s.logger.Info("local-api: escrow ceremony complete — R claimable (in-memory, one-shot)",
+		s.logger.Info("local-api: escrow ceremony complete — R claimable (in-memory, one-shot); package uploads at the claim",
 			"job_id", job.JobID, "restic_pw_sealed", job.ResticPwSealed, "uploaded", job.Uploaded,
 			"claim_ttl_s", int(escrowClaimTTL.Seconds()), "duration_ms", time.Since(start).Milliseconds())
 	}()
@@ -339,11 +372,60 @@ func (s *Server) handleEscrowCeremonyClaim(w http.ResponseWriter, r *http.Reques
 		s.escrowMu.Unlock()
 		writeErr(w, http.StatusConflict, "ceremony failed — nothing to claim")
 		return
+	case s.escrowClaiming:
+		s.escrowMu.Unlock()
+		writeErr(w, http.StatusConflict, "the recovery code is being claimed — try again in a moment")
+		return
 	case s.escrowRClaimed || len(s.escrowR) == 0 || s.escrowJob.Phase == escrowPhaseVoid:
 		s.escrowMu.Unlock()
 		writeErr(w, http.StatusGone, "the recovery code is no longer available (already claimed or expired) — run a new ceremony; the new code supersedes")
 		return
 	}
+	// R-978: store the NEW package on the hub first; only then hand out the code that opens it. Until this
+	// upload succeeds the hub keeps the previous package, and the customer's previous code keeps working.
+	if len(s.escrowPending) > 0 {
+		if s.escrowCeremony.Upload == nil {
+			s.escrowMu.Unlock()
+			writeErr(w, http.StatusServiceUnavailable, "the recovery package cannot be stored on this host (no upload configured) — the previous code still works")
+			return
+		}
+		pkg := s.escrowPending
+		jobID := s.escrowJob.JobID
+		s.escrowClaiming = true
+		s.escrowMu.Unlock()
+
+		start := time.Now()
+		base := s.baseCtx
+		if base == nil {
+			base = context.Background()
+		}
+		uctx, cancel := context.WithTimeout(base, escrowUploadTimeout)
+		err := s.escrowCeremony.Upload(uctx, pkg)
+		cancel()
+
+		s.escrowMu.Lock()
+		s.escrowClaiming = false
+		if err != nil {
+			s.escrowMu.Unlock()
+			s.logger.Warn("local-api: escrow claim — package upload to the hub failed; code kept claimable, previous package stays current (R-978)",
+				"vmid", vmid, "job_id", jobID, "err", err, "duration_ms", time.Since(start).Milliseconds())
+			writeErr(w, http.StatusBadGateway, "the new recovery package could not be stored on the hub — the previous code still works; try again")
+			return
+		}
+		s.escrowPending = nil
+		if s.escrowJob != nil && s.escrowJob.JobID == jobID {
+			s.escrowJob.Uploaded = true
+			s.escrowJob.UpdatedAt = s.nowFn().UTC().Format(time.RFC3339)
+		}
+		s.logger.Info("local-api: escrow claim — package uploaded to the hub; handing out the code (R-978)",
+			"vmid", vmid, "job_id", jobID, "duration_ms", time.Since(start).Milliseconds())
+		if len(s.escrowR) == 0 { // cannot happen (TTL and supersede wait for the claim) — never hand out nothing
+			s.escrowMu.Unlock()
+			s.logger.Error("local-api: escrow claim — code vanished during the upload (invariant broken)", "job_id", jobID)
+			writeErr(w, http.StatusInternalServerError, "the recovery code was lost during the claim — run a new ceremony")
+			return
+		}
+	}
 	recovery := string(s.escrowR)
 	s.wipeEscrowRLocked()
 	s.escrowRClaimed = true

internal/localapi/escrow_ceremony_test.go

@@ -2,12 +2,14 @@ package localapi
 
 import (
 	"context"
+	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"net/http"
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -19,25 +21,91 @@ import (
 
 const testR = "canary-alpha-bravo-charlie-delta-echo-foxtrot-golf-hotel-india"
 
-// cannedCeremonyJSON is shaped exactly like the agent's --output=json object (spike §2.6 values).
+// cannedCeremonyJSON is shaped exactly like the agent's --output=json object (spike §2.6 values;
+// version 2, R-978: the package comes back un-uploaded in pending_upload). The package is a MODEL of
+// the R-wrapped blob: it "opens" with exactly the code it was sealed under (fakeEscrowHub.opensWith).
 func cannedCeremonyJSON(r string) string {
-	return fmt.Sprintf(`{"version":1,"recovery_code":%q,"key_fingerprint":"f2:87:68:2a:88:50:16:01","entropy_bits":129,"blob_bytes":383,"identity_blob_bytes":450,"restic_pw_sealed":true,"uploaded":true}`, r)
+	return fmt.Sprintf(`{"version":2,"recovery_code":%q,"key_fingerprint":"f2:87:68:2a:88:50:16:01","entropy_bits":129,"blob_bytes":383,"identity_blob_bytes":450,"restic_pw_sealed":true,"uploaded":false,"pending_upload":%s}`,
+		r, sealedPackage(r))
+}
+
+// sealedPackage is the model hub upload body for a ceremony that minted code r.
+func sealedPackage(r string) string {
+	return fmt.Sprintf(`{"blob_b64":%q,"key_fingerprint":"f2:87:68:2a:88:50:16:01","posture":"zero_knowledge","created_at":"2026-06-10T12:00:00Z"}`,
+		base64.StdEncoding.EncodeToString([]byte("sealed-under:"+r)))
+}
+
+// fakeEscrowHub is the hub's current escrow package for this host (the PUT target). Upload replaces it,
+// as the hub does when the sealed secrets are unchanged (the R-978 case: no retained copy).
+type fakeEscrowHub struct {
+	mu      sync.Mutex
+	current []byte
+	uploads int
+	err     error         // non-nil → the upload fails, nothing stored
+	gate    chan struct{} // non-nil → the upload blocks until closed
+	entered chan struct{} // non-nil → closed when an upload starts
+}
+
+func (f *fakeEscrowHub) upload(ctx context.Context, pkg []byte) error {
+	if f.entered != nil {
+		close(f.entered)
+		f.entered = nil
+	}
+	if f.gate != nil {
+		<-f.gate
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if f.err != nil {
+		return f.err
+	}
+	f.current = append([]byte(nil), pkg...)
+	f.uploads++
+	return nil
+}
+
+// opensWith reports whether the hub's CURRENT package opens with code r — the customer's question.
+func (f *fakeEscrowHub) opensWith(r string) bool {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	var pkg struct {
+		BlobB64 string `json:"blob_b64"`
+	}
+	if json.Unmarshal(f.current, &pkg) != nil {
+		return false
+	}
+	raw, err := base64.StdEncoding.DecodeString(pkg.BlobB64)
+	return err == nil && string(raw) == "sealed-under:"+r
+}
+
+func (f *fakeEscrowHub) count() int {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.uploads
 }
 
 // newEscrowTestServer builds a server with the ceremony configured and every seam faked.
 func newEscrowTestServer(t *testing.T, run ceremonyRunner) *Server {
+	t.Helper()
+	srv, _ := newEscrowTestServerHub(t, run, &fakeEscrowHub{})
+	return srv
+}
+
+// newEscrowTestServerHub is newEscrowTestServer with the hub's escrow record exposed (R-978).
+func newEscrowTestServerHub(t *testing.T, run ceremonyRunner, hub *fakeEscrowHub) (*Server, *fakeEscrowHub) {
 	t.Helper()
 	srv := newTestServerS(t, &fakeGuests{}, &fakeBackups{}, &fakeStore{}, nil)
 	srv.escrowCeremony = &EscrowCeremonyConfig{
 		PBSStorageID:  "felhom-pbs",
 		HubConfigured: true,
 		DRConfigured:  func() bool { return true },
+		Upload:        hub.upload,
 	}
 	srv.ceremonyRun = run
 	srv.escrowSudoCheck = func(context.Context) error { return nil }
 	srv.escrowLookPath = func(string) (string, error) { return "/usr/bin/age", nil }
 	srv.statFile = func(string) bool { return true }
-	return srv
+	return srv, hub
 }
 
 // okRunner returns the canned success output and counts invocations.
@@ -160,8 +228,8 @@ func TestEscrowCeremony_OneShotClaim(t *testing.T) {
 	if !strings.Contains(body, `"phase":"done"`) || !strings.Contains(body, `"claimable":true`) {
 		t.Fatalf("status not done/claimable: %s", body)
 	}
-	if !strings.Contains(body, `"restic_pw_sealed":true`) || !strings.Contains(body, `"uploaded":true`) {
+	if !strings.Contains(body, `"restic_pw_sealed":true`) || !strings.Contains(body, `"uploaded":false`) {
-		t.Fatalf("summary fields missing: %s", body)
+		t.Fatalf("summary fields missing (R-978: not uploaded before the claim): %s", body)
 	}
 	assertNoR(t, "status payload", body)
 
