diff --git a/documentation/backlog/ROADMAP.md b/documentation/backlog/ROADMAP.md
index 299565c..39087e2 100644
--- a/documentation/backlog/ROADMAP.md
+++ b/documentation/backlog/ROADMAP.md
@@ -124,6 +124,13 @@ Each attempt runs the **full quiesce cycle**, so every customer app stack is STO
 
 Self-resolves the moment the target answers (the storage read succeeds, sees the archive, tier stops being due) — which is why it can hide indefinitely: it needs an offsite outage to appear at all. **Shape:** distinguish *storage unreachable* from *storage readable and empty*. Unreachable is UNKNOWN — defer the due-verdict rather than resolving it either way, exactly as R-81 made the hub do with a missing report. Only a target that is reachable AND has no archive is genuinely due. **Fix the window bypass in the same slice:** `AgeSecs == nil` must stop meaning "run now regardless of the hour". Either the agent distinguishes *never backed up* from *cannot tell* in what it reports, or `scheduledRunAllowed` gates on the former only — otherwise any future nil-age path re-opens the same hole. Note this does NOT weaken R-84's fail-safe intent: a tier whose storage is merely slow or briefly unreadable should still err toward backing up — it is specifically the **cold-store + unreachable** pair that must defer, because there the fallback has no information at all, only an empty default that looks like a fact. |
 | R-89 | **Retention is a COMMERCIAL attribute — it belongs to the hub, not to ep0 or a box** | M | idea — operator ruling 2026-07-27, first increment SHIPPED same day | **Ruling (2026-07-27):** retention is a per-customer *commercial* attribute (a paid tier may buy longer retention), so the **hub owns the policy** and ep0 merely executes it. Execution stays **server-side**: a reconciler writes a **PBS prune job** and PBS's own scheduler runs it, so hub downtime leaves the last-known policy running rather than silently stopping retention. **Increment 1 SHIPPED 2026-07-27** (`runbooks/RUNBOOK-pbs-prune-serverside-2026-07-27.md`): boxes no longer attempt prune (`keep_last: 0` → `allowPBSPrune=false`, config only — no code, no grant), and per-namespace prune jobs run on ep0 daily 03:30 UTC (`keep-last 2`), dry-run gated and verified `TASK OK`. This also closed a live defect — **every** demo-hp PBS backup since the tier was created on 07-26 had reported `job errors` while the data landed correctly, because `DatastoreBackup` grants `Datastore.Backup` but not `Datastore.Prune`; a tier that cries wolf on every success makes a genuine failure invisible, which is exactly what happened during the 07-27 migration window. **Remaining work:** retention becomes a per-customer attribute on the hub — a **policy** (keep-last/daily/weekly/monthly), NOT a bare number, since a paid tier will want "12 weekly + 6 monthly"; a reconciler converges it into the ep0 prune job in the same descriptor-and-converge shape as `pbs_dr`. **SECURITY PROPERTY TO PRESERVE — do not "fix" a future prune error by widening the grant:** box tokens stay **write-only** (`DatastoreBackup`), never `DatastorePowerUser`/`DatastoreAdmin`. A compromised box must not be able to delete its own offsite backups — that is the scenario offsite DR exists to survive. **PARALLEL QUESTION, unanswered and arguably the more urgent half: does the restic key on `storage-box-pool-1` (`u629488`) have DELETE rights?** If so the daily app-data offsite tier carries the identical exposure, and restic's **append-only** mode is the equivalent answer. **Rule once for both tiers.** Flips the capability-map row for offsite retention (currently: retention configured but unenforceable) |
+| R-95 | **The restic offsite tier's credential CAN DELETE — R-89's "parallel question", now ANSWERED** | M | idea — established read-only 2026-07-27 | **The exposure closed on the weekly PBS tier is fully open on the daily restic tier**, which holds the customer's actual documents and photos and is the only tier that survives losing the box. Established without mutating anything: **(1) Identity** — a per-customer *subaccount* on `storage-box-pool-1` (box 611714, bx11, `u629488`): `u629488-sub1` home `felhom-demo-felhom`, `sub2` peti-felhom, `sub3` demo-hp, each labelled `felhom-customer`. Auth is an **SSH key stored ON THE