[Karanvir1729]

Four hunts on 2026-11-04 ran genclass_explore next to their own page checks. On fixed builds where the page was right it said RACE in 85 of 120 reordered runs, and it missed real bugs. Two independent reviews of this branch then found rules that hid real races main reports, and one that reported from a single sample. All are fixed here; the numbers are after those fixes.

• An age ("13 minutes ago", "an hour ago", "just now", "10m ago") is one line when the later load is older by about the time between the two, plus a minute. "26 minutes ago" then "27" is one line; "just now" against "3 days ago", and "12 minutes ago" then "2" ten minutes on, are not. A clock, "today", "in 2 days" and a bare "3d" are values a page can get wrong, compared as they are unless the baselines show them varying.
• A POST one side never sent stays a race, however it was sent. Only a different non-zero count, with the page the same, needs evidence: two identical requests in flight, or a user action sending it directly a different number of times. No URL lists.
• A field name some baseline lacks compares with itself by shape (el-id-*-32, *-input). A name every baseline has is a name, whatever row beside it is new, so qty-100 and qty-200 swapping values is a race.
• With slow_ms the plan adds a baseline with no added delay (runs + 3 loads). What both slow baselines show the same and it does not is reported as what a slow response alone does, and says "on one load's word" when no other load sides with it. What one slow baseline alone does stays noise, as on main, and is named under "worth a look", as is a line only that load shows.
• Noise blanks only the tokens that varied. "version-1" against "version-2" stays a difference, and another number on that line may drift a little and no further.
• The report orders page content first, and the headline names that line. A report that compared nothing says NO VERDICT, never beside a race finding.
• Routes group by shape, and an action only some loads did no longer splits a session.
• Safe requests to other sites leave the counts. Writes stay: from localhost a hosted API of the page's own looks like anyone's. A bare count is internal only when it stands for a store field one such response alone wrote, with no overwrite on it, so a stale cart count from a hosted API stays a race. They still join the schedule, which needs the runtime.
• Matching redactions no longer backtracks: a page could hang the bridge for over 15 s with one value.
• The review's inputs are tests, and bench/explore/planted.mjs has pages for wrong values that read like a time and for a lost POST.

Measured (RESULTS section 27):

• Fixed builds, rescored from stored reports: false reports 85 of 120 to 0 of 120. 49 of the 85 had capped lists (Open WebUI 43 of 43), so listed evidence clears 36.
• Sessions whose page check was right: 211 of 466 runs to 24 of 466, over 94 sessions, 7 of which compared nothing.
• Sessions with a wrong run that get a RACE verdict: 35 of 37 to 25 of 37. Nine are OpenCode verdicts that rested on a read-only POST count, the same on the fixed build, and one needs the new load.
• OpenCode 2.2.5, recorded runs (slow_ms 600): file-refresh RACE as a slow response's where main said 5 of 6 on counts; switch and palette no race where main said 3, 6, 6 and 5 of 6; new-session still RACE, possibly false.
• Planted pages, false runs on correct variants: 49 of 360 to 18 of 396 (a save queue's PUT count and clocks moving, both false on main too).
• Planted bugs given a verdict: 38 of 54 to 48 of 54, with 28 false runs to 0.
• The first review's inputs: the 17 wrong values that read like a time are reported in 17 of 17, as main, up from 1 of 17 on this branch's first version; hidden.mjs reports all of them in Chromium, and the lost POST.
• The second review's inputs: the stale write from a hosted API and the swap beside a new row are reported, 7 of 7, up from 1 of 7. What stays unreported is listed in RESULTS: ages that can have moved, and names made for each load swapping values.
• RTK and TanStack: 32 of 32 found, 0 of 36 false, unchanged.

Cost:

• A duplicate POST behind another request or a timer, with no overlap and the page unchanged, is internal: dbl-timer 2 of 3 sessions to 0 of 3, dbl-chain 0 of 3 both ways, all 6 named under "worth a look". "No race found" says when a POST count differed, that a duplicate write may have happened, and to check the server. Of the stored reports' 113 POST lines, 106 stay demoted, which is what moved the nine OpenCode verdicts above, and 7 stay visible, which moved no verdict.
• Clocks and bare forms are compared as values: 5 more false runs on correct stored sessions (note-autosave, "Today at 3:27 PM"), 3 on a session with a real bug.
• A bug that shows in only one slow baseline is noise, as on main, and named under "worth a look": OpenCode's refresh when its baselines split, Langflow's New Flow at 600 ms.

Limits: the no-delay baseline has run on one real app only (OpenCode). Both slow baselines landing on one value of a random line can read as a slow response's race (tip: 1 of 3 sessions). Field names made for each load pool their values, so two of them swapping is not seen. The rescoring reads text and has no send steps, so "sent directly" cannot fire in it. No earlier row measured a time-shaped wrong value or a lost POST.

Skill text changes (--repeat-each 9, "worth a look"), so the plugin re-syncs. No runtime change, and no model-visible text changed.

🤖 Generated with Claude Code