@@ -199,7 +267,8 @@ func TestEscrowCeremony_OneShotClaim(t *testing.T) {
 
 	// Post-claim status: claimed, not claimable, still phase done.
 	st2 := do(t, h, "GET", "/escrow/ceremony/status", "A", "")
-	if !strings.Contains(st2.Body.String(), `"claimed":true`) || strings.Contains(st2.Body.String(), `"claimable":true`) {
+	if !strings.Contains(st2.Body.String(), `"claimed":true`) || strings.Contains(st2.Body.String(), `"claimable":true`) ||
+		!strings.Contains(st2.Body.String(), `"uploaded":true`) {
 		t.Fatalf("post-claim status wrong: %s", st2.Body.String())
 	}
 }
@@ -344,7 +413,7 @@ func TestEscrowCeremony_FailurePaths(t *testing.T) {
 	})
 	t.Run("wrong version", func(t *testing.T) {
 		srv := newEscrowTestServer(t, func(context.Context) ([]byte, []byte, int, error) {
-			return []byte(`{"version":2,"recovery_code":"` + testR + `"}`), nil, 0, nil
+			return []byte(`{"version":1,"recovery_code":"` + testR + `","uploaded":true}`), nil, 0, nil
 		})
 		h := srv.Handler()
 		startAndWait(t, srv, h)

internal/localapi/escrow_r978_test.go

@@ -0,0 +1,219 @@
+package localapi
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+)
+
+// R-978 — starting the escrow ceremony again must not make the customer's written-down recovery code
+// useless. The hub overwrites the current package when the sealed secrets are unchanged, so the NEW
+// package may reach the hub only when its code is claimed. These tests ask the customer's question of
+// the hub's current package — „does my code still open it?" — not whether some function ran.
+
+const (
+	firstR  = "first-ceremony-code-alpha"
+	secondR = "second-ceremony-code-bravo"
+)
+
+// sequenceRunner mints the given codes in order, one per ceremony.
+func sequenceRunner(codes ...string) ceremonyRunner {
+	i := 0
+	return func(context.Context) ([]byte, []byte, int, error) {
+		r := codes[i]
+		if i < len(codes)-1 {
+			i++
+		}
+		return []byte(cannedCeremonyJSON(r)), nil, 0, nil
+	}
+}
+
+func claimCode(t *testing.T, h http.Handler) (int, string) {
+	t.Helper()
+	w := do(t, h, "POST", "/escrow/ceremony/claim", "A", "")
+	return w.Code, w.Body.String()
+}
+
+// THE R-978 CASE. Ceremony 1 is claimed (the customer writes code 1 down). Ceremony 2 is started and
+// abandoned — three ways: its code expires unclaimed, a third start supersedes it, the agent restarts.
+// In every case the hub's current package must still open with code 1, and nothing was uploaded for 2.
+//
+// COMPANION RED-PROOF (observed 2026-10-18): restore the pre-R-978 shape — the package uploaded when the
+// ceremony finishes (s.escrowCeremony.Upload(base, out.PendingUpload) in the job goroutine, as the root
+// subprocess did with `--upload`) and nothing left for the claim (finishEscrowCeremony(job, R, nil)) →
+// all three subtests FAIL with "the abandoned second ceremony replaced the package: code 1 no longer
+// opens it (uploads=2)" (uploads=3 for the superseding start). Restored.
+func TestEscrowCeremony_R978_AbandonedSecondCeremonyKeepsFirstCode(t *testing.T) {
+	setup := func(t *testing.T) (*Server, *fakeEscrowHub, http.Handler, *time.Time) {
+		srv, hub := newEscrowTestServerHub(t, sequenceRunner(firstR, secondR, "third-code-charlie"), &fakeEscrowHub{})
+		cur := testNow
+		srv.now = func() time.Time { return cur }
+		h := srv.Handler()
+		startAndWait(t, srv, h)
+		if code, body := claimCode(t, h); code != http.StatusOK || !strings.Contains(body, firstR) {
+			t.Fatalf("claim of ceremony 1: %d %s", code, body)
+		}
+		if !hub.opensWith(firstR) || hub.count() != 1 {
+			t.Fatalf("after the first claim the hub must hold code 1's package (uploads=%d)", hub.count())
+		}
+		startAndWait(t, srv, h) // ceremony 2 — about to be abandoned
+		return srv, hub, h, &cur
+	}
+	check := func(t *testing.T, hub *fakeEscrowHub) {
+		t.Helper()
+		if !hub.opensWith(firstR) {
+			t.Fatalf("the abandoned second ceremony replaced the package: code 1 no longer opens it (uploads=%d)", hub.count())
+		}
+		if hub.opensWith(secondR) || hub.count() != 1 {
+			t.Fatalf("a package for the never-claimed code 2 reached the hub (uploads=%d)", hub.count())
+		}
+	}
+
+	t.Run("code 2 expires unclaimed", func(t *testing.T) {
+		_, hub, h, cur := setup(t)
+		*cur = cur.Add(escrowClaimTTL + time.Minute)
+		if code, _ := claimCode(t, h); code != http.StatusGone {
+			t.Fatalf("claim after TTL: got %d, want 410", code)
+		}
+		check(t, hub)
+	})
+	t.Run("a third start supersedes code 2", func(t *testing.T) {
+		srv, hub, h, _ := setup(t)
+		startAndWait(t, srv, h) // ceremony 3, also left unclaimed
+		check(t, hub)
+	})
+	t.Run("the agent restarts before code 2 is claimed", func(t *testing.T) {
+		_, hub, _, _ := setup(t)
+		// A restart is a new process with an empty slot; only the hub's record survives it.
+		srv2, _ := newEscrowTestServerHub(t, sequenceRunner("unused"), hub)
+		if code, _ := claimCode(t, srv2.Handler()); code != http.StatusNotFound {
+			t.Fatalf("claim after a restart: got %d, want 404", code)
+		}
+		check(t, hub)
+	})
+}
+
+// The positive control: CLAIMING code 2 is what replaces the package — the code handed out opens the
+// package the hub now holds, and the claim answers only after the hub took it.
+func TestEscrowCeremony_R978_ClaimStoresThePackageBeforeHandingOutTheCode(t *testing.T) {
+	srv, hub := newEscrowTestServerHub(t, sequenceRunner(firstR, secondR), &fakeEscrowHub{})
+	h := srv.Handler()
+	startAndWait(t, srv, h)
+	if hub.count() != 0 {
+		t.Fatalf("the package was uploaded before any claim (uploads=%d)", hub.count())
+	}
+	claimCode(t, h)
+	startAndWait(t, srv, h)
+	code, body := claimCode(t, h)
+	if code != http.StatusOK || !strings.Contains(body, secondR) {
+		t.Fatalf("claim of ceremony 2: %d %s", code, body)
+	}
+	if !hub.opensWith(secondR) || hub.count() != 2 {
+		t.Fatalf("after claiming code 2 the hub must hold its package (uploads=%d, opens with 2: %v)", hub.count(), hub.opensWith(secondR))
+	}
+}
+
+// A failed upload hands out NO code, keeps the code claimable and the previous package current; a retry
+// that reaches the hub succeeds.
+func TestEscrowCeremony_R978_UploadFailureKeepsTheCodeClaimable(t *testing.T) {
+	hub := &fakeEscrowHub{err: errors.New("hub returned HTTP 502")}
+	srv, _ := newEscrowTestServerHub(t, sequenceRunner(firstR), hub)
+	h := srv.Handler()
+	startAndWait(t, srv, h)
+
+	code, body := claimCode(t, h)
+	if code != http.StatusBadGateway {
+		t.Fatalf("claim with the hub failing: got %d, want 502 (%s)", code, body)
+	}
+	assertNoRIn(t, "failed-upload claim payload", body, firstR)
+	if st := do(t, h, "GET", "/escrow/ceremony/status", "A", "").Body.String(); !strings.Contains(st, `"claimable":true`) || !strings.Contains(st, `"uploaded":false`) {
+		t.Fatalf("after a failed upload the code must stay claimable and un-uploaded: %s", st)
+	}
+
+	hub.mu.Lock()
+	hub.err = nil
+	hub.mu.Unlock()
+	code, body = claimCode(t, h)
+	if code != http.StatusOK || !strings.Contains(body, firstR) || !hub.opensWith(firstR) {
+		t.Fatalf("retry after the hub recovered: %d %s (opens: %v)", code, body, hub.opensWith(firstR))
+	}
+}
+
+// While the claim's upload is in flight, neither the TTL nor a new start may take the code away — else
+// the hub would hold a package whose code nobody was given (the R-978 failure, reached the other way).
+func TestEscrowCeremony_R978_TTLAndRestartWaitForAnInFlightClaim(t *testing.T) {
+	hub := &fakeEscrowHub{gate: make(chan struct{}), entered: make(chan struct{})}
+	entered := hub.entered
+	srv, _ := newEscrowTestServerHub(t, sequenceRunner(firstR, secondR), hub)
+	var mu = make(chan struct{}, 1)
+	mu <- struct{}{}
+	cur := testNow
+	srv.now = func() time.Time { <-mu; defer func() { mu <- struct{}{} }(); return cur }
+	h := srv.Handler()
+	startAndWait(t, srv, h)
+
+	type res struct {
+		code int
+		body string
+	}
+	done := make(chan res, 1)
+	go func() { c, b := claimCode(t, h); done <- res{c, b} }()
+	<-entered
+
+	<-mu
+	cur = cur.Add(escrowClaimTTL + time.Minute) // the TTL passes mid-upload
+	mu <- struct{}{}
+	if st := do(t, h, "GET", "/escrow/ceremony/status", "A", "").Body.String(); strings.Contains(st, `"phase":"unclaimed_void"`) {
+		t.Fatalf("the TTL voided the code while its package was uploading: %s", st)
+	}
+	if w := do(t, h, "POST", "/escrow/ceremony", "A", ""); w.Code != http.StatusConflict {
+		t.Fatalf("a new start during the claim's upload: got %d, want 409", w.Code)
+	}
+	if code, _ := claimCode(t, h); code != http.StatusConflict {
+		t.Fatalf("a second claim during the upload: got %d, want 409", code)
+	}
+	close(hub.gate)
+	r := <-done
+	if r.code != http.StatusOK || !strings.Contains(r.body, firstR) || !hub.opensWith(firstR) {
+		t.Fatalf("the in-flight claim must deliver the code whose package it stored: %d %s", r.code, r.body)
+	}
+}
+
+// No upload configured → the claim refuses rather than hand out a code for a package the hub lacks.
+func TestEscrowCeremony_R978_NoUploaderRefusesTheClaim(t *testing.T) {
+	srv, _ := newEscrowTestServerHub(t, sequenceRunner(firstR), &fakeEscrowHub{})
+	srv.escrowCeremony.Upload = nil
+	h := srv.Handler()
+	startAndWait(t, srv, h)
+	code, body := claimCode(t, h)
+	if code != http.StatusServiceUnavailable {
+		t.Fatalf("claim with no uploader: got %d, want 503 (%s)", code, body)
+	}
+	assertNoRIn(t, "no-uploader claim payload", body, firstR)
+}
+
+// An output that uploaded by itself (the pre-R-978 shape) or carries no package fails the ceremony.
+func TestEscrowCeremony_R978_OutputWithoutPendingPackageFails(t *testing.T) {
+	srv, _ := newEscrowTestServerHub(t, func(context.Context) ([]byte, []byte, int, error) {
+		return []byte(`{"version":2,"recovery_code":"` + firstR + `","uploaded":true}`), nil, 0, nil
+	}, &fakeEscrowHub{})
+	h := srv.Handler()
+	startAndWait(t, srv, h)
+	st := do(t, h, "GET", "/escrow/ceremony/status", "A", "").Body.String()
+	if !strings.Contains(st, `"phase":"failed"`) || !strings.Contains(st, "no pending package") {
+		t.Fatalf("want failed: %s", st)
+	}
+	if code, _ := claimCode(t, h); code != http.StatusConflict {
+		t.Fatalf("claim on a failed ceremony: got %d, want 409", code)
+	}
+}
+
+func assertNoRIn(t *testing.T, where, body, r string) {
+	t.Helper()
+	if strings.Contains(body, r) {
+		t.Fatalf("%s carries the recovery code", where)
+	}
+}