BOX** (`…/felhom-controller-data/_data/data/offbox/ssh_key`, 0600, beside `repo_password` + a pinned `known_hosts`) — customer-side, not hub-side, so a compromised guest holds it. **(2) Read-write: YES** — the API reports **`readonly=False` on all three subaccounts**, and it is not merely latent: the controller runs `restic forget --group-by host,tags --keep-daily 7 --keep-weekly … --prune` **from the box** (`backup/offbox.go:984`, also `:1070`). Delete rights are exercised on every run. **(3) Append-only: NO, and not expressible** — the repo is built as `sftp:` (`offbox.go:482`); restic's append-only mode requires the **REST server** backend, which plain SFTP cannot provide. **(4) A zero-code mitigation exists and is unused:** the box type carries `snapshot_limit=10` and the API reports `snapshot_plan=null` with **0 snapshots** and `size_snapshots=0`. Hetzner Storage Box snapshots are taken **server-side, outside the SFTP namespace** — an SFTP subaccount cannot delete them — so they are a genuine immutability layer at no extra cost and with no code change. **Rule once for both tiers, per R-89.** Options, cheapest first: enable a snapshot plan (operator click, immediate); split backup-write from prune so pruning runs somewhere the box cannot reach; or move the repo to restic's REST server with `--append-only`. Flips the capability-map row for offsite immutability |
+| R-94 | **The hub hands out host-install `1.19.0` while `1.20.0` is what ships R-82's backup default** | XS | idea — found 2026-07-27 | `hub/internal/web/configs.go:28` pins `const hostInstallVersion = "1.19.0"`, and that constant renders the customer page's install command (`configs.go:487` `ScriptVersion`). `scripts/felhom-host-install.sh:187` is `SCRIPT_VERSION="1.20.0"` — the version R-82 shipped to default a fresh box to **local-daily + offsite-weekly**. So a new install driven from the hub still gets the **pre-R-82** default. This is the same constant hub `v0.73.2` existed to sync (→1.19.0); it went stale again one installer release later, which is the actual finding: **a hand-synced version constant in a second repo drifts every time the first one ships.** v0.73.2 itself is NOT an undeployed artifact — its content is inside the live 0.77.0. Related to R-82's "REMAINING GATE" note, but that gate is about the fleet flip; this is the hub simply pointing at the older script |
+| R-90 | **ep0 has 3.8 GB and NO swap — it OOMs under a restore-test, and that gates R-86** | S | BLOCKED on Hetzner CX33 availability (operator, 2026-07-27) | Verified read-only: `Swap: 0B`, nothing in `/etc/fstab`, no swap unit, no `/swapfile` — **swap was never added.** On 2026-07-27 06:58:12 a global OOM fired while `proxmox-backup-proxy` held a 3.2 GB peak serving a 14.46 GB restore-test read and the datastore-migration `rsync` held 1.9 GB; the kernel killed the rsync and PBS stopped serving for ~15 minutes. That outage is what exposed **R-88** downstream. Still exposed today: 3.8 GB, no swap, one kill on record. **This gates R-86** — backup-aligned restore-testing puts a multi-GB read on this box *on a schedule*, so the headroom question must be settled before that lands. Interim lever if CX33 stays unavailable: a swapfile, which needs no console |
+| R-91 | **The old 13 GB datastore copy is still on ep0's root disk** | XS | WATCHING — gated on demo-felhom's first post-migration PBS backup | The datastore moved to a Hetzner Cloud Volume on 2026-07-27 (`/dev/sdb`, 100 GiB, attached 06:29:40 UTC, now `/mnt/pbs-datastore`, 13 G used of 98 G). The pre-migration copy survives at **`/srv/pbs-felhom`, 13 G**, on `/` (38 G total, 16 G used, 21 G free). **Do not delete yet:** demo-hp has landed two post-migration snapshots (07-27 08:25:47Z, 09:37:29Z) but **demo-felhom's newest is 2026-07-26T12:21:48Z — before the migration**, so the new volume has not yet proven a write for that namespace. Delete once it has. **Doc drift to fix in the same commit:** `CONTEXT.md:1018` still records the datastore at `/srv/pbs-felhom` |
+| R-92 | **The hub's PBS-DR gauge is 0.1 GB-granular, so small deltas are unverifiable** | XS | idea — 2026-07-27 | The PBS-DR box card rounds to 0.1 GB, which is coarser than the changes an operator wants to confirm after a prune or a GC — a successful prune of a small namespace moves the number by less than one displayed digit, so the UI cannot distinguish "it worked" from "nothing happened". Cosmetic today; it becomes load-bearing the moment retention (R-89) is customer-visible and someone needs to see that a policy change took effect |
