INCIDENT + registers: ep0's PBS proxy served nobody for 9.5h (R-336..R-338)

Two whole_guest_backup_failed alerts at 04:30 and 04:32 CEST were one
incident, and not on either customer box: ep0's proxmox-backup-proxy was
active, holding its listening socket, and accepting nothing.

Root cause: accept() returning EMFILE. The process held exactly 1024 fds
-- its systemd-default soft RLIMIT_NOFILE -- of which 1016 were sockets
and 547 connections sat in CLOSE-WAIT. The 1024-deep accept backlog had
overflowed (Recv-Q 1025), so every client timed out. It was wedged from
its own loopback too, which is what moved this from a network problem to
a process problem.

Fed by ~85k requests/day (a flat 3,538/hour) against an endpoint written
to weekly, that leak reached the ceiling in 14 days of uptime.

Fix: LimitNOFILE=65536 drop-ins for both PBS units, restart, verified from
both boxes (200 in ~0.1s, felhom-pbs active), then re-drove the missed
backups through the product path -- POST /backup?target=felhom-pbs on each
agent's local API, not a hand-run vzdump.

  demo-felhom  ct/9201/2026-08-18T03:57:43Z  4.10 GB  36.4s
  demo-hp      ct/9201/2026-08-18T03:58:43Z  4.29 GB  41.5s

Both host reports now carry felhom-pbs success=true, so the hub is green on
the evidence rather than on a restart having been performed. No data lost,
no backup skipped: the daily local tier was never affected and the PBS tier
is weekly, so the window cost exactly one attempt.

Evidence copied off ep0 BEFORE the restart, per standing rule 5.

Filed: R-336 (the ~1 req/s poll rate is the real defect; the raised ceiling
is mitigation, not a cure), R-337 (a status endpoint that trailed its own
artifact by minutes then caught up -- WATCHING, downgraded from the defect
I first wrote, because it self-corrected), R-338 (demo-hp is not on the
R-50 island at all and nodes.md says it is; its local API is bound to the
customer LAN).

R-334 updated: still open, now one version wider (controller 0.216.0 vs
golden 0.214.0). golden-currency is the only failing gate and is inherited
-- it reads files this session did not touch -- so this push used
--no-verify, stated per .claude/rules/gates.md.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016p1PTCzb8rF5G9Aa1qhpBN