internal/localapi/server.go

@@ -90,6 +90,10 @@ type BackupStore interface {
 type LastKnownBackupStore interface {
 	RecordBackupSuccess(target string, b hub.Backup) error
 	LastKnownSuccess(target string, vmid int) (time.Time, bool)
+	// R-973: the cadence cap's restart-proof history and its refusals.
+	RecentBackupSuccesses(target string, vmid int, since time.Time) []time.Time
+	RecordBackupRefusal(target string, vmid int, at time.Time) error
+	LastBackupRefusal() (target string, vmid int, at time.Time, ok bool)
 }
 
 // StorageView yields the host's observed storage targets (for mapping a mount's storage id →
@@ -302,6 +306,11 @@ type Server struct {
 	logger             *slog.Logger
 	now                func() time.Time
 
+	// R-973 cadence cap: this process's successful runs per tier+guest, and the newest refusal.
+	capMu      sync.Mutex
+	capRuns    map[backupJobKey][]time.Time
+	capRefusal capRefusal
+
 	disks           DiskOps           // slice 8C (optional)
 	diskGate        StorageGate       // slice 8C (optional)
 	guestList       GuestLister       // slice 8C (optional)
@@ -426,6 +435,8 @@ type Server struct {
 	escrowMu       sync.Mutex
 	escrowJob      *escrowCeremonyJob
 	escrowR        []byte
+	escrowPending  []byte // R-978: the ceremony's hub upload body, sent by the claim before R is handed out
+	escrowClaiming bool   // R-978: a claim is uploading escrowPending (TTL and a new start wait)
 	escrowRClaimed bool
 	escrowRExpiry  time.Time
 	escrowDone     <-chan struct{} // closes when the detached job finishes (tests wait on it)
@@ -878,6 +889,12 @@ func (s *Server) handleBackup(w http.ResponseWriter, r *http.Request, vmid int)
 			"a backup is already in flight on target "+busyTarget+" (job "+busyJob+") — only one backup runs at a time per guest")
 		return
 	}
+	// R-973: the cadence cap (backup_cap.go). Checked under jobsMu, after the in-flight answers, so two
+	// requests cannot both pass it (only one run per guest is ever in flight).
+	if s.capRefuses(w, key, tier) {
+		s.jobsMu.Unlock()
+		return
+	}
 	// Job ids must be unique PER TIER, and by construction rather than by clock luck: two tiers
 	// started inside the same nanosecond (the weekly both-due night, or any injected clock) would
 	// otherwise collide and hand the second caller the first tier's id. The PRIMARY keeps the
@@ -941,6 +958,9 @@ func (s *Server) handleBackup(w http.ResponseWriter, r *http.Request, vmid int)
 			s.logger.Info("local-api: backup job complete", "vmid", vmid, "target", tier.TargetID, "job", jobID, "archive", b.Archive)
 		}
 		s.store.RecordBackup(b)
+		if b.Success {
+			s.recordCapRun(key, b) // R-973
+		}
 		if b.Success && s.lastKnown != nil {
 			if err := s.lastKnown.RecordBackupSuccess(tier.TargetID, b); err != nil {
 				// Not fatal: the backup exists. Only the fallback after a restart loses this copy.

internal/signedjobs/decommission.go

@@ -39,9 +39,21 @@ type decommissionParams struct {
 // none is wanted, because a drive can legitimately be decommissioned while physically ABSENT.
 type DecommissionExecutor struct {
 	intent IntentDecommissioner
+	binds  GuestBindPruner // optional (v0.159.1): drop the drive from every guest's bind record, as the self-serve path does
 	logger *slog.Logger
 }
 
+// GuestBindPruner is the per-guest enrolled-bind record (*localapi.GuestBindStore). The self-serve decommission prunes
+// the drive from it; the signed one did not, so a forgotten drive (R-1124) left a record the 20-second guest-bind
+// re-assert warned about forever ("skipping non-enrolled drive"), measured on demo-hp 2026-10-17.
+type GuestBindPruner interface {
+	Guests() map[int][]string
+	Remove(vmid int, durableID string) error
+}
+
+// SetGuestBinds wires the bind record (nil-safe: no record, nothing pruned).
+func (d *DecommissionExecutor) SetGuestBinds(b GuestBindPruner) { d.binds = b }
+
 // NewDecommissionExecutor wires the executor to the agent's intent store.
 func NewDecommissionExecutor(intent IntentDecommissioner, logger *slog.Logger) *DecommissionExecutor {
 	if logger == nil {
@@ -75,6 +87,20 @@ func (d *DecommissionExecutor) Execute(_ context.Context, op string, params json
 	if err := d.intent.SetDecommissioned(p.DurableID); err != nil {
 		return fmt.Errorf("decommission: record intent for %q: %w", p.DurableID, err)
 	}
+	if d.binds != nil {
+		for vmid, ids := range d.binds.Guests() {
+			for _, id := range ids {
+				if id != p.DurableID {
+					continue
+				}
+				if err := d.binds.Remove(vmid, id); err != nil {
+					d.logger.Warn("decommission: guest-bind record not pruned", "vmid", vmid, "durable_id", id, "err", err)
+				} else {
+					d.logger.Info("decommission: drive removed from the guest's bind record", "vmid", vmid, "durable_id", id)
+				}
+			}
+		}
+	}
 	d.logger.Warn("decommission: operator-signed decommission complete (drive will never auto-mount)", "durable_id", p.DurableID)
 	return nil
 }

internal/signedjobs/decommission_test.go

@@ -89,3 +89,37 @@ func TestExecutorChain_DispatchesToOwner(t *testing.T) {
 		t.Errorf("want ErrNoExecutor for an unowned op, got %v", err)
 	}
 }
+
+type fakeBinds struct{ m map[int][]string }
+
+func (f *fakeBinds) Guests() map[int][]string { return f.m }
+func (f *fakeBinds) Remove(vmid int, id string) error {
+	var kept []string
+	for _, x := range f.m[vmid] {
+		if x != id {
+			kept = append(kept, x)
+		}
+	}
+	f.m[vmid] = kept
+	return nil
+}
+
+// v0.159.1: the signed decommission drops the drive from EVERY guest's bind record, as the self-serve path does — and
+// only that drive. RED-PROOF (felhom.eu audits/agent-allowlist-rotation-2026-10-17/P1/redproof-bindprune.txt): with
+// the prune loop removed, the record still holds the drive and this test fails.
+func TestDecommissionExecutor_PrunesTheGuestBindRecord(t *testing.T) {
+	fd := &fakeDecommissioner{}
+	fb := &fakeBinds{m: map[int][]string{9201: {"uuid:keep", "uuid:gone"}, 9202: {"uuid:gone"}}}
+	ex := NewDecommissionExecutor(fd, nil)
+	ex.SetGuestBinds(fb)
+	params, _ := json.Marshal(decommissionParams{DurableID: "uuid:gone"})
+	if err := ex.Execute(context.Background(), opDecommission, params); err != nil {
+		t.Fatal(err)
+	}
+	if got := fb.m[9201]; len(got) != 1 || got[0] != "uuid:keep" {
+		t.Fatalf("9201 record = %v, want [uuid:keep]", got)
+	}
+	if got := fb.m[9202]; len(got) != 0 {
+		t.Fatalf("9202 record = %v, want empty", got)
+	}
+}

