## Unreleased (2026-11-14 night) — built and NOT released: a clock step is not a new boot (felhom.eu R-1395); HSTS one year (R-1050); the web certificate's end date in the report (R-1379) **MinAgent: 0.158.1** (unchanged — the new field is optional; without it the controller behaves as before). - **A step of the host's clock no longer makes the controller recreate every drive-backed app (R-1395).** Found on the N100 on 2026-11-14: its time service came back, stepped the clock by +32.7 s, and radarr, sonarr and qbittorrent were recreated. **The boot token is the AGENT's, not the controller's:** `guest_boot_id` in the agent's `/disks` answer is `-`, and the kernel computes the boot time from the wall clock. The only reader in this repository is `web.processGuestBootChange` (it compares the token with `settings.LastGuestBootID`). It now decides through `guestBootVerdict`, which prefers the agent's new `guest_boot_key` (`-`; an agent after 0.174.0 — built there, not released): the key equal to the record is the same boot whatever the old token says; the first report with a key moves the record to the key WITHOUT a recreate when the old token says „same boot”; an agent that sends no key gets the old rule, byte for byte. Tests `TestGuestBootVerdict_AClockStepIsNotANewBoot` (14 cases), `TestGuestBootVerdict_IsWhatTheBootPassReads`; 8 mutations seen red (felhom.eu `documentation/audits/night-build-2026-11-14/C/redproof-controller-boot-key.txt`). **Not proven on a box.** With an older agent a clock step still recreates the apps once. - **The report carries the web certificate's end date (felhom.eu R-1379, D77, `09` §3 decision 298; `08` §12).** Only the box knows it: the certificate lives in Traefik's store in the guest (`/traefik/acme.json`), and nothing read it — a renewal that failed was silent until the browser warning on the end date. `infra.ReadWebCerts` decodes the store into a shape with NO key field, parses the certificate only, and returns names, end dates and the issuer's name (at most 8, soonest first); the report's `web_certs` carries them. A missing or empty store is no field; a broken store is a WARN line that does not quote the file, and the report still goes. The hub's warning (14 days before) is built in its own repository, not released. Tests `TestReadWebCerts_*` (real certificates and keys made in the test; the key, the account key and the e-mail must not appear in what is reported), `TestWebCertsForReport_PathFieldAndBrokenStore`; 8 mutations seen red (one stayed green at first: the test had no certificate in the folder — a real one and a decoy in the wrong folder were added). **Not proven on a box.** - **After review (2026-11-14): only certificates of the household's CURRENT domain are reported** (`web_certs`; the name, `*.`, or a name under it). Traefik's store keeps a certificate for a name the box no longer serves; it can never renew, and reporting it would warn the operator about a name nobody uses. 2 mutations seen red. - **HSTS is one year (R-1050, felhom.eu `09` §3 decision 249).** `infra.HSTSMaxAgeSeconds` 86400 → 31536000; still no `includeSubDomains`, no `preload` (the test asserts both). The ruling's condition — a month with no https trouble since 2026-10-15 — was searched on 2026-11-14: the open and closed registers (ASCII fragments, with a positive and a negative control), the operator's mailbox, the hub's log since its start today; no report of a name that does not open and none of a box answering http after the floor. **The release must be built on or after 2026-11-15** (the ruled date). A browser that has seen the new value keeps „https only” for that name for a year; going back to one day in a later release does not shorten what a browser already holds. ## v0.352.1 (2026-11-14) — Go 1.27.4: the security fix GO-2026-6859 (felhom.eu R-1391) — RELEASED (image `sha256:1a6e8945…`, from `8747744b`, CI run id 2637; rc.1 on scratch guest 9202); per-box floors 09:18 UTC, global floor 2026-11-14 09:21 UTC (MinAgent 0.158.1); golden 0.352.1 baked and vouched the same day ### 2026-11-14 **MinAgent: 0.158.1** (unchanged). **Why.