documentation/backlog/OPEN-ITEMS.md

@@ -634,5 +634,8 @@ class (an image `VOLUME` at an unmounted path) is still live — `immich-server`
 | **R-331** | **Disk health Phase 3 — growth-rate detection, and retiring the static 64.** The v0.215.0 count backstop (64 unreadable sectors → Hiba) is **a judgement from ONE drive**: the observed benign excursion peaked at 16 and cleared inside an hour, and the terminal run passed 64 at 13 Aug 11:28 and never came back. It is deliberately a backstop BEHIND the sustain rule, not the primary signal, but it is still a magic number tuned on a single sample and it will be wrong for some drive. With Phase 2's history the box can ask the question that actually matters — *is this count climbing, and how fast* — which distinguishes a drive with eight stable aging sectors from one adding forty a day, something no static threshold can do. Revisit 64 when that exists | **READY (M) — NEW 2026-08-14** | R-330 | Growth-rate rule over persisted samples; re-derive or delete the static 64 | CC |
 | **R-332** | **The new Hiba-from-counters path has never fired on real hardware.** v0.215.0's whole point is a verdict the product could not previously reach, and it is proven only against the committed fixture's values in unit tests (12 scenario groups, 11 of 12 red-proofs failing as required). The live validation on demo-hp proved the **negative** — three healthy disks still read Rendben across the deploy, no false alert — and the **severity wire** end to end, but no live disk has actually reached Hiba. **This is the honest gap and it must not be closed by pointing at the fixture tests**: the drive that produced the fixture is in DooPlex, which is Tier 2 and never a drill target, and the demo boxes are all-flash and healthy | **WATCHING — NEW 2026-08-14, NARROWED same day.** One item originally in this gap is now PROVEN LIVE: the **persisted state surviving a controller restart**. The v0.215.0→v0.216.0 redeploy destroyed and rebuilt the container, and the new one read back a `changed_at` written by the PREVIOUS version (`2026-08-14T07:23:14.640216851Z`, still intact at 09:31:35Z) instead of re-baselining — Scenario L on real hardware, not just the production-path unit test. **What remains unproven is the verdict itself, plus the stronger restart half: an already-ALERTED disk not re-alerting** | a real degrading disk, or an injection harness | **Closing condition:** a live disk reaching Hiba from counters, OR a deliberate injection through the REAL pipeline (agent `/disks` → controller check → hub event), not a hand-set verdict | CC |
 | **R-333** | **Two disk-health questions the deploy raised and did NOT act on.** **(a) The 55/60 °C bands are SPINNING-DISK bands applied to NVMe.** They were adopted unchanged from the operator's Prometheus config so the two systems cannot disagree — a deliberate, stated decision — but **measured on demo-hp 2026-08-14 the healthy Toshiba KXG50PNV1T02 NVMe idles at 53 °C, two degrees below Figyelmeztetés and seven below Hiba**, and NVMe routinely exceeds 60 °C under load with no fault whatever. As it stands a healthy customer NVMe under sustained write can be reported as **Hiba** — the single worst outcome this feature can produce. **(b) The agent runs bare `smartctl -a -j` with no `-n standby`** (`felhom-agent/internal/storage/hostops.go:368`), so every poll WAKES a spun-down drive; going 6h → hourly multiplies that by six. demo-hp is all-flash so the cadence measurement could not reveal it, and it was recorded rather than acted on per the task's own instruction. Mitigating datum from the fixture: the failing drive logged only **3375 load cycles in 60505 hours** (~one per 18h), i.e. that duty cycle barely spins down at all | **READY (S each) — NEW 2026-08-14** | — | (a) split the temperature bands by device class, or drop them for NVMe and rely on `critical_warning`; (b) add `-n standby` to the agent's smartctl invocation (an agent change, so fold it into R-330's session) | Viktor decides (a); CC does (b) |
-| **R-334** | **WAIVER + open item: controller v0.215.0 is released and deployed, and NO golden carries it.** Convicted by `golden_currency_gate.py` on the 2026-08-14 push: newest released controller **0.215.0**, newest golden bake **0.214.0** (`documentation/tests/golden-0.214.0-2026-08-12`). **A machine installed right now receives 0.214.0** — i.e. a brand-new box would ship WITHOUT the R-328 severity fix and would keep emailing nobody about a failing disk. The running fleet is unaffected (demo-hp guest 9201 is on 0.215.0 and healthy); this is purely the day-0 install path. **Not baked in this session deliberately:** the task scoped deployment to demo-hp only, and the second half of the fix — vouching the bake in the hub's day-0 artifact manifest — is **operator-password-gated, so CC cannot complete it**; a baked-but-unvouched golden is worse than none. **The push was made with `git push --no-verify` and it is stated here and in the session report**, per `.claude/rules/gates.md` — the gate has no waiver parser, so recording a waiver does not clear it | **READY (S) — NEW 2026-08-14** | operator availability for the vouch step | Bake a golden on 0.215.0 per `runbooks/RUNBOOK-manual-build.md` §4.1, then vouch it — a THREE-field change (`golden_version` + `agent_version` + `min_agent`). Until then every NEW install lacks the severity fix | CC bakes; **Viktor vouches** |
+| **R-334** | **WAIVER + open item: controller v0.215.0 is released and deployed, and NO golden carries it.** Convicted by `golden_currency_gate.py` on the 2026-08-14 push: newest released controller **0.215.0**, newest golden bake **0.214.0** (`documentation/tests/golden-0.214.0-2026-08-12`). **A machine installed right now receives 0.214.0** — i.e. a brand-new box would ship WITHOUT the R-328 severity fix and would keep emailing nobody about a failing disk. The running fleet is unaffected (demo-hp guest 9201 is on 0.215.0 and healthy); this is purely the day-0 install path. **Not baked in this session deliberately:** the task scoped deployment to demo-hp only, and the second half of the fix — vouching the bake in the hub's day-0 artifact manifest — is **operator-password-gated, so CC cannot complete it**; a baked-but-unvouched golden is worse than none. **The push was made with `git push --no-verify` and it is stated here and in the session report**, per `.claude/rules/gates.md` — the gate has no waiver parser, so recording a waiver does not clear it. **STILL OPEN and now one version WIDER, 2026-08-18:** the newest released controller is **0.216.0** (v0.215.0's own follow-up fix, R-335) and the newest bake is still **0.214.0**, so a new install now misses *two* releases. Re-convicted on this date's documentation-only push, which was likewise made with `--no-verify`; the gate reads `felhom-controller/CHANGELOG.md` and `documentation/tests/golden-*`, **neither of which that session touched** — the conviction is inherited, not caused | **READY (S) — NEW 2026-08-14, re-confirmed 2026-08-18** | operator availability for the vouch step | Bake a golden on **0.216.0** per `runbooks/RUNBOOK-manual-build.md` §4.1, then vouch it — a THREE-field change (`golden_version` + `agent_version` + `min_agent`). Until then every NEW install lacks the severity fix | CC bakes; **Viktor vouches** |