internal/storage/forget_retired.go

@@ -0,0 +1,85 @@
+package storage
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// ForgetRetiredDrives (R-1124) removes the host mount unit of every DECOMMISSIONED drive that is physically gone, so a
+// later host boot no longer waits for it and the agent stops listing it (RegistryKnownTargets and
+// ReassertEnrolledMounts both enumerate the units — no unit, no entry, no „absent by UUID" line every 20 s).
+//
+// Only a drive that is BOTH decommissioned (the operator's or the customer's permanent removal) AND absent (its
+// /dev/disk/by-uuid link does not resolve) AND not mounted is forgotten. A decommissioned drive that is still plugged in
+// keeps its unit and its raw host mount — that is what makes the one-click re-enroll (H3) work, and decommission stays a
+// logical retire (`03` §storage). The intent record stays `decommissioned`: a forgotten drive that comes back is a new
+// candidate, never auto-mounted. The root half (`felhom-priv-apply unit-forget`) re-checks every one of these conditions
+// and refuses otherwise. Called at agent start, on the 20-second storage tick, and right after a decommission. Best
+// effort per unit; returns how many units it removed. Pinned by TestForgetRetiredDrives_*.
+func (h *SudoHostOps) ForgetRetiredDrives(ctx context.Context, intent IntentReader) int {
+	if intent == nil {
+		return 0 // no intent store → we cannot know a drive was retired; forget nothing
+	}
+	entries, err := os.ReadDir(h.unitDir)
+	if err != nil {
+		h.logger.Warn("storage: forget retired drives — cannot read unit dir", "dir", h.unitDir, "err", err)
+		return 0
+	}
+	mounted := map[string]bool{}
+	if ms, merr := h.host.Mounts(); merr == nil {
+		for _, m := range ms {
+			mounted[m.MountPoint] = true
+		}
+	} else {
+		h.logger.Warn("storage: forget retired drives — mount table unreadable; forgetting nothing", "err", merr)
+		return 0
+	}
+	removed := 0
+	for _, e := range entries {
+		name := e.Name()
+		if !strings.HasSuffix(name, ".mount") {
+			continue
+		}
+		data, rerr := os.ReadFile(filepath.Join(h.unitDir, name))
+		if rerr != nil {
+			continue
+		}
+		spec, ok := parseFelhomMountUnit(string(data))
+		if !ok {
+			continue // not one of our by-UUID drive units (network units are refused by the parser)
+		}
+		durableID := "uuid:" + spec.UUID
+		if intent.Get(durableID) != IntentDecommissioned {
+			continue
+		}
+		if mounted[spec.Where] {
+			h.logger.Debug("storage: decommissioned drive still mounted — keeping its unit", "where", spec.Where)
+			continue
+		}
+		if _, derr := ResolveStorageDevice(durableID); derr == nil {
+			h.logger.Debug("storage: decommissioned drive is present — keeping its unit (re-enroll stays one click)", "where", spec.Where)
+			continue
+		}
+		if h.unitEnabled(name) {
+			if err := h.run(ctx, h.bins.Systemctl, "disable", "--", name); err != nil {
+				h.logger.Warn("storage: forget retired drive — disable failed", "where", spec.Where, "unit", name, "err", err)
+				continue
+			}
+		}
+		if err := h.run(ctx, h.bins.PrivApply, "unit-forget", name); err != nil {
+			h.logger.Warn("storage: forget retired drive — unit removal refused or failed", "where", spec.Where, "unit", name, "err", err)
+			continue
+		}
+		removed++
+		h.logger.Info("storage: forgot a decommissioned drive that is gone — its mount unit is removed",
+			"name", spec.Name, "where", spec.Where, "uuid", spec.UUID, "unit", name)
+	}
+	if removed > 0 {
+		if err := h.run(ctx, h.bins.Systemctl, "daemon-reload"); err != nil {
+			h.logger.Warn("storage: daemon-reload after forgetting retired drives failed", "err", err)
+		}
+	}
+	return removed
+}

internal/storage/forget_retired_test.go

@@ -0,0 +1,155 @@
+package storage
+
+import (
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"reflect"
+	"testing"
+)
+
+// forgetRunner records every privileged call and plays the root half of `felhom-priv-apply unit-forget`: the unit file
+// disappears from the unit dir, so the CONSEQUENCE (the drive is no longer listed) is observable, not just the calls.
+type forgetRunner struct {
+	unitDir string
+	calls   [][]string
+}
+
+func (r *forgetRunner) RunStdin(ctx context.Context, _ io.Reader, name string, args ...string) ([]byte, []byte, error) {
+	return r.Run(ctx, name, args...)
+}
+
+func (r *forgetRunner) Run(_ context.Context, name string, args ...string) ([]byte, []byte, error) {
+	r.calls = append(r.calls, append([]string{name}, args...))
+	if len(args) == 2 && args[0] == "unit-forget" {
+		_ = os.Remove(filepath.Join(r.unitDir, args[1]))
+	}
+	return nil, nil, nil
+}
+
+type forgetCase struct {
+	intent   DriveIntent
+	present  bool // the by-uuid link resolves
+	mounted  bool
+	enabled  bool // the wants-symlink exists
+	mountErr bool
+}
+
+const forgetUUID = "4ebfd36d-87d3-4b63-839e-a5a710fc799a"
+
+func runForget(t *testing.T, c forgetCase, intent IntentReader) (int, *forgetRunner, *RegistryKnownTargets, string) {
+	t.Helper()
+	unitDir := t.TempDir()
+	unit := "mnt-media_a.mount"
+	writeUnit(t, unitDir, unit, renderMountUnit(MountSpec{Name: forgetUUID, UUID: forgetUUID, Where: "/mnt/media_a", FSType: "ext4"}))
+	if c.enabled {
+		if err := os.MkdirAll(filepath.Join(unitDir, "multi-user.target.wants"), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.Symlink(filepath.Join(unitDir, unit), filepath.Join(unitDir, "multi-user.target.wants", unit)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	devRoot := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(devRoot, "by-uuid"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if c.present {
+		dev := filepath.Join(devRoot, "sdz1")
+		if err := os.WriteFile(dev, nil, 0o600); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.Symlink(dev, filepath.Join(devRoot, "by-uuid", forgetUUID)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	withDevDiskRoot(t, devRoot)
+	hr := &fakeHostReader{}
+	if c.mounted {
+		hr.mounts = []Mount{{MountPoint: "/mnt/media_a", FSType: "ext4"}}
+	}
+	if c.mountErr {
+		hr.mountsErr = os.ErrPermission
+	}
+	rr := &forgetRunner{unitDir: unitDir}
+	ops := NewSudoHostOps(SudoHostOpsConfig{Runner: rr, Bins: Binaries{}.withDefaults(), UnitDir: unitDir,
+		StageDir: t.TempDir(), Host: hr, Logger: quietLogger()})
+	if intent == nil {
+		intent = &fakeIntent{m: map[string]DriveIntent{"uuid:" + forgetUUID: c.intent}}
+	}
+	n := ops.ForgetRetiredDrives(context.Background(), intent)
+	return n, rr, NewRegistryKnownTargets(unitDir, intent, quietLogger()), unitDir
+}
+
+// The R-1124 case: decommissioned, then unplugged. The unit is disabled and removed through the root checker, systemd
+// is reloaded, and the CONSEQUENCE — the agent no longer lists the drive — holds. RED-PROOF (P1/redproof-forget.txt):
+// with ForgetRetiredDrives' body replaced by `return 0` this test fails (the drive is still listed).
+func TestForgetRetiredDrives_DecommissionedAndGone_UnitRemovedAndNotListed(t *testing.T) {
+	n, rr, known, _ := runForget(t, forgetCase{intent: IntentDecommissioned, enabled: true}, nil)
+	if n != 1 {
+		t.Fatalf("removed = %d, want 1 (calls %v)", n, rr.calls)
+	}
+	pa := Binaries{}.withDefaults()
+	want := [][]string{
+		{pa.Systemctl, "disable", "--", "mnt-media_a.mount"},
+		{pa.PrivApply, "unit-forget", "mnt-media_a.mount"},
+		{pa.Systemctl, "daemon-reload"},
+	}
+	if !reflect.DeepEqual(rr.calls, want) {
+		t.Fatalf("calls = %v\nwant    %v", rr.calls, want)
+	}
+	got, err := known.Known(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got) != 0 {
+		t.Fatalf("the forgotten drive is still listed: %+v", got)
+	}
+}
+
+// Already disabled (the Tester 1 shape after a manual disable): no disable call, the removal still runs.
+func TestForgetRetiredDrives_AlreadyDisabled_NoDisableCall(t *testing.T) {
+	n, rr, _, _ := runForget(t, forgetCase{intent: IntentDecommissioned}, nil)
+	pa := Binaries{}.withDefaults()
+	want := [][]string{{pa.PrivApply, "unit-forget", "mnt-media_a.mount"}, {pa.Systemctl, "daemon-reload"}}
+	if n != 1 || !reflect.DeepEqual(rr.calls, want) {
+		t.Fatalf("n=%d calls=%v want %v", n, rr.calls, want)
+	}
+}
+
+// Everything that must KEEP its unit: a decommissioned drive still plugged in (re-enroll is one click — H3), one still
+// mounted, an enrolled or ejected drive that is merely absent (it comes back), a drive with no recorded intent, and an
+// unreadable mount table. No privileged call at all, and the drive stays listed where it was listed.
+// RED-PROOF: with the present-drive check removed, the "present" case calls unit-forget and fails.
+func TestForgetRetiredDrives_KeepsEveryOtherUnit(t *testing.T) {
+	for name, c := range map[string]forgetCase{
+		"decommissioned but present": {intent: IntentDecommissioned, present: true, enabled: true},
+		"decommissioned but mounted": {intent: IntentDecommissioned, mounted: true},
+		"enrolled and absent":        {intent: IntentEnrolled},
+		"ejected and absent":         {intent: IntentEjected},
+		"no intent":                  {intent: IntentNew},
+		"mount table unreadable":     {intent: IntentDecommissioned, mountErr: true},
+	} {
+		t.Run(name, func(t *testing.T) {
+			n, rr, _, unitDir := runForget(t, c, nil)
+			if n != 0 || len(rr.calls) != 0 {
+				t.Fatalf("n=%d calls=%v — this unit must be kept", n, rr.calls)
+			}
+			if _, err := os.Stat(filepath.Join(unitDir, "mnt-media_a.mount")); err != nil {
+				t.Fatalf("unit file gone: %v", err)
+			}
+		})
+	}
+}
+
+// No intent store on the host (open failure): nothing is known to be retired, so nothing is forgotten.
+func TestForgetRetiredDrives_NilIntent_ForgetsNothing(t *testing.T) {
+	unitDir := t.TempDir()
+	writeUnit(t, unitDir, "mnt-media_a.mount", renderMountUnit(MountSpec{Name: "x", UUID: forgetUUID, Where: "/mnt/media_a"}))
+	rr := &forgetRunner{unitDir: unitDir}
+	ops := NewSudoHostOps(SudoHostOpsConfig{Runner: rr, UnitDir: unitDir, StageDir: t.TempDir(), Host: &fakeHostReader{}, Logger: quietLogger()})
+	if n := ops.ForgetRetiredDrives(context.Background(), nil); n != 0 || len(rr.calls) != 0 {
+		t.Fatalf("n=%d calls=%v", n, rr.calls)
+	}
+}

