backlog: record #1211's second excursion -- the margin is widening, not repeating

[wshallwshall]

Docs-only. #1211 stays OPEN. No new number -- this item already exists for this exact metric, and a second instance belongs on it rather than beside it.

#1211's own text says a recurrence gets evidence, not another re-run. This is that evidence.

Two excursions, hours apart, on unrelated changes

PR #343   fixed_per_conn@N=24: 36    < prior 48.4 * 0.75 (= 36.30)   short by 0.30
PR #355   fixed_per_conn@N=24: 34.9  < prior 48.2 * 0.75 (= 36.15)   short by 1.25

Both windows-2025, both test_connscale_smoke.py::test_connscale_smoke_end_to_end, both the empty_claims_monotonic band.

Attribution is ruled out on blast radius, not on the word "flake"

#343 touched scripts/hooks/worktree_gate.ps1 and three gate test files. #355 touched a docs guard, its tests, ci.yml, and four lines of CLAUDE.md. Neither has any overlap with connscale, the store, or the pipeline. Neither change could reach the subsystem that failed.

The new information is the trend, not the repeat

The miss went 0.30 -> 1.25 against a prior that barely moved (48.4 -> 48.2). One excursion is a tight band; two excursions drifting further out is the ratio behaving differently, which is what this item was left open to catch.

And it is not the clock

The suite grew 11,961 -> 12,025 passing tests between the runs, and the failing step took 40:46 against a 55:00 cap -- the leg had headroom and was not being killed. Runner slowness would surface as a cap kill, not as an SLO miss.

I did not re-run it to get a green. A green re-run here is consistent with the diagnosis rather than evidence against it, and banking one would have destroyed the second data point that makes the trend visible.