docs(ups): record a pack that is fitted and not yet proven

[Gerrrt]

What changed

The documents and the UPS dashboard stop saying mjolnir has no battery, because as of 2026-08-28 it has one. They do not start saying the UPS works — they say a pack is fitted and unproven, and they name the two steps that would settle it.

Nothing about the running stack changes: no rule, expression, threshold, route or receiver is edited. The ups.rules.yaml rules are byte-for-byte the same, as that file was designed for.

Why

#93. The APCRBC115 went in today, so every file asserting an empty bay became false at that moment and had to move. But the runbook merged in #158 gates the full flip on verification, and verification has not happened — the pack was plugged in and nothing else.

That gap is not pedantry here. A management card that cannot see a pack which is present — badly seated, or faulty out of the box — emits exactly the five values it emitted over an empty bay: batteryNormal, 100% charge, 63 minutes, 48.0 V, upsAlarmsPresent = 0. A healthy-looking dashboard separates those two worlds not at all. Only a passing self-test, and charge/runtime that have moved off the pre-fit baseline, do.

The urgent part is not in this diff, because it cannot be. The UpsSelfTestFailed silence 54f1715c-e57b-4322-8a6d-5435bc8e1bd8 runs until 2026-09-20 and is now suppressing the one rule that could report a bad new pack. The runbook predicted this exact window; we are inside it. It lives in Alertmanager's state, not in the repo, so this PR names it everywhere it is relevant instead of fixing it:

curl -sS -X DELETE \
  http://localhost:9093/api/v2/silence/54f1715c-e57b-4322-8a6d-5435bc8e1bd8

Delete that before running the self-test, not after.

Detail worth flagging:

• The dashboard banner is retitled, not deleted. It now reads "Battery fitted 2026-08-28 — not yet proven" and carries the remaining steps. Keeping the panel holds the count at 84, so the check_docs.py panel-count coupling the runbook warns about is not triggered by this commit — it is deferred to the one that deletes the banner.
• The five (fabricated — no battery fitted) panel titles become (unproven — self-test pending).
• ups.rules.yaml keeps its long header rather than being trimmed. The pre-fit readings stay in it deliberately: step 1's baseline capture was never run live, so those recorded values are the baseline step 5 compares against.
• security.md keeps the finding and its lesson — "the device lied, and the rules trusted it" — and stops claiming the bay is empty. "Mains power loss" moves from not defended to partly defended, which is the honest reading of an untested pack.
• The screenshot is left alone. It is a real render of what the dashboard said on 2026-08-22, and docs/images/README.md refuses mocked-up images; both READMEs now say what it predates and when to re-shoot it.
• #110 is untouched — the shelf and the switch move did not happen, so prometheus and oracle still go deaf on a mains cut. The runbook's flip list is split three ways: done now, done after the self-test, done after the shelf.

Blast radius

Documentation and one Grafana dashboard's cosmetic fields (titles, descriptions, one text panel). No host, VLAN or service behaviour changes. Grafana re-provisions the dashboard from the file; the panel count and every PromQL expression are unchanged.

• No change to network segmentation or firewall rules
• No new port published to a VLAN that could not already reach the service
• No credential added outside secrets/*.sops.yaml

Verification

Run in this environment, which has no route to the lab:

$ python3 scripts/check_docs.py
docs OK — 35 Prometheus + 13 Loki rules, 5 dashboards, 84 panels, 6 assertions

$ python3 scripts/check_dashboards.py
5 dashboards OK (84 panels, 83 PromQL expressions)

$ yamllint stacks/observability/prometheus/rules/ups.rules.yaml
(clean)

promtool and markdownlint-cli2 are not installed here, so those two legs of make validate were not run — CI covers both. The rules file is unchanged below its header comment, and MD013 is disabled in .markdownlint-cli2.yaml.

• make validate passes — partially: check_docs, check_dashboards and yamllint pass locally; promtool and markdownlint left to CI
• Deployed to the lab and confirmed working — not done, and it is the point of the PR. Delete the silence, run one self-test, confirm upsTestResultsSummary = 1 from both the NMC and Prometheus, compare charge and runtime against the pre-fit baseline, then enable scheduled self-tests
• Docs updated (docs/, service README, or docs/roadmap.md)

---

Generated by Claude Code