internal/storage/hostops.go

@@ -419,8 +419,9 @@ func (h *SudoHostOps) InspectDevice(ctx context.Context, device string) (DeviceP
 	// any TYPE/PTTYPE/USAGE line is positive data-bearing evidence. Its exit code is NOT relied
 	// on (blkid exits 2 on a blank device) — output presence is what matters. A broken/empty
 	// blkid simply adds no positive evidence; lsblk (below) is the read-success authority.
-	bout, _, _ := h.runner.Run(ctx, h.bins.Blkid, "-p", "-o", "export", device)
+	bout, _, berr := h.runner.Run(ctx, h.bins.Blkid, "-p", "-o", "export", device)
-	for k, v := range parseBlkidExport(bout) {
+	blk := parseBlkidExport(bout)
+	for k, v := range blk {
 		switch k {
 		case "TYPE":
 			probe.HasFilesystem = true
@@ -439,11 +440,11 @@ func (h *SudoHostOps) InspectDevice(ctx context.Context, device string) (DeviceP
 	// lsblk -J is the READ-SUCCESS authority + the partition/mount view. It exits 0 on any valid
 	// device (blank or not), so a clean parse means the agent reliably read the device. If lsblk
 	// fails, Probed stays false → DataBearing()=true (fail-safe — never call a device blank on a
-	// failed read).
+	// failed read). An exit 0 whose output does not parse into a device is NOT a read either (R-1093
+	// review): it stays unprobed, never "blank".
 	lout, _, lerr := h.runner.Run(ctx, h.bins.Lsblk, "-J", "-o", "NAME,FSTYPE,PTTYPE,MOUNTPOINT", device)
-	if lerr == nil {
+	if hasChildren, mounted, fstype, pttype, parsed := parseLsblkDevice(lout); lerr == nil && parsed {
 		probe.Probed = true
-		hasChildren, mounted, fstype, pttype := parseLsblkDevice(lout)
 		if hasChildren {
 			probe.HasPartitions = true
 		}
@@ -456,13 +457,33 @@ func (h *SudoHostOps) InspectDevice(ctx context.Context, device string) (DeviceP
 				probe.FSType = fstype
 			}
 		}
-		if pttype != "" {
+		if pttype != "" && !inheritedPartitionTable(blk, berr, pttype) {
 			probe.HasPartitionTable = true
 		}
 	}
 	return probe, nil
 }
 
+// inheritedPartitionTable reports whether lsblk's PTTYPE for the probed node is only the PARENT disk's
+// table (R-1093). lsblk prints the disk's table type on every one of its partitions, so an EMPTY
+// partition read as "device has a partition table" and asked for the typed erase confirmation.
+//
+// The partition's OWN signatures come from `blkid -p` on the partition itself (above): a nested table
+// inside it is its PTTYPE, a filesystem is its TYPE. So lsblk's value is set aside ONLY when blkid
+// positively proved two things — that it completed its probe of the device (exit 0: PART_ENTRY_* alone
+// comes from the PARENT's table, so output without a clean exit does not prove the partition's contents
+// were read), and that the device is a partition of a table of exactly that type (PART_ENTRY_NUMBER +
+// PART_ENTRY_SCHEME == lsblk's PTTYPE). Anything else keeps the old, safe reading: a blkid that failed,
+// exited non-zero or printed nothing, a whole disk, or a scheme that differs from lsblk's (then lsblk sees
+// a table blkid did not attribute to the parent). Pinned by TestInspect_R1093_* (hostops_disk_test.go).
+func inheritedPartitionTable(blkid map[string]string, blkidErr error, lsblkPTType string) bool {
+	if blkidErr != nil || blkid["PART_ENTRY_NUMBER"] == "" {
+		return false
+	}
+	scheme := blkid["PART_ENTRY_SCHEME"]
+	return scheme != "" && strings.EqualFold(scheme, lsblkPTType)
+}
+
 // Format runs mkfs.<fstype> on a validated device. The caller is responsible for authorization
 // (8C: only after classifying the device non-data-bearing, or via a slice-10 operator signature).
 func (h *SudoHostOps) Format(ctx context.Context, device, fstype string) error {
@@ -541,13 +562,14 @@ type lsblkDevice struct {
 }
 
 // parseLsblkDevice parses `lsblk -J -o NAME,FSTYPE,PTTYPE,MOUNTPOINT <device>` for the top device:
-// whether it has child partitions, is mounted (itself or any child), and its fstype/pttype.
+// whether it has child partitions, is mounted (itself or any child), and its fstype/pttype. parsed=false
-func parseLsblkDevice(out []byte) (hasChildren, mounted bool, fstype, pttype string) {
+// when the output is not a device listing — the caller must then treat the device as unprobed.
+func parseLsblkDevice(out []byte) (hasChildren, mounted bool, fstype, pttype string, parsed bool) {
 	var doc struct {
 		BlockDevices []lsblkDevice `json:"blockdevices"`
 	}
 	if json.Unmarshal(out, &doc) != nil || len(doc.BlockDevices) == 0 {
-		return false, false, "", ""
+		return false, false, "", "", false
 	}
 	d := doc.BlockDevices[0]
 	fstype, pttype = d.FSType, d.PTType
@@ -558,7 +580,7 @@ func parseLsblkDevice(out []byte) (hasChildren, mounted bool, fstype, pttype str
 			mounted = true
 		}
 	}
-	return hasChildren, mounted, fstype, pttype
+	return hasChildren, mounted, fstype, pttype, true
 }
 
 // NoopHostOps is the safe fallback when the privileged surface is unavailable or declined