+| R-93 | **`drill-r50` is both a blocked customer and the only drift fixture** | XS | idea — 2026-07-27 | The drill customer is blocked in the hub (so it stops alarming) yet it is also the only record exercising the endpoint-drift path R-77 added. Blocking hides it from `GetActiveCustomerIDs`, so the fixture it provides is silently inert — a monitor with no live subject reads exactly like a monitor that passes. Decide: retire it and build a synthetic fixture, or unblock it and silence per-customer instead (the operator has a per-alert silencing feature planned). Related to the R-50 drill VM, now shut down |
+| R-96 | **Two standing rules were agreed in chat and never committed** | XS | idea — found 2026-07-27 | Verified by grep across every `CLAUDE.md`, `CONTEXT.md`, `REUSE.md` and `skills/`: **(a) "never combine a test run and a commit in one command"** — zero hits, despite three separate occurrences (the worst pushed a red suite because `packages ok: 28` was read while `rc=1` was not); **(b) "a 'no access' claim must list what was tried"** — zero hits, despite a wrong "no access exists" verdict on ep0 on 2026-07-27 where the working route was documented in memory and only one path had been attempted. **The third rule IS committed** — `CONTEXT.md:8` S-1, N.5's third leg (architecture docs same-session coupled). A rule that lives only in a chat log binds nobody; both belong in `CONTEXT.md` beside S-1/S-2 |
 | R-76 | **FileBrowser-created folders break the setgid chain, and a drop-zone's mode is not stable** | S | idea (surfaced by the R-75 spike, 2026-07-26) | Two related findings from `audits/SPIKE-catalog-data-paths-2026-07-26.md` P3/P5, both **pre-existing** and deliberately left alone by that spike. **(a)** FileBrowser Quantum 1.3.3 creates files `0644` and folders `0755` and does **not** propagate the setgid bit — even though the entrypoint wrapper's `umask 002` really is in effect (`/proc/1/status` `Umask: 0002`). Group inheritance itself works (a file uploaded into a 2775 group-100 dir landed group 100, not the process gid 1000), so the convention's *group* half holds and only its *mode* half is lost. The consequence is proven with a control: inside a UI-created `0755` folder a gid-1000 process's file landed group **1000**, while the identical write into the 2775 parent landed group **100**. So **any folder a customer creates through FileBrowser breaks the shared-group chain one level down.** Latent today — every userdata-touching catalog app that declares an identity declares uid/gid **1000**, the same uid FileBrowser runs as, so owner permissions mask it; it bites the day a content app runs as a different non-root uid with gid 1000. The comment at `infra/infra.go:156` is right that the image ignores `-e UMASK` but does not say the wrapper fails to achieve the intended mode either. **(b)** `import/calibre` is live on demo-felhom at `755 1000:1000` where every sibling is `2775 root:1000` — with `media/books` on the same box, same app, same deploy, at `2775` as the control, and no parser asymmetry (checked: `ParseComposeUserdataMounts` picks up both calibre-web binds). Consistent with the consuming app rewriting the mode of its own ingest dir after the deploy belt sets it; **not confirmed causally**. Together they mean **a drop-zone directory's mode is not stable against either the customer or the consuming app** — verify before building anything that assumes `import/*` stays 2775. Flips no capability-map row today (latent); would become customer-visible the moment a non-1000 content app enters the catalog |
