docs(connscale): retract superseded figures in the CI log, and pin the coverage claim that a disarm would falsify

[wshallwshall]

Authored by the PM seat, Dispatcher-routed. Pushed and opened by the LANDER. Doc-only, two files, one line each.

1. CI-FAILURE-LOG.md:70 — figures retracted in place

The row carried mean 1.05 / sd 0.243 and "the band is the defect". BACKLOG #1211 withdrew those the same day they were written — the truncation estimate inverted the censoring under a single-normal assumption, and #1211 states in terms that the model is misspecified and its numbers should not be used.

Struck through and annotated rather than deleted, because that file's own rule at line 52 is "correct it in place and say so on the row".

It deliberately does not restate #1211's replacement model. That model is itself under active review — a reviewing seat has found the bimodal reading may be an OS split rather than two modes. A log carrying a second copy of a live figure is exactly how the first one outlived its own retraction, so the row points at #1211 and stops.

The direction conclusion survives; the mechanism and every number did not.

2. 17-performance-and-scale.md:76 — not corrected, because it is not wrong

The row names the empty-claim monotonicity assertion as covered. That is true today. It becomes false only if the leg is disarmed.

So it gains a coverage pin: any disarm, skip or delete must edit this row in the same change.

the disarm looks complete while the false claim survives here

The failure it guards is silent. A disarm that satisfies its own reviewer leaves the test plan asserting coverage nothing provides, and nothing links the two. Wording carried verbatim from the Dispatcher's request, which asked that it survive into the dispatch text rather than be discovered afterwards.

Verification

• 1 ahead / 0 behind origin/main 302e2ea8f; two files, 1+/1- each.
• Author's gates: ledger gate, PHI leak guard, control-characters, secrets, username-as-access-key all passed; ruff/bandit skipped (no .py).
• LANDER read both diffs before landing: the strikethrough, the retraction annotation and the #1211 pointer are present, and the row does not restate the withdrawn figures.
• The branch existed only on a private rescue tag — no origin ref — with a positive control confirming the instrument reads remote refs.

No test, band, or #1211 figure is touched. The statistics remain the reviewing seat's, attributed and un-re-derived.