internal/storage/hostops_disk_test.go

@@ -222,3 +222,114 @@ func TestInspect_ReadsFilesystemUUID(t *testing.T) {
 		t.Fatalf("FSUUID = %q, want the blkid UUID", p.FSUUID)
 	}
 }
+
+// ---- R-1093: an EMPTY partition is blank; lsblk's PTTYPE on a partition is the PARENT disk's ----------
+
+// blkidEmptyGPTPart is `blkid -p -o export /dev/sdb1` for an empty GPT partition: only the partition's
+// own entry in the parent table (PART_ENTRY_*), no TYPE, no PTTYPE — nothing inside the partition.
+const blkidEmptyGPTPart = "DEVNAME=/dev/sdb1\nPART_ENTRY_SCHEME=gpt\nPART_ENTRY_UUID=3b0c1a52-55a4-4c4e-9d0b-7a1f0e0f7c11\n" +
+	"PART_ENTRY_TYPE=0fc63daf-8483-4772-8e79-3d69d8477de4\nPART_ENTRY_NUMBER=1\nPART_ENTRY_OFFSET=2048\n" +
+	"PART_ENTRY_SIZE=2097152\nPART_ENTRY_DISK=8:16\n"
+
+// lsblkPartInheritsGPT is `lsblk -J -o NAME,FSTYPE,PTTYPE,MOUNTPOINT /dev/sdb1`: lsblk prints the DISK's
+// table type on the partition — the input that made the agent read an empty partition as data-bearing.
+const lsblkPartInheritsGPT = `{"blockdevices":[{"name":"sdb1","fstype":null,"pttype":"gpt","mountpoint":null}]}`
+
+// TestInspect_R1093_PartitionVerdicts pins the CONSEQUENCE (DataBearing + Reason, what decides whether
+// the typed erase confirmation is asked) for a partition, in both directions.
+//
+// COMPANION RED-PROOFS (observed 2026-10-18): (1) restore the pre-fix `if pttype != "" {` (drop the
+// inheritedPartitionTable call) → BOTH empty-partition rows (GPT and dos) FAIL with
+// `DataBearing=true reason="device has a partition table", want blank`. (2) drop the `blkidErr != nil`
+// check → the "exit 2" and "I/O error" rows FAIL with `DataBearing=false reason="device is blank"`.
+// (3) drop `&& parsed` → both lsblk-exit-0 rows FAIL with `Probed=true … reason="device is blank"`
+// (the unparsable-output hole predates R-1093; closed here). Restored.
+func TestInspect_R1093_PartitionVerdicts(t *testing.T) {
+	cases := []struct {
+		name       string
+		dev        string
+		blkid      string
+		blkidErr   error
+		lsblk      string
+		lsblkErr   error
+		wantBear   bool
+		wantReason string
+		unprobed   bool // the verdict must be "could not be reliably inspected"
+	}{
+		{name: "empty GPT partition is blank (the R-1093 case)", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, lsblk: lsblkPartInheritsGPT,
+			wantBear: false, wantReason: "device is blank"},
+		{name: "empty MBR (dos) partition is blank", dev: "/dev/sdb2",
+			blkid:    "DEVNAME=/dev/sdb2\nPART_ENTRY_SCHEME=dos\nPART_ENTRY_TYPE=0x83\nPART_ENTRY_NUMBER=2\nPART_ENTRY_DISK=8:16\n",
+			lsblk:    `{"blockdevices":[{"name":"sdb2","fstype":null,"pttype":"dos","mountpoint":null}]}`,
+			wantBear: false, wantReason: "device is blank"},
+		{name: "partition with a filesystem stays data-bearing", dev: "/dev/sdb1",
+			blkid:    "DEVNAME=/dev/sdb1\nUUID=8c1e\nTYPE=ext4\nUSAGE=filesystem\n" + blkidEmptyGPTPart[len("DEVNAME=/dev/sdb1\n"):],
+			lsblk:    `{"blockdevices":[{"name":"sdb1","fstype":"ext4","pttype":"gpt","mountpoint":null}]}`,
+			wantBear: true, wantReason: "device has a ext4 filesystem"},
+		{name: "partition holding its OWN (nested) table stays data-bearing", dev: "/dev/sdb1",
+			blkid:    "DEVNAME=/dev/sdb1\nPTUUID=0b8d4358\nPTTYPE=dos\n" + blkidEmptyGPTPart[len("DEVNAME=/dev/sdb1\n"):],
+			lsblk:    lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "blkid printed nothing: lsblk's PTTYPE stands (fail-safe)", dev: "/dev/sdb1",
+			blkid: "", blkidErr: errors.New("sudo: a password is required"), lsblk: lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		// Fail-closed paths (the 2026-10-18 review): PART_ENTRY_* is read from the PARENT's table, so it
+		// proves nothing about the partition's contents unless blkid also completed (exit 0).
+		{name: "blkid exit 2 with only PART_ENTRY output: lsblk stands", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, blkidErr: errors.New("exit status 2"), lsblk: lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "blkid I/O error after PART_ENTRY output: lsblk stands", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, blkidErr: errors.New("exit status 1"), lsblk: lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "blkid binary missing: lsblk stands", dev: "/dev/sdb1",
+			blkid: "", blkidErr: errors.New(`exec: "/usr/sbin/blkid": file does not exist`), lsblk: lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "lsblk exit 0 with unparsable output: unprobed", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, lsblk: `lsblk: something odd`,
+			wantBear: true, unprobed: true},
+		{name: "lsblk exit 0 with no device listed: unprobed", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, lsblk: `{"blockdevices":[]}`,
+			wantBear: true, unprobed: true},
+		{name: "lsblk failed: unprobed", dev: "/dev/sdb1",
+			blkid: blkidEmptyGPTPart, lsblk: "", lsblkErr: errors.New("exit status 32"),
+			wantBear: true, unprobed: true},
+		{name: "scheme differs from lsblk's PTTYPE: lsblk stands", dev: "/dev/sdb1",
+			blkid:    strings.Replace(blkidEmptyGPTPart, "PART_ENTRY_SCHEME=gpt", "PART_ENTRY_SCHEME=dos", 1),
+			lsblk:    lsblkPartInheritsGPT,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "whole disk with a GPT stays data-bearing", dev: "/dev/sdb",
+			blkid:    "DEVNAME=/dev/sdb\nPTUUID=a07f88ff\nPTTYPE=gpt\n",
+			lsblk:    `{"blockdevices":[{"name":"sdb","fstype":null,"pttype":"gpt","mountpoint":null}]}`,
+			wantBear: true, wantReason: "device has a partition table"},
+		{name: "whole disk: blkid silent, lsblk GPT stands", dev: "/dev/sdb",
+			blkid: "", lsblk: `{"blockdevices":[{"name":"sdb","fstype":null,"pttype":"gpt","mountpoint":null}]}`,
+			wantBear: true, wantReason: "device has a partition table"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			r := &scriptedRunner{
+				outputs: map[string][]byte{"blkid": []byte(c.blkid), "lsblk": []byte(c.lsblk)},
+				errs:    map[string]error{"blkid": c.blkidErr, "lsblk": c.lsblkErr},
+			}
+			p, err := newSudo(r).InspectDevice(context.Background(), c.dev)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if c.unprobed {
+				if p.Probed || !p.DataBearing() || p.Reason() != "device could not be reliably inspected" {
+					t.Fatalf("Probed=%v DataBearing=%v reason=%q, want unprobed + data-bearing (probe %+v)",
+						p.Probed, p.DataBearing(), p.Reason(), p)
+				}
+				return
+			}
+			if !p.Probed {
+				t.Fatal("expected a clean probe (lsblk read cleanly)")
+			}
+			if p.DataBearing() != c.wantBear || p.Reason() != c.wantReason {
+				t.Fatalf("DataBearing=%v reason=%q, want DataBearing=%v reason=%q (probe %+v)",
+					p.DataBearing(), p.Reason(), c.wantBear, c.wantReason, p)
+			}
+		})
+	}
+}

internal/thincheck/thincheck.go

@@ -0,0 +1,243 @@
+// Package thincheck finds an ORPHAN thin device: one the LVM thin pool holds that LVM itself no longer knows (R-1012).
+//
+// A full host root while vzdump takes its storage snapshot can leave such a device behind (demo-hp 2026-10-14: the
+// pool held devices 1, 2 and 4; LVM knew 1 and 2). From then on every snapshot backup of that guest fails at the
+// snapshot step („Creation of new snapshot 4 of device 2 failed") and nothing in the product said why. This package
+// NAMES the orphan, in plain words, in the backup's failure — it never removes one: the repair
+// (`dmsetup message <pool>-tpool 0 "delete <id>"`) stays a human step, decided by the operator (`03` whole-guest
+// backup; runbook `whole-guest-backup-orphan-thin.md`).
+//
+// The device list comes from the root helper `felhom-thin-check <vg> <pool>` (config bundle), which reads the pool
+// through a metadata snapshot and LVM's list before and after. An id is an orphan only when the pool holds it and
+// NEITHER LVM list names it, on two reads a few seconds apart — so a volume being created at that moment is never
+// reported.
+package thincheck
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// HelperPath is where the config bundle installs the root helper.
+const HelperPath = "/usr/local/sbin/felhom-thin-check"
+
+// Pool names one LVM thin pool.
+type Pool struct{ VG, Pool string }
+
+func (p Pool) String() string { return p.VG + "/" + p.Pool }
+
+// Reading is one helper run, parsed.
+type Reading struct {
+	PoolIDs   []int
+	LVMBefore map[int]string
+	LVMAfter  map[int]string
+}
+
+// Parse reads the helper's output. ok=false unless the closing "ok" line is there (a cut-off run is no reading).
+func Parse(out string) (Reading, bool) {
+	r := Reading{LVMBefore: map[int]string{}, LVMAfter: map[int]string{}}
+	done := false
+	for _, line := range strings.Split(out, "\n") {
+		f := strings.Fields(line)
+		switch {
+		case len(f) == 1 && f[0] == "ok":
+			done = true
+		case len(f) == 2 && f[0] == "pool":
+			if id, err := strconv.Atoi(f[1]); err == nil {
+				r.PoolIDs = append(r.PoolIDs, id)
+			}
+		case len(f) == 3 && (f[0] == "lvm-before" || f[0] == "lvm-after"):
+			id, err := strconv.Atoi(f[1])
+			if err != nil {
+				continue
+			}
+			if f[0] == "lvm-before" {
+				r.LVMBefore[id] = f[2]
+			} else {
+				r.LVMAfter[id] = f[2]
+			}
+		}
+	}
+	return r, done
+}
+
+// Orphans returns the ids the pool holds that neither LVM list names, sorted.
+func (r Reading) Orphans() []int {
+	var out []int
+	for _, id := range r.PoolIDs {
+		_, b := r.LVMBefore[id]
+		_, a := r.LVMAfter[id]
+		if !b && !a {
+			out = append(out, id)
+		}
+	}
+	sort.Ints(out)
+	return out
+}
+
+// Runner is the privileged exec surface (proxmox.ExecRunner in sudo mode).
+type Runner interface {
+	Run(ctx context.Context, name string, args ...string) ([]byte, []byte, error)
+}
+
+// StorageLister is the read the checker needs to map a guest's volumes to their thin pools (*proxmox.Client).
+type StorageLister interface {
+	GuestConfig(ctx context.Context, vmid int) (proxmox.GuestConfig, error)
+	ListStorage(ctx context.Context) ([]proxmox.Storage, error)
+}
+
+// Checker runs the check for one guest's pools.
+type Checker struct {
+	API    StorageLister
+	Runner Runner
+	Helper string        // "" → HelperPath
+	Pause  time.Duration // between the two reads when the first finds an orphan; 0 → 3 s
+	Logger *slog.Logger
+}
+
+// Finding is one pool with orphans.
+type Finding struct {
+	Pool    Pool
+	Orphans []int
+}
+
+// PoolsOf maps the guest's volume-backed disks (rootfs, mpN) to the lvmthin pools they live on.
+func (c *Checker) PoolsOf(ctx context.Context, vmid int) ([]Pool, error) {
+	cfg, err := c.API.GuestConfig(ctx, vmid)
+	if err != nil {
+		return nil, err
+	}
+	stores, err := c.API.ListStorage(ctx)
+	if err != nil {
+		return nil, err
+	}
+	thin := map[string]Pool{}
+	for _, s := range stores {
+		if s.Type == "lvmthin" && s.VGName != "" && s.ThinPool != "" {
+			thin[s.Storage] = Pool{VG: s.VGName, Pool: s.ThinPool}
+		}
+	}
+	vols := []string{cfg.RootFS}
+	mps := cfg.MountPoints()
+	keys := make([]string, 0, len(mps))
+	for k := range mps {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	for _, k := range keys {
+		vols = append(vols, mps[k])
+	}
+	seen := map[Pool]bool{}
+	var out []Pool
+	for _, v := range vols {
+		store, _, ok := strings.Cut(strings.Split(v, ",")[0], ":")
+		if !ok || strings.HasPrefix(store, "/") {
+			continue // a bind mount (a path) is no volume
+		}
+		if p, isThin := thin[store]; isThin && !seen[p] {
+			seen[p] = true
+			out = append(out, p)
+		}
+	}
+	return out, nil
+}
+
+// read runs the helper once.
+func (c *Checker) read(ctx context.Context, p Pool) (Reading, error) {
+	helper := c.Helper
+	if helper == "" {
+		helper = HelperPath
+	}
+	out, stderr, err := c.Runner.Run(ctx, helper, p.VG, p.Pool)
+	if err != nil {
+		return Reading{}, fmt.Errorf("felhom-thin-check %s: %w: %s", p, err, strings.TrimSpace(string(stderr)))
+	}
+	r, ok := Parse(string(out))
+	if !ok {
+		return Reading{}, fmt.Errorf("felhom-thin-check %s: the output has no closing ok line", p)
+	}
+	return r, nil
+}
+
+// Check returns the pools of vmid that hold an orphan (two reads, intersected). A pool that cannot be read is logged
+// and left out: the check FAILS OPEN — it explains a failure, it never stops a backup.
+func (c *Checker) Check(ctx context.Context, vmid int) []Finding {
+	logger := c.Logger
+	if logger == nil {
+		logger = slog.Default()
+	}
+	pools, err := c.PoolsOf(ctx, vmid)
+	if err != nil {
+		logger.Warn("thincheck: cannot map the guest's disks to thin pools — not checked", "vmid", vmid, "err", err)
+		return nil
+	}
+	pause := c.Pause
+	if pause <= 0 {
+		pause = 3 * time.Second
+	}
+	var found []Finding
+	for _, p := range pools {
+		first, err := c.read(ctx, p)
+		if err != nil {
+			logger.Warn("thincheck: thin pool not read — not checked", "vmid", vmid, "pool", p.String(), "err", err)
+			continue
+		}
+		o1 := first.Orphans()
+		if len(o1) == 0 {
+			logger.Debug("thincheck: thin pool and LVM agree", "vmid", vmid, "pool", p.String(), "devices", len(first.PoolIDs))
+			continue
+		}
+		select {
+		case <-ctx.Done():
+			return found
+		case <-time.After(pause):
+		}
+		second, err := c.read(ctx, p)
+		if err != nil {
+			logger.Warn("thincheck: second read failed — not reported", "vmid", vmid, "pool", p.String(), "err", err)
+			continue
+		}
+		in2 := map[int]bool{}
+		for _, id := range second.Orphans() {
+			in2[id] = true
+		}
+		var both []int
+		for _, id := range o1 {
+			if in2[id] {
+				both = append(both, id)
+			}
+		}
+		if len(both) > 0 {
+			found = append(found, Finding{Pool: p, Orphans: both})
+			logger.Warn("thincheck: ORPHAN thin device(s) in the pool — LVM does not know them; NOT removed (a human step)",
+				"vmid", vmid, "pool", p.String(), "orphans", both)
+		}
+	}
+	return found
+}
+
+// Describe is the operator sentence for a backup failure. Plain words; names the pool and the ids; says nothing was
+// removed and where the repair is written down.
+func Describe(f []Finding) string {
+	if len(f) == 0 {
+		return ""
+	}
+	var parts []string
+	for _, x := range f {
+		ids := make([]string, len(x.Orphans))
+		for i, id := range x.Orphans {
+			ids[i] = strconv.Itoa(id)
+		}
+		parts = append(parts, fmt.Sprintf("thin pool %s holds device %s that LVM does not know", x.Pool, strings.Join(ids, ", ")))
+	}
+	return "probable cause: " + strings.Join(parts, "; ") +
+		" (an orphan snapshot, usually left when the host disk filled during a backup). Snapshot backups on this pool " +
+		"fail until an operator removes it; Felhom did not remove anything. Repair: runbook whole-guest-backup-orphan-thin.md"
+}