 | R-21 | **Bare-metal Felhom ISO** — per-PVE-release auto-install ISO for blank customer hardware → first-boot wrapper (invokes `felhom-host-install.sh`) → universal secret-free / operator-bind (option C) | XL | **SHIPPED + PHYSICALLY CLOSED (slices A+B+C; rehearsal executed 2026-07-18)** | **PHYSICAL CLOSURE 2026-07-18** (`tests/VALIDATION-n100-rehearsal-2026-07-18.md`): the generic pairing ISO v1.20.0 (`--loader mkimage`, SB off) **booted the very AMI board that F1 blocked**, installed unattended, and the box self-registered as an unclaimed appliance the same second it first booted (16:17:14) → self-bind → credential → day-0 SUCCESS 16:32:32 → floor-lifted to current. **F1 is closed on physical hardware and the rehearsal dependency on this item is discharged.** Two residual notes stay open, neither blocking: PXE/network-boot is still unbuilt (the third F1 option, for boards where even USB-mkimage fails), and mkimage remains unsigned → **SB must be OFF**. The installer's GRUB menu still offers interactive installers → new item **R-38**. — Prior: **PHYSICAL RUN 2026-07-16 (`tests/VALIDATION-n100-baremetal-2026-07-16.md`):** demo N100 reinstalled clean-slate from a pipeline ISO → chain reached **rc-0 first try on real hardware** (closes slice A's operator-gated boundary), serial-filter safety proven on metal, PBS-DR reconciler self-healed on the reused peer, DMI verdict = key on MAC+UUID. **F1 (HIGH, slice-B input):** this cheap AMI `AN3PLUS 0.01` firmware won't UEFI-boot the ISO's GRUB from USB (`relocation 0x0`) — SB-off/shim-bypass don't help; worked around live with a `grub-mkimage` loader built from the box's own GRUB. Pipeline must ship a firmware-compatible loader / PXE path. Reused-customer edges (F2 claim re-issue, F3 offsite re-issue, F4 non-default-storage-id ACL 403) feed R-1/Peti. UX: F6 drive-init doesn't mount+attach, F5 guest-RAM not configurable, F7 back-route. — **Slice A (build pipeline + first-boot bootstrap) DONE + validated on VM 310:** build gate/red-proof, disk-filter fail-safe, stub→retry-unit→real public-channel host-install fetch+invoke→retry, resume-decision, exactly-once, no-net retry+recovery all GREEN. Operator-gated remainder: host-install rc-0 terminal success (drill customer needs the password-gated create-UI). **Slice B — SHIPPED (scripts v1.18.0, 2026-07-17):** the F1 firmware fix is now a first-class pipeline mode `build-felhom-iso.sh --loader shim|mkimage` (default shim; `mkimage` = monolithic grub-mkimage loader from the ISO's own GRUB, recipe from the run evidence). RUNBOOK-B legs on nested VM 311 proved it: shim boots+installs under OVMF SB-enforcing + SeaBIOS; mkimage boots+installs under OVMF SB-off; mkimage under SB-enforcing FAILS `Access Denied` (unsigned → **SB must be OFF**, documented); surgery byte-identical payload. **Physical N100 boot on the real board still pending** → folds into the supervised rehearsal (R-1; an `n100-safety` match-nothing ISO is built + sha-recorded for a zero-risk pre-flight). **PXE/network-boot** (the third F1 option, for boards where even USB-mkimage fails) stays a deferred note under this item — not built. **Slice C — SHIPPED (hub v0.62.0 + scripts v1.19.0, 2026-07-17):** the GENERIC secret-free universal ISO (`build-felhom-iso.sh --pairing`). The box self-registers as an unclaimed appliance (keyed by SMBIOS-uuid + MAC set — the DMI-verdict tiebreaker), the operator BINDS it to a customer on the Hosts page, and the hub delivers customer-id + retrieval passphrase ONCE (`/api/v1/appliance/register` + one-shot poll, 404-no-oracle — all live-verified through the public ingress); the bootstrap then falls through to the slice-A direct path. **One unit, two modes** (direct = byte-identical, regression-proven zero-appliance-calls). Artifact proven secret-free (baked env = hub URL only; manifest `secret-bearing: no`). The **bind is operator-password-gated** → the live boot→register→bind→day-0 composition (with a Viktor-created drill customer) + the physical N100 boot fold into the supervised rehearsal — **which now runs the COMPLETE final product flow in one pass: RESET the demo → boot the generic ISO → bind → day-0**. SSH-host-key pinning: stored + fingerprints displayed (attaching to the host on bind = future, no clean hand-off surface today). **Customer-facing self-bind page = R-27 (future).** Origin spike `audits/SPIKE-baremetal-iso-2026-07-16.md`: every mechanism GREEN on nested virt (VM 310 on felhom-pve). Zero-touch install BIOS **and** UEFI incl. **Secure Boot enforcing** (no MOK/keypress); first-boot hook `fully-up` = root + working pvesh/pct, exactly-once via `pending-first-boot-setup` flag; post-install **webhook** carries SMBIOS-UUID + management-MAC + host SSH keys → the unclaimed-appliance record; disk-filter installs only the target (canary byte-identical) and **fails-safe** on match-nothing / bad disk; `from-url` + `cert-fingerprint` **fails CLOSED**. OPEN (needs ONE real bare-metal run): vendor DMI serials (empty on virt), real firmware/NIC quirks, a pre-existing-LVM wipe step. Pipeline notes: assistant pairs to the ISO by Debian codename, ~11.5 s/ISO on DooPlex, **gate on `validate-answer` output not `$?