felhom.eu/REPORT.md

REPORT — the operator's ten answers built (2026-10-06, rulings 09 §3 139–148)

Part	Result
A — the box changes (1, 4, 5)	done — R-444 weekly trim (measured on demo-hp first, then built, delivered, seen working by itself), R-645 version skip, R-856 crash-boot grace; agent v0.149.0 + bundle, controller v0.300.0 + golden 0.300.0, hub v0.139.0, all on demo-hp, demo-felhom, Tester 1
B — the backup leftovers (2)	done — runbook written; read-only listing of ep0: no leftover exists, nothing deleted, real counts unchanged; the box's warning names the runbook
C — the page sentences (6, 10)	done — mealie and Karakeep, hu + en, live in the catalog
D — the test tools (3, 7, 8, 9)	done — R-618 closed by ruling; R-734 marker list; R-624 bench seed proven on a recreated bench; R-502 full-run gate, first real run green with its decoy convicted

Rows before	Rows after	Opened	Closed
150	142	2 (R-890, R-891)	10

Closed: R-444, R-99, R-618, R-645, R-856, R-747, R-734, R-624, R-502, R-774.

Baselines and rulings

Verified 11:07: felhom.eu 8f40b3ce26 (the operator's own „cleaned reports" commit on top of fe998b8), controller 13bda270c3 (v0.299.0), agent 37e98f452b (v0.148.0), catalog d1a148408f, register 150. The ten rulings were recorded first (09 §3 139–148, 8c65ff0c), with the note that CC's own pick differed on 2, 8 and 9.

The operator's cleanup removed every REPORT*.md. Two felhom.eu checks read REPORT.md: the decoy test now plants a missing file and removes it again (8c65ff0c), and this file is the session's own report, as the repo rule says.

Part A

• R-444, measured first (demo-hp, 09:14Z, audits/ten-answers-2026-10-06/r444-measure.txt): pct fstrim 9201 rc 0 in 24.4 s; thin pool 65.53 % → 33.40 %; 18/18 app probes 200, slowest 1.1 s. Built (agent ee71abd): weekly, due Wednesday from 10:00 host-local, starts only 10:00–20:59, under the one-heavy-op gate (backup, restore-test, OS steps), 3 tries a week, persisted, reported as guest_disk_trim; ONE sudo rule /usr/sbin/pct ^fstrim [0-9]+$. Hub (a411cde7): System page „Last disk trim". Live: sudo -l on demo-hp and demo-felhom allows the trim and refuses --ignore-mountpoints, ;x, a second vmid and pct destroy. The job's first catch-up try ran 2 minutes after the agent update and failed (the bundle had not landed — expected); the hourly retry at 12:56 local logged fstrim: guest 9201 trimmed 2.1 GiB in 2.4s and the hub page read „10 min ago · 2.1 GiB" (r444-live.txt).
• R-645 (controller 2d63714): every night leg skips an app whose pin is not what it runs; one amber line on the backups page. Red-proved on the hand-lift shape. Residual, stated: once the boot reconciler starts the app on the new version, pin = running again — a design question, not built.
• R-856 (controller c393d85 + agent route GET /host/crash-guard): after a crash boot, app mails wait 15 min. Live through the real route: demo-hp logged the normal 90 s because its last crash boot (2026-10-05) was not this start; demo-felhom logged a clean boot. The crash branch was not shown live (no crash allowed); tests cover it.

Part B — ep0, read only

Every snapshot of datastore felhom-offsite, both from the server's API and from the directories: 9 snapshots in 5 namespaces, all ≥ 369,808,250 B, all verification ok, every directory with its manifest. No phantom; nothing deleted. Runbook runbooks/pbs-phantom-cleanup.md (the list command, the two-part test, one api delete per proven phantom, the before/after count control) + runbooks/pbs-phantom-list.py; the agent's WARN now names the runbook (be398f9).

Part C and D

• R-747, R-774 (catalog ec72c9d, live 1938921): one first-step sentence each, hu + en, freeze updated.
• R-734 (catalog b0939cf): the update test ignores listed marker files (immich's six .immich), each with a reason, only when changed/added and ≤ 64 bytes; an unlisted file still counts (red-proved).
• R-624 (catalog aed80ee): bench-only vaultwarden seed through its admin invite. The bench LXC 9401 no longer existed; recreated by its recorded recipe (stopped again afterwards). Proven: admin sign-in, invite and registration 200; the seed read back before and after; .env shredded, secret greps 0 with a working control; without the run flag the admin route is not tried (inconclusive). The move used to drive it (1.36.0-alpine → 1.36.0) failed its health check — the non-alpine image fails the alpine health check; that is the chosen target, not the seed. New question: R-890.
• R-502 (felhom.eu 9d39faab): iso-bootstrap, full runs only, NOT CHECKED without docker or the image; the image felhom-iso-assistant:trixie was built on DooPlex; first real run: 73 checks green, the built-in decoy convicted.
• R-618: closed by the ruling; nothing to build.

Said plainly

• CI run 1423 (felhom.eu, the hub-manifest commit) was red on the golden gate: it read controller v0.300.0 a few minutes before the golden 0.300.0 record was committed. Run 1425 on the next push was green.
• The agent update's first trim attempt failed by design (the sudo rule rides the bundle, which came 10 minutes later). On a box that gets the agent but not the bundle, the trim warns weekly and its capability reads degraded.
• R-891: felhom.eu CLAUDE.md still says --fast runs every gate — untrue since iso-bootstrap. An instruction file; left for the operator.

CI, last commit of every repo

felhom.eu: this commit (checked by its commit after the push); controller 36088fd → 1426 success; agent cefdc73 → 1427 success; catalog 65130c6 → 1428 success.

Teardown

Drill VM: build guest destroyed, secrets shredded, off, virgin. Bench 9401: stopped (kept; its evidence copied off). Scratch 9202: not used. ep0: read only. demo-hp / demo-felhom: the deliveries, one measured trim, sudo -l checks. Hub: the deploy, the vouches, three floors per release. Scratch secrets shredded at the end.