** Go 1.27.3 has GO-2026-6859: an HTTP/1 server can crash with a runtime panic when a `MaxBytesReader` limit is hit while the handler reads the request body and writes a response header at the same time. Fixed in Go 1.27.4 (published; checked on go.dev and on Docker Hub before the change). **What changed.** `controller/go.mod` `go 1.27.4`; the builder image `golang:1.27.4-trixie@sha256:b54b2639de8c1bfe538ec5ffc45f29b1932235ac5f6d63c059853646677112d9` (version and digest, as before). No source change. **Reach, read 2026-11-14 (20-minute read).** Seven places wrap a body in `MaxBytesReader` (`api/router.go limitBody`, `web/poweroff_handlers.go`, `web/handler_export_upload.go` ×4, `web/escrow_handlers.go`). Each one reads the body to its end in the handler's own goroutine and writes the answer after it; none starts a goroutine that reads the body, and the controller has no `TimeoutHandler` and no `EnableFullDuplex`. So no controller handler matches the crash pattern. The release is hygiene, not an urgent fix. **Not in this release: HSTS one year (felhom.eu R-1050, `09` §3 decision 249).** The ruling allows it on or after 2026-11-15; this release was built on 2026-11-14 and the operator's words „one day early is OK” were not given. **Tests.** `go build`, `go vet`, `go test ./...` green on Go 1.27.4; `controller_gates.py --fast` green. **Released.** rc.1 on scratch guest 9202 (healthy, 0 ERROR lines). Tester 1, demo-hp and demo-felhom took 0.352.1 by per-box floors within a minute; on each: every container healthy, 0 ERROR lines, the running binary reads `go1.27.4`, `https_only_check.py` exit 0, and every public name still answers `strict-transport-security: max-age=86400` and nothing more (26 names). Global floor in its own save; 35 minutes later: all three on the release, no new event but `controller_updated`/`controller_started`. Per-box floors cleared. Evidence: felhom.eu `documentation/audits/release-1-dates-2026-11-14/controller/`. ## v0.352.0 (2026-11-10) — R-1371: one fill warning per filesystem — RELEASED (image `sha256:3c785dd3…`, from `3dbb68c1`, CI 2615; rc.1 on scratch guest 9202); per-box floors 15:16 UTC, global floor 2026-11-10 15:31 UTC (MinAgent 0.158.1) ### 2026-11-10 **MinAgent: 0.158.1** (unchanged). **Why (measured 2026-11-10, demo-hp guests 9201 and 9202 and the N100, read-only; felhom.eu `documentation/audits/space-list-2026-11-10/evidence/extra/fs-identity.txt`).** The fill watch has two built-in targets: "/" (the controller's overlay, which lies on the Docker data volume) and `/mnt/sys_drive`. Since the guest's two data volumes became one they are ONE filesystem: the same size, the same free space. The target list was de-duplicated by path only, so one fill would send two warnings with two names („Alkalmazások területe", „Rendszer- és mentési terület"). Not seen on a box: no disk was filled. **The fix.** `fillTargets` now keeps one target per filesystem (`fillTargetsBy`); the first one stays, in the order app area, system area, drives. A path whose filesystem cannot be identified is kept (one warning too many, never one too few). **The identity is NOT the device number alone** (`system.FilesystemKey`, `internal/system/fskey.go`). Inside the container "/" is an overlay: it has its own device number and its own filesystem id (139 against 64519 on demo-hp), so a plain `st_dev` compare would have passed its tests and changed nothing on a box. Equal size-and-free numbers are no identity either: two reads in one loop already differed by one block. An overlay path is identified by the device of a file Docker keeps on its data-root and bind-mounts into the container (`/etc/hostname`, `/etc/hosts`, `/etc/resolv.conf`) — measured equal to `/mnt/sys_drive`'s device on all three guests. **Tests.