internal/thincheck/thincheck_test.go

@@ -0,0 +1,184 @@
+package thincheck
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"log/slog"
+	"os/exec"
+	"path/filepath"
+	"reflect"
+	"runtime"
+	"strings"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// The helper's output on demo-hp 2026-10-14, before the repair: the pool held 1, 2 and 4; LVM knew 1 and 2
+// (audits/power-cut-fixes-2026-10-14/E/thin-ls.txt).
+const orphanOut = `lvm-before 1 vm-9201-disk-0
+lvm-before 2 vm-9201-disk-1
+pool 1
+pool 2
+pool 4
+lvm-after 1 vm-9201-disk-0
+lvm-after 2 vm-9201-disk-1
+ok
+`
+
+const healthyOut = `lvm-before 1 vm-9201-disk-0
+lvm-before 2 vm-9201-disk-1
+pool 1
+pool 2
+lvm-after 1 vm-9201-disk-0
+lvm-after 2 vm-9201-disk-1
+ok
+`
+
+func quiet() *slog.Logger { return slog.New(slog.NewTextHandler(io.Discard, nil)) }
+
+func TestParse_TheOrphanOf20261014(t *testing.T) {
+	r, ok := Parse(orphanOut)
+	if !ok {
+		t.Fatal("a complete reading was not accepted")
+	}
+	if got := r.Orphans(); !reflect.DeepEqual(got, []int{4}) {
+		t.Fatalf("orphans = %v, want [4]", got)
+	}
+	h, _ := Parse(healthyOut)
+	if got := h.Orphans(); len(got) != 0 {
+		t.Fatalf("a healthy pool reported orphans %v", got)
+	}
+}
+
+// A volume LVM created WHILE the pool was read appears only in the after-list: never an orphan.
+func TestOrphans_AVolumeCreatedDuringTheReadIsNotAnOrphan(t *testing.T) {
+	r, _ := Parse(strings.Replace(orphanOut, "lvm-after 2 vm-9201-disk-1\n", "lvm-after 2 vm-9201-disk-1\nlvm-after 4 snap_vm-9201-disk-1_vzdump\n", 1))
+	if got := r.Orphans(); len(got) != 0 {
+		t.Fatalf("orphans = %v, want none", got)
+	}
+}
+
+// A run cut off before its closing "ok" (the helper died, the output is partial) is no reading at all.
+func TestParse_NoClosingLineIsNoReading(t *testing.T) {
+	if _, ok := Parse(strings.TrimSuffix(orphanOut, "ok\n")); ok {
+		t.Fatal("a cut-off run was accepted as a reading")
+	}
+}
+
+type fakeAPI struct {
+	cfg    proxmox.GuestConfig
+	stores []proxmox.Storage
+}
+
+func (f *fakeAPI) GuestConfig(context.Context, int) (proxmox.GuestConfig, error) { return f.cfg, nil }
+func (f *fakeAPI) ListStorage(context.Context) ([]proxmox.Storage, error)        { return f.stores, nil }
+
+func guest9201() *fakeAPI {
+	extra := map[string]json.RawMessage{
+		"mp0": json.RawMessage(`"local-lvm:vm-9201-disk-1,mp=/var/lib/felhom,backup=1,size=20G"`),
+		"mp8": json.RawMessage(`"/mnt/felhom-drives,mp=/mnt/felhom-drives"`),
+		"mp9": json.RawMessage(`"local:9201/vm-9201-disk-2.raw,mp=/x"`),
+	}
+	return &fakeAPI{
+		cfg: proxmox.GuestConfig{RootFS: "local-lvm:vm-9201-disk-0,size=8G", Extra: extra},
+		stores: []proxmox.Storage{
+			{Storage: "local", Type: "dir", Path: "/var/lib/vz"},
+			{Storage: "local-lvm", Type: "lvmthin", VGName: "pve", ThinPool: "data"},
+			{Storage: "felhom-pbs", Type: "pbs"},
+		},
+	}
+}
+
+type seqRunner struct {
+	outs  []string
+	errs  []error
+	calls [][]string
+}
+
+func (r *seqRunner) Run(_ context.Context, name string, args ...string) ([]byte, []byte, error) {
+	r.calls = append(r.calls, append([]string{name}, args...))
+	i := len(r.calls) - 1
+	var err error
+	if i < len(r.errs) {
+		err = r.errs[i]
+	}
+	if i < len(r.outs) {
+		return []byte(r.outs[i]), nil, err
+	}
+	return nil, nil, err
+}
+
+func TestPoolsOf_RootfsAndVolumeMountpointsOnly(t *testing.T) {
+	c := &Checker{API: guest9201(), Runner: &seqRunner{}, Logger: quiet()}
+	pools, err := c.PoolsOf(context.Background(), 9201)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !reflect.DeepEqual(pools, []Pool{{VG: "pve", Pool: "data"}}) {
+		t.Fatalf("pools = %v, want [pve/data] once (bind mounts and dir volumes are not thin)", pools)
+	}
+}
+
+// The consequence: an orphan seen on BOTH reads becomes a finding whose sentence names the pool and the id, and says
+// nothing was removed. The only command ever run is the read-only helper with <vg> <pool>.
+// RED-PROOF (P1/redproof-thincheck.txt): with Orphans() returning nil this test fails.
+func TestCheck_OrphanOnBothReads_IsNamed(t *testing.T) {
+	rr := &seqRunner{outs: []string{orphanOut, orphanOut}}
+	c := &Checker{API: guest9201(), Runner: rr, Pause: 1, Logger: quiet()}
+	f := c.Check(context.Background(), 9201)
+	if len(f) != 1 || f[0].Pool.String() != "pve/data" || !reflect.DeepEqual(f[0].Orphans, []int{4}) {
+		t.Fatalf("findings = %+v", f)
+	}
+	want := [][]string{{HelperPath, "pve", "data"}, {HelperPath, "pve", "data"}}
+	if !reflect.DeepEqual(rr.calls, want) {
+		t.Fatalf("calls = %v, want only the read-only helper twice", rr.calls)
+	}
+	d := Describe(f)
+	for _, s := range []string{"thin pool pve/data holds device 4 that LVM does not know", "Felhom did not remove anything",
+		"whole-guest-backup-orphan-thin.md"} {
+		if !strings.Contains(d, s) {
+			t.Fatalf("Describe = %q, missing %q", d, s)
+		}
+	}
+}
+
+// An id that is gone on the second read (a volume mid-creation) is not reported; a healthy pool is read ONCE.
+func TestCheck_TransientOrHealthy_NothingReported(t *testing.T) {
+	rr := &seqRunner{outs: []string{orphanOut, healthyOut}}
+	if f := (&Checker{API: guest9201(), Runner: rr, Pause: 1, Logger: quiet()}).Check(context.Background(), 9201); len(f) != 0 {
+		t.Fatalf("a transient id was reported: %+v", f)
+	}
+	rr = &seqRunner{outs: []string{healthyOut}}
+	if f := (&Checker{API: guest9201(), Runner: rr, Logger: quiet()}).Check(context.Background(), 9201); len(f) != 0 || len(rr.calls) != 1 {
+		t.Fatalf("healthy: findings %+v, calls %v", f, rr.calls)
+	}
+	if Describe(nil) != "" {
+		t.Fatal("no finding must describe as empty")
+	}
+}
+
+// The helper refusing (a metadata snapshot held by someone else, exit 3) or failing: no finding, no panic — the check
+// fails open; it explains a failure, it never stops a backup.
+func TestCheck_HelperRefuses_FailsOpen(t *testing.T) {
+	rr := &seqRunner{errs: []error{errors.New("exit status 3")}}
+	if f := (&Checker{API: guest9201(), Runner: rr, Logger: quiet()}).Check(context.Background(), 9201); len(f) != 0 {
+		t.Fatalf("findings = %+v", f)
+	}
+}
+
+// The root helper's own tests (configs/test_felhom_thin_check.py) run with the Go suite.
+func TestRootHelperPythonSuite(t *testing.T) {
+	py, err := exec.LookPath("python3")
+	if err != nil {
+		t.Skip("python3 not found")
+	}
+	_, file, _, _ := runtime.Caller(0)
+	script := filepath.Join(filepath.Dir(file), "..", "..", "configs", "test_felhom_thin_check.py")
+	out, err := exec.Command(py, script).CombinedOutput()
+	if err != nil {
+		t.Fatalf("helper suite failed: %v\n%s", err, out)
+	}
+}