 | **R-335** | **One physical disk was walked TWICE per run, and the second walk sustained it against itself.** Found on live hardware ~2h after the v0.215.0 deploy, **by noticing the release's own positive observable disagreed with its own persisted artefact**: the hourly check logged *"3 disk(s) evaluated"* while `disk-health-state.json` held **two** records. Cause: demo-hp's `c11-scratch` and `felhom-backup` are the same physical NVMe (`/dev/nvme0n1`) and resolve to the same `diskKey`. **Not cosmetic** — `RunDiskHealthCheck` writes a disk's new record before the next entry reads it, so the SECOND copy consumed the FIRST copy's write as its prior: the disk **sustained against itself and reached Hiba on a FIRST sighting**, defeating truth-table row 6 — the exact rule separating a one-hour benign excursion from a false critical — and would have emitted **two identical events** for one drive. **Latent, not active, on demo-hp** (all three entries healthy, zero counters), but any aliased disk developing a single pending sector would have gone straight to Hiba. **This is the shape standing rule 3 warns about: an absent alarm was not evidence — the two artefacts had to be read AGAINST each other** | **CLOSED — controller v0.216.0, 2026-08-14.** Each `diskKey` is evaluated once per run; both entries stay marked `seen` so neither looks like a disappeared disk, and the card still renders both storage rows (the dedup is about state and alerts, not display). Pinned by `TestDiskCheck_SameDiskTwiceIsEvaluatedOnce`; companion red-proof run and reverted — deleting the guard makes the first sighting emit `Kind:2` (Hiba-from-sectors) at 8 sectors | — | — | CC |
+| **R-336** | **The offsite DR endpoint is polled about once per second, and that is what turned a slow leak into an outage.** ep0's PBS proxy served **~85,000 requests/day** — a flat **3,538/hour**, every hour, from two boxes: `74,445 GET /api2/json/admin/datastore` (`libwww-perl`, i.e. PVE's `pvestatd`) and `73,171 GET /admin/datastore/felhom-offsite/status` (`proxmox-backup-client`). Two pollers asking substantially the same question at the same rate. On 2026-08-18 this walked a connection leak in the proxy to its 1024-fd soft limit in **14 days**, wedging the offsite tier for 9½ hours (`audits/INCIDENT-ep0-pbs-fd-exhaustion-2026-08-18.md`). The `LimitNOFILE=65536` drop-in applied that morning raises the ceiling **but does not fix the leak** — it converts a fortnightly outage into a multi-year one, which is mitigation, not a fix. A DR endpoint that is written to weekly does not need to be asked about every second | **READY (M) — NEW 2026-08-18** | — | Find what polls `felhom-pbs` this hard (PVE storage status is the prime suspect, and its interval is tunable) and cut it; then confirm the fd count between restarts stops climbing — the positive observable, per standing rule 3 | CC |
+| **R-337** | **`/backup/status` lagged a completed backup by minutes on one box and not the other — and it RESOLVED ITSELF, which is why this is WATCHING and not a defect.** During the R-336 recovery on 2026-08-18, `demo-hp`'s snapshot landed on ep0 at **03:58:43Z** (complete manifest; the host's own task index says `OK`) — yet `GET /backup/status` was **still serving the superseded 03:27:00Z failure at ~04:03Z**, four-plus minutes later. `demo-felhom` showed its new result within ~40 s of completion. **The lag cleared on its own:** demo-hp's 04:07:35Z host report carries `felhom-pbs success=true, 4.29 GB`, and the hub is green for both boxes. **The first draft of this row claimed the success was "still reported as failed" — that was written before the next report arrived and it was wrong; the corrected claim is a several-minute skew between the two boxes, not a stuck value.** It is recorded because a status field that can trail its own artifact by minutes will, during an incident, be read as a second failure — this session nearly did — and because the asymmetry between the two boxes is unexplained | **WATCHING — NEW 2026-08-18** | another observation, ideally during an incident rather than constructed | **Do not open a fix on this as written.** First establish the intended refresh path for `/backup/status` after an out-of-schedule run; only if the skew is not simply collection cadence is there anything to pin. If it is cadence, close this row and say so | CC |
+| **R-338** | **`demo-hp` is not on the R-50 island at all, and `operations/nodes.md` states that it is.** The page records both fleet boxes as island-migrated 2026-07-25. True of `felhom-pve`; **false of `demo-hp`**, whose `agent.json` has `listen_addr: 192.168.0.87:8443` — the customer LAN address — and **no `island_bridge`/`island_guest_addr` keys at all**, whose guest 9201 has `net0` only (no `eth1`), and whose `vmbr9` exists with **zero members**. The controller's `controller.yaml` points at the LAN address, so the box works; this is inventory drift, not breakage. **Two costs.** A session trusting the page addresses the wrong endpoint — that happened on 2026-08-18 and the resulting timeout was briefly read as a fault. And the agent's local API is **bound to the customer LAN on this box** rather than to a point-to-point island, which is the exposure R-50 was built to remove — so a documented security property is claimed for a box that does not have it | **READY (S) — NEW 2026-08-18** | — | Decide which is true: migrate `demo-hp` to the island, or correct `nodes.md`. Leaving both is the one option that keeps the doc lying | Viktor decides; CC executes |