** `TestR1371_OneFillSendsOneWarning` (the real watcher and push closure: one warning, then one critical), `TestR1371_TargetsByFilesystem`, `TestR1371_ProductionListUsesTheRealIdentity`, `TestFilesystemKey_AnOverlayRootIsTheDockerDataRoot` (the measured numbers), `TestFilesystemKey_RealSyscalls`. **Red seen three times:** no filesystem check (2 warnings); the production list not wired to the real identity; an overlay keeping its own device number. **Proven on boxes, 2026-11-10** (felhom.eu `documentation/audits/release-space-fixes-2026-11-10/controller/`). Scratch guest 9202, filled past the warning line with one test file: the version before this one sent TWO warnings for the one fill (two names); `0.352.0-rc.1` sent ONE, and cleared it silently when the file was gone. Tester 1, demo-hp and demo-felhom took the release through per-box floors: each now checks 2 filesystems where it checked 3, all containers healthy, the box walk and the https-only check green. Then the global floor, in its own save. **Correction to the sentence above („Not seen on a box"):** it is now seen on a box, as two warnings on 9202. **Not proven.** A real fill on a household's box. The critical step (95 %) on a box; a test covers it. **Not fixed here:** the guest's own root disk (`/opt/docker/stacks` lies on it) has no fill watcher — a design point in felhom.eu `16-storage-physical-identity.md`. ## v0.351.0 (2026-11-09) — old test builds and old infrastructure images are deleted (decision 56 amended, decision 285) — RELEASED (image `sha256:1d90563c…`, from `27360d6a`, CI 2603; rc.1 on Tester 1 by hand); per-box floors 12:14 UTC, global floor 2026-11-09 12:18 UTC on the operator's yes (MinAgent 0.158.1) ### 2026-11-09 **MinAgent: 0.158.1** (unchanged). **Why (measured 2026-11-09, demo-hp guest 9201, `docker system df -v`):** the whole-guest backup copies `/var/lib/docker` (mp0, `backup=1`), and demo-hp's local tier was refused for space again (felhom.eu R-548). Of 27.5 GB of images, 13.2 GB ran in containers, 9.8 GB were the apps' undo images (decision 53, kept on purpose), 2.2 GB were 28 old controller images and 1.35 GB were old infrastructure and helper images. Nothing deleted the last two kinds. **Controller images: a test build is a version (decision 56 amended).** `RetainControllerImages` left every `-rc` tag alone, so 25 test builds stayed. Tags, the running version and the previous now parse with `util.ParseVersionPre` (an rc orders below its release): an old rc goes like an old release; an rc above the running one or between the previous and the running one stays; `latest` stays. A box that RUNS a test build is judged (it was skipped). The fallback previous prefers a release over a newer test build. Tests `TestControllerRetention_OldTestBuildsGo`, `_BoxOnATestBuildIsJudged`; the existing test's rc expectation changed with the rule. **Infrastructure images (decision 285, new).** `RetainInfraImages(infra.Images())`, 3 minutes after every start: per pinned repository the pin, newer versions, the previous pin and every in-use image stay; older and untagged ones go by exact name/ID; never forced, never pruned; not while an app pulls (R-863, retried every 2 minutes, 30 tries) or while the controller swaps itself. Tests `TestInfraRetention_*` (4). Customer boxes get a new traefik or cloudflared pin every few weeks (~65–195 MB each); this is the part a customer box needs. Red-proofs: six code mutants, each test red (the previous-pin keep, the in-use check, the R-863 lock, the tag parser, the running-version parser, the release preference). Green gate: `go build`, `go vet`, `go test ./...` rc 0; `controller_gates.py --fast` OK. ## v0.350.0 (2026-11-03) — Traefik v3.7.17 (GHSA-xv5x-242w-7hf7); a restore forgets the update-held flag (R-1347); a failed put-back at start-up mails (R-1348); three false start-up lines (R-1339, R-1351, R-1323) ### 2026-11-03 **MinAgent: 0.158.1** (unchanged). **Load test, scratch 9202 only, from the LAN, 2026-11-03** (felhom.eu `audits/release-2026-11-03/part1/`): the same many short TLS connections with an `Authorization: NTLM` header to the dashboard route. v3.7.16: Traefik's anonymous memory 18.7 → 79.0 MiB after 920 connections, not released (stopped at twice its start, as the test's bound says). v3.7.17: 16.2 → 25-26 MiB, flat for 25,116 connections and the whole 300 s. **Traefik v3.7.17 (GHSA-xv5x-242w-7hf7, high, published 2026-11-02).