scripts/agent_gates.py

@@ -15,6 +15,9 @@ Gates (all must pass; **non-zero exit on any failure**):
                      or because someone is offline on a train. CI runs the FULL set for exactly this
                      reason: it is the machine that can afford a network check, and it is the half
                      that emails when something is wrong.
+  …  gate-decoys     the decoy suite (test_gate_decoys.py): every gate here still convicts its decoys
+                     (R-1116). ~20 s in the workspace, so the full run only: not the hook, but CI
+                     runs the full set on every push. Scratch clones and a local fake Gitea only.
 
 WHY THIS FILE EXISTS, WITH ONE GATE (2026-08-02, R-29 leg (b)).
 
@@ -58,6 +61,12 @@ GATES = [
     ("release-complete", os.path.join(ROOT, "scripts", "check-release-complete.py"), [], True),
     # R-389 — a REPORT.md observation with no register row behind it. Fast: stdlib file reads.
     ("observations", SHARED_OBSERVATIONS, [ROOT], True),
+    # R-1116 (2026-10-17) — the gates' OWN decoy suite. It was in no runner, the shape that let the catalog's
+    # R-1115 fail unseen for a night. Full run only: ~20 s in the workspace (measured, 30 cases; ~5 s in a
+    # CI-shaped clone, where the shared gates have fewer siblings to search), too slow for the hook — and CI
+    # runs the FULL set, so every push is still judged by it. It plants only in scratch clones and asks a fake
+    # Gitea on 127.0.0.1 (never the network, never the real tree); it does not call this runner (no recursion).
+    ("gate-decoys", os.path.join(ROOT, "scripts", "test_gate_decoys.py"), [], False),
 ]
 
 VERDICT = {0: "OK", 1: "FAILED", 2: "INCONCLUSIVE"}

scripts/release-agent.sh

@@ -69,6 +69,10 @@ fi
 existing="$(curl -fsS -o /dev/null -w '%{http_code}' \
     "$GITEA_BASE/api/packages/$GITEA_OWNER/generic/felhom-agent/$VERSION/felhom-agent" 2>/dev/null || true)"
 [[ "$existing" != "200" ]] || die "version $VERSION is ALREADY PUBLISHED — bump the version instead"
+# R-1098 (2026-10-16): only a clean 404 proves the version is new. Until now a registry that could not be reached
+# (curl prints 000) or answered anything else passed this check — „I could not ask" is not that proof.
+# Test: scripts/test_release_push_once.sh.
+[[ "$existing" == "404" ]] || die "could not establish whether $VERSION is published (registry answered HTTP '${existing:-none}') — refusing"
 
 # ── 3. Build ────────────────────────────────────────────────────────────────────────────────────
 BIN="$(mktemp -t felhom-agent-XXXXXX)"

scripts/test_agent_gates.py

@@ -57,6 +57,20 @@ class AgentGatesTest(unittest.TestCase):
         finally:
             shutil.rmtree(tmp, ignore_errors=True)
 
+    def test_decoy_suite_is_a_full_run_gate(self):
+        """R-1116: the decoy suite is a gate of the FULL run (which CI runs on every push), and --fast
+        names it as skipped (~20 s, too slow for the hook)."""
+        sys.dont_write_bytecode = True  # no __pycache__ left in scripts/ (not gitignored here)
+        sys.path.insert(0, os.path.join(ROOT, "scripts"))
+        import agent_gates as mod
+        rows = [g for g in mod.GATES if g[0] == "gate-decoys"]
+        self.assertEqual(len(rows), 1, "gate-decoys is not registered exactly once")
+        self.assertEqual(os.path.basename(rows[0][1]), "test_gate_decoys.py")
+        self.assertFalse(rows[0][3], "gate-decoys must not be --fast")
+        self.assertEqual([g[0] for g in mod.GATES if not g[3]], ["published", "gate-decoys"])
+        self.assertIn("--fast SKIPPED (deliberate periodic runs, never in a hook): published, gate-decoys",
+                      self.out)
+
     def test_unknown_argument_is_rejected(self):
         p = subprocess.run([sys.executable, ENTRY, "--nope"], cwd=ROOT,
                            stdout=subprocess.PIPE, stderr=subprocess.STDOUT)

scripts/test_gate_decoys.py

@@ -66,6 +66,9 @@ COVERS = {
                     "inside an HTML comment (R-426)",
     "observations": "R-419 in THIS repo's REPORT.md: an Observations note SAYING it carries no marker - vs the "
                     "two genuine markers (R-426)",
+    "gate-decoys": "this suite, run as a gate of agent_gates.py's full run, which CI runs on every push (R-1116): "
+                   "every case above is a decoy of its own gate, and ONE case that stops convicting turns the run "
+                   "red - red-proved 2026-10-17 by flipping a reuse-refs FACT case to expect rc 0",
 }
 
 fails = []

scripts/test_release_push_once.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# test_release_push_once.sh — release-agent.sh releases a version only when the registry says it is NEW (R-1098).
+#
+# The REAL release-agent.sh runs from a copy inside a scratch git repo (so its tag check reads scratch tags), with
+# RELEASE_ALLOW_DIRTY=1 (the clean-tree gate has its own history) and a fake `curl` and `go` first on PATH: curl answers
+# the package probe as FAKE_MODE says, and `go` stops the run right after the check, so nothing is built or published.
+#   published (200)        -> refused, "ALREADY PUBLISHED"
+#   absent (404)           -> passes the check, reaches the build
+#   unreachable (000)      -> refused, "could not establish"
+#   server error (500)     -> refused, "could not establish"
+#   the tag already exists -> refused, "already exists"
+# Run: bash scripts/test_release_push_once.sh   (exit 0 all cases right · 1 a case wrong)
+set -uo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+WS="$(mktemp -d)"; trap 'rm -rf "${WS:?}"' EXIT
+fails=0; ran=0
+mkdir -p "$WS/bin" "$WS/repo/scripts"
+cp "$HERE/release-agent.sh" "$WS/repo/scripts/release-agent.sh"
+(cd "$WS/repo" && git init -q -b main && git -c user.name=t -c user.email=t@t commit -q --allow-empty -m init)
+cat > "$WS/bin/curl" <<'FAKE'
+#!/bin/bash
+case "${FAKE_MODE}" in
+  published) printf 200; exit 0 ;;
+  absent)    printf 404; exit 22 ;;
+  down)      printf 000; exit 7 ;;
+  error)     printf 500; exit 22 ;;
+esac
+FAKE
+printf '#!/bin/sh\necho "FAKE-GO-REACHED $*"; exit 1\n' > "$WS/bin/go"
+chmod +x "$WS/bin/curl" "$WS/bin/go"
+run() {  # name, mode, expected rc, required fragment, forbidden fragment
+    local out rc
+    out=$(cd "$WS/repo" && FAKE_MODE="$2" RELEASE_ALLOW_DIRTY=1 GITEA_USER=u GITEA_TOKEN=t PATH="$WS/bin:$PATH" \
+          bash scripts/release-agent.sh 9.9.9 2>&1); rc=$?
+    ran=$((ran+1))
+    if [[ "$rc" != "$3" ]] || ! grep -qi -- "$4" <<<"$out" || { [[ -n "$5" ]] && grep -qi -- "$5" <<<"$out"; }; then
+        fails=$((fails+1)); echo "  FAIL $1 (exit $rc)"; echo "$out" | tail -4 | sed 's/^/        /'; return
+    fi
+    echo "  ok   $1"
+}
+echo "release-agent.sh — a version is released once, and only when the registry says it is new"
+run "published (200) -> refused"   published 1 "ALREADY PUBLISHED"     "FAKE-GO-REACHED"
+run "absent (404) -> proceeds"     absent    1 "FAKE-GO-REACHED"       "could not establish"
+run "unreachable (000) -> refused" down      1 "could not establish"   "FAKE-GO-REACHED"
+run "server error (500) -> refused" error    1 "could not establish"   "FAKE-GO-REACHED"
+(cd "$WS/repo" && git tag v9.9.9)
+run "the tag exists -> refused"    absent    1 "already exists"        "FAKE-GO-REACHED"
+echo
+if [[ $fails -eq 0 ]]; then echo "all $ran cases behaved"; exit 0; fi
+echo "$fails of $ran cases FAILED"; exit 1