`** (exit 0 on failure). Would flip a new capability-map MISSING row "customer self-installs on bare hardware" once spec'd. *(brief called this R-22)* | **SECOND-HARDWARE PROOF 2026-07-21 (slice C, demo-hp): the pairing flow is no longer a one-board result.** The universal secret-free ISO was booted on a completely different machine — an **HP t740 (Ryzen V1756B, AMI M42 firmware)** versus the N100 it was proven on — and the whole chain ran on **virgin hardware in one pass**: armed install → self-registration as an unclaimed appliance → operator bind → day-0, ending with a running customer guest 9201 and agent 0.92.1 checking in as host `demo-hp-bb76ea`. **Two things generalise from the second board specifically:** the **shim** loader booted with **Secure Boot ENABLED** (`mokutil --sb-state` → `SecureBoot enabled`), confirming the mkimage/SB-off dance is an N100-firmware workaround and NOT a Felhom requirement; and the exact-serial disk filter selected the SanDisk system SSD while leaving the box's **1TB NVMe untouched and unenrolled** (its prior NTFS partition is still intact, unmounted, in no LVM/ZFS) — the destructive path stayed inside its filter on hardware it had never seen. **Not clean, and the failures are filed:** the install got no DHCP on the 4-port NIC and baked a static fallback rather than aborting (**R-59**), which cost a cable move and a hand-repair; and the console was unreachable because the baked root password is unknowable (**R-61**)
 | R-29 | **The design-v2 green gates are not enforced anywhere — one has been RED for 16 releases.** `controller/scripts/docker_run_volume_path_gate.py` has failed continuously since **2026-07-14 (v0.129.0)** and nobody noticed until R-7b's close-out ran it by hand at v0.145.0. Two separable parts. **(a) The finding itself is benign and the fix is 3 lines.** The flagged call is `internal/appexport/estimate.go:179` `docker run --rm -v <volumeName>:/vol:ro alpine du` — a **NAMED-VOLUME** mount, i.e. daemon-side with no host path, which is the *safe* shape and byte-for-byte the same pattern as three entries already on the gate's ALLOWLIST (`export.go` `volName+":/vol"`, `backup.go` `volName+":/vol:ro"`, `restore.go` `volName+":/vol"`). It is NOT the v0.124.0 path-strand class the gate exists to catch — the author of the v0.129.0 F-A fix explicitly avoided that class (see the function's own comment) and simply never added the allowlist entry. So the fix is an ALLOWLIST addition WITH ITS WHY, **not** a docker-cp rewrite; anyone who 'fixes' this by rewriting the call has misread the gate. **(b) The systemic half is the real item:** the gates run only when a human remembers to run them, so a gate can sit red across 16 releases while every REPORT says 'green'. This is the SECOND instance of the class — cf. the v0.123.0 note *'Windows green gate silently red (read-only fsync)'*. Decide where they run (pre-push hook, `build.sh` step, or a CI job) and make a red gate block the train the way the Go green gate does. | S (a) / M (b) | idea | Origin: R-7b close-out, `felhom-controller` REPORT §4(f) — CC correctly left it alone as out-of-scope and pre-existing, and verified by stashing that it fails identically on the unmodified tree. Flips no capability-map row (engineering hygiene, no customer-visible behaviour). Affected gates to audit for the same rot: controller `template_id_gate` / `emoji_gate` / `native_confirm_gate` / `offbox_rename_gate` / `mojibake_gate` / `app_row_dedup_gate` / `docker_run_volume_path_gate`, hub `hub_confirm_gate`, manifests `manifest_bearer_gate`, website `site_gates`. **Do not bundle (a) into an unrelated feature commit** — it is a one-line behavioural claim about a mount's safety and deserves its own reviewed diff. **2026-07-18 rehearsal note:** the run's finding list independently re-raised "assign the pre-existing `docker_run_volume_path_gate` failure its ID so red stops normalizing" — **that is this item; no second ID was minted.** |