# REPORT — release run 2026-11-03 (attended): the Traefik fix first, then the golden, agent, hub, catalog

## Five lines

1. **Every box runs controller 0.350.0 with Traefik v3.7.17:** yes. The global floor at 08:18:16 CET; the three boxes moved at 08:05 CET on per-box floors; 9202 by hand.
2. **The load test (9202 only, from the LAN):** v3.7.16 grew from 18.7 to 79.0 MiB after 920 connections and did not release it (stopped at 11 s by the bound). v3.7.17 stayed near 25 MiB for 25,116 connections and the whole 300 s. Proven by behaviour.
3. **Golden 0.350.0 installed fresh and vouched:** yes, 10:18:10 CET. The install was the closest path, not the hub's: the public installer with one added line (the hub refuses an unvouched golden, R-297).
4. **Agent 0.172.0 vouched:** yes, same save. **Hub 0.179.0 live:** yes, 09:05 CET.
5. **Catalog pushed:** yes, `f4c5c79`; Tester 1 read it.

## Versions and digests

| What | Version | Digest / sha256 | Built from |
|---|---|---|---|
| Controller | 0.350.0 (rc.1 before it) | `9ee53908e6d95f0340c41bf4eaba954280084a8e6927de59779440e3551df261` (rc.1 `cbf4d42f…`) | `5c336d82` (rc.1 `13e34d4`) |
| Traefik | v3.7.17 | image id `a72cb2f4a2109f1ae52261c86d56b7d33a27dc5c4e211ce3c7835b772338b94f` | |
| Golden | 0.350.0 | `082e2e040de5615300da3ec8bc48fe6761ee0e6375dcc732e5cba82086e5e06b` | `build-golden.sh` v3.2.0 |
| Agent | 0.172.0 | `2c0222cd01713f65c627d6d17c18b4a8b2b1e0c39881db0df7ccb17de48e24c7`, bundle `cc34b0bc…` | `ecbb0cc` |
| Hub | 0.179.0 | `9c52c1c3427dcc493e007ec94d021802eab7dd267ea2a199639ff047b2767011` | `57d46201` |

Evidence: `documentation/audits/release-2026-11-03/` (each part has a README), `documentation/tests/golden-0.350.0-2026-11-03/`.

## Said plainly

- **My mistakes.** The first fresh install did not use the test installer: I set the variable in a place the bootstrap blanks, and I read the result only at the end. Then I told the operator a 45-minute wait would open the reinstall; it does not (R-1159), and 12 retries got the same refusal. It cost one operator act (the host delete) and about 90 minutes.
- **Scripted on purpose:** the uninstall's typed confirmation went through `ssh -tt`, on my own test VM only.
- **Proof limits.** R-1351, R-1323 and R-1352 have test proof only: their conditions did not fire. R-1347 and R-1348 were proven on the test image of the same code; the release build ran a plain restore. The file manager sign-in was not driven on the N100, which is lived-in; its anonymous page answered 200. R-1302 was staged: a plain first start on the appliance path reaches neither line.
- **ep0.** I changed nothing there by hand. The fresh installs made the hub re-issue Tester4's PBS token on ep0, twice: the product's own cascade on a customer's new host.

## Observations

- FILED (appended): R-548 — demo-hp's local backup tier is refused since 10-30, the disk has too little room.
- FILED (appended): R-1361 — the night walk's lost `db_dump_completed` became a false `expected_dbdump_missed` mail on Tester 1.
- NOT-A-FINDING: `Tester4: passphrase_keys_refused`, `host_stale`, `node_stale`, `offsite_delivery_stuck` mails — CC's own removal of a test install and its retries; the guard was right.
- NOT-A-FINDING: Traefik's repo digest on 9202 differs from the other boxes — the image id is the same on all; 9202 pulled the tag on the night of 11-02, the others this morning. The tag's index was not read.
- NOT-A-FINDING: `FELHOM_INSTALL_URL` in a unit drop-in is ignored — the bootstrap reads it only from its env file, by design.
- NOT-A-FINDING: ArgoCD's `Ingress/contact-mailer` OutOfSync — the live object matches git field by field (the same reading as 2026-10-30).
- NOT-A-FINDING: the slow-statement WARN lines on the hub — R-1353's class, already a row.

## Teardown

VM 9383 destroyed with its disks; drill VM off, back on `virgin`; bench 9401 stopped; 9202 on 0.350.0, its `radicale` hold left for the operator; per-box floors cleared; DooPlex's clones and scratch secrets removed. **Hub:** host `Tester4-e286d3` waits for the operator, from about 11:00 CET.