** v3.7.14–v3.7.16: an unauthenticated client that sends a `Negotiate` or `NTLM` `Authorization` header gets a dedicated backend transport per connection, and it is not released when the connection closes: many short connections exhaust Traefik's memory, whether or not an NTLM backend exists. Traefik fronts every box. `TraefikImage` v3.7.16 → v3.7.17; a running box takes it by the reconcile at start (`TestEnsureTraefik_ANewImageRecreatesTraefik`). The upgrade notes name Go 1.27 (GODEBUG TLS settings: we set none; 500 request-header values at most, else 431) and the ingress-nginx provider (not used). Found by reading the release itself: the watch had flagged it on a word-only line (felhom.eu R-1310). **R-1347 — a restore that lifts an update hold forgets the „update held" flag.** `clearUpdateHoldAfterRestore` cleared the hold in settings, but `stacks.Manager` kept `updateHeld`: the crash-loop watcher did not watch the app and the night leg skipped it until the controller restarted. New seam `backup.SetUpdateHeldForgetter` (`stackMgr.ForgetUpdateHeld`). Tests `TestR1347_*` (the seam, the wiring, and the consequence: the same restarts trip the watcher once the flag is gone). **R-1348 — a failed put-back at start-up mails the operator.** `holdPutBackFailed` writes the D20 hold at `Recover`, before any notifier exists, so only the hold reporter's info and the hub's later reminder said anything. `AppStopGuard.SetHeldNotify` hands over what `Recover` held, once, and mails a later failure (a drive that returned) at once: the R-1340 sentence for `MixedCauseInterruptedPutBack`, from the hold record. The mail's run name says „restore" for it, not „offsite-reconstitute" (`heldRunKind`, seen on the live proof). Tests `TestR1348_*`. **Live proof 2026-11-03, scratch 9202, test image of c4b6844** (felhom.eu `audits/security-2026-11-02/part3/`): a restore cut at „the restore copy is recorded", the volume held by a created-only container, the controller started: hold `restore_mixed` / `interrupted_put_back`, the event `backup_run_failures` with the sentence in the sink. **R-1347's live proof, same image:** an update with its undo copy cut off → hold `update_failed`; the household's restore → „the update hold is CLEARED", „its update-held flag is forgotten"; six `kill -9` → the crash-loop watcher stops the app (`unhealthy_stop`), so it watches again. **R-1339 — no false „drive … not mounted".** A per-app folder under a mounted drive is never a mountpoint itself; nine false lines at every start on 9202. The folder still gets no bind and no skeleton of its own (the row's „not skipped" would write a `userdata/` inside the app's folder and recreate the file manager, so the behaviour stays and only the line changes). A drive that is really absent keeps its line, and a folder of it names the drive. `TestR1339_*`. **R-1351 — FileBrowser's two start paths at boot take one lock.** `EnsureBaseStack`'s `compose up` and `web.syncFileBrowserMounts`' `compose up --force-recreate` in the same second gave „Conflict. The container name … is already in use", two ERROR lines and a first-boot WARN for an app that then ran healthy. `stacks.Manager.LockFileBrowserStack`. `TestR1351_*`. **R-1323 — „No such container" is not „every app can still reach it".** The checked container was replaced between the check and the leave; the line now reads the CURRENT container's networks and says the truth. A current container on the app network, an unreadable one, or any other failure keeps the alarm. `TestR1323_*`. ## v0.349.0 (2026-10-29) — the household line after an operator reboot and the busy file for it (D65 = A); a hold clear records the wanted stop (R-1350) ### 2026-10-29 (evening) **MinAgent: 0.158.1** (unchanged). The new `GET /guest/operator-reboot` is asked once at start; an older agent answers 404, and such an agent cannot reboot from the hub either — nothing to show, nothing broken. The busy file is read BY the agent (≥ the reboot_guest commit); an older agent never reads it. - **D65 = A — the busy file (`internal/busymark`).** `/ctl/busy`, rewritten every 15 s while a backup, a restore, an off-site run, the night catch-up, an app update or an app export/import runs; removed when idle and at start. The agent refuses the operator's „Reboot guest" while it is fresh. Tests `TestBusyFileFollowsTheWork`, `TestBusyMarkPathIsTheOneTheAgentReads` (the path is the agent's guest path), `TestBusyMarkIsWiredAtStartup`. - **D65 = A — „Felhom support restarted your box at