From 7c9419ea2f7903989b6359380f04244120027b5e Mon Sep 17 00:00:00 2001 From: wshallwshall Date: Wed, 12 Aug 2026 19:49:08 -0500 Subject: [PATCH] backlog: record #1211's second excursion -- the margin is widening, not repeating #1211 said a recurrence gets evidence rather than another re-run. This is the evidence. Two excursions on windows-2025, same test, same SLO, 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 NEITHER CHANGE COULD REACH THE SUBSYSTEM. #343 was the worktree gate and its tests; #355 was a docs guard, ci.yml and four CLAUDE.md lines. Zero overlap with connscale, the store or the pipeline in either case, so attribution to the diff is ruled out on blast radius rather than on "flake". 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 the band being tight; two excursions drifting further out is the ratio behaving differently, which is exactly what this item was left open to catch. AND IT IS NOT THE CLOCK: the suite grew 11,961 -> 12,025 passing between the runs, and the failing step took 40:46 against a 55:00 cap, so the leg had headroom and was not being killed. Slowness would have shown up as a cap kill, not as an SLO miss. I did NOT re-run to make it green, and I did not file a new number -- this item already exists for this exact metric, and a second instance belongs on it. --- docs/BACKLOG.md | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/docs/BACKLOG.md b/docs/BACKLOG.md index 72254ae0..2797d757 100644 --- a/docs/BACKLOG.md +++ b/docs/BACKLOG.md @@ -7802,6 +7802,19 @@ only to dismiss it as a control for time dilation - correct about dilation, sile CPU does"; an independent verification pass sharpened it to the above and corrected the remedy - the gate is the wrong shape, validation of the walk is the right one. ## 1211. `empty_claims_per_msg` is not contention-immune: the ratio form excursions past its own SLO band on a hosted runner +> +> ⚠️ **SECOND AND THIRD OCCURRENCES RECORDED 2026-08-13 -- the item said a recurrence gets evidence rather than another re-run, so here it is.** Both on `windows-2025`, both the same test and the same SLO, 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 +> ``` +> +> **Neither change could reach this subsystem.** #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`. **Zero overlap with connscale, the store, or the pipeline in either case.** +> +> **THE MARGIN IS SHRINKING, AND THAT IS THE NEW INFORMATION.** The first excursion missed the band by **0.30**; the second by **1.25**, against a `prior` that barely moved (48.4 -> 48.2). A ratio drifting further outside its band across independent runs is a different signal from a single excursion, and it is what this item was left open to catch. +> +> **Also measured, and it argues against "the runner got slower":** the suite grew from **11,961** to **12,025** passing tests between the two runs, and the second failure took **40:46** against a 55:00 cap -- so the leg had headroom and was not being killed. The failure is the SLO, not the clock. > 🔢 **Filed 2026-08-10 -- measured, with the confirming re-run recorded before it ran.** Value **5/10** · Difficulty **3/10** · _fill-in_. `_empty_claims_per_msg` (`harness/load/connscale/runner.py`) is `Δempty_claims / Δread` over the same first-to-last window, and its docstring states the design claim in terms: *"both inputs are Δ/span over the SAME first->last in-hold samples, so span cancels ... that is why it survives runner contention: slowing the run scales numerator and denominator identically."* **That claim is measured FALSE on a GitHub hosted runner.** On PR #311 the `empty_claims_monotonic` SLO failed on windows-2025 at `fixed_per_conn@N=24: 33.7 < prior 46.8 * 0.75` -- an excursion of roughly 28% against a +/-25% band -- and a re-run of the IDENTICAL commit went green.