Test night provenance (night 2026-10-22 → 23, Part 4)
- VM 9344 `night1023-soak` on demo-hp (4 vCPU, 8 GiB, 256 GB raw system disk + 200 GB raw data disk on nvme-scratch,
  not the thin pool), created 18:46:17 UTC; installed from the public ISO felhom-installer-1.30.0-pve9.2-1.iso
  (Install pressed 18:51:58 UTC; the ISO fetched the current installer at run time).
- Customer Tester4 (the kept fresh-install customer; no old host at the start — read from the hub DB). Bound 19:0x UTC
  with the bind link minted 2026-10-22 07:37 UTC (unconsumed until tonight). Host enrolled as Tester4-16fdef (19:13 UTC);
  guest 9201 (tester4), 192.168.0.151; host 192.168.0.128. Golden controller 0.339.0 → the box's first controller.
- Root: the day-0 install rotated the root password into the hub vault; read through the operator's break-glass path
  (stored in sec as NIGHT_2026_10_23_VAULTED_ROOTPW, never printed); DooPlex's public key added to root.
- Fixtures (by hand, release candidates, not releases):
  controller 0.340.0-rc.6 (= felhom-controller 943c517) 19:30:44 UTC → rc.8 (= 1babfe9) 22:03:31 UTC;
  agent 0.163.0-night.1 (= felhom-agent ae69ad4, D30) 19:30:07 UTC → 0.162.99 (= b9b831f, D30 hardened) 21:51:40 UTC
  (the first name is a pre-release string the controller cannot compare — the off-site unlock refused; see 07a);
  the D4 drop-in by hand (the bundle's own steps: sshd -t -f, install, sshd -t, try-reload-or-restart, sshd -T).
- Drive: the 200 GB disk initialised through the product as `hdd_1` (the second-drive target); the apps on `belso`.
- Apps (the rehearsal's 8, install API): nextcloud, immich, paperless-ngx, vaultwarden, mealie, actualbudget, radicale,
  bookstack — all running (05-installs.txt).
- D30 live proof ran on this box BEFORE the uploads (19:45–20:16 UTC; P2-D30/).
- Household data: gen_household.py with SEED 20261023 (3414 files, 3.5 GB, MANIFEST in the loader CT 9345).
- Load: soakrun.sh 120 on CT 9345 from 20:04:16 UTC (load.py + load2.py cycles); the Nextcloud tree stream restarted
  20:09:21 UTC with a hosts entry. A first attempt at 19:55:59 UTC was VOID (the tool used the dashboard cookie for the
  setup gate; the gate now hands out its own cookie) — stopped after 2 min, logs kept in the loader's /root/void-1955/.
- Off-site: the box held for the household's recovery code (offsite.state=awaiting_recovery_key); unlocked 21:52 UTC
  with the household's recovery code (07-unlock.txt).
- Recorders: rec.py on the box host (10 s) from 19:47 UTC; outside.sh on DooPlex (10 s /api/health + qemu I/O) from
  20:09 UTC.
- LOAD CORRECTION (found 22:14 UTC): the VOID attempt's load2.py (pid 4514, „load2.py 120") was NOT stopped at 19:58 —
  the `pkill -f load2.py` ran inside a `bash -c` whose own command line held the pattern and killed its own shell first
  (the known pgrep-matches-itself trap). It ran from 19:55:59 to 22:14 UTC beside the planned load: 406 Paperless cycles
  (upload / delete / empty trash), Immich and Nextcloud cycles, with the same household credentials. So the box carried
  MORE than the planned load for ~2 h. Stopped 22:14 UTC (kill by pid), with its orphaned rclone (pid 51651). Its log:
  the loader's /root/void-1955/proof.log (3908 DONE lines).
