[Flotapponnier]

I took this bench to draft last night (#3522) because its same-block tab published a 0 on every row, and 0 is the best score here, so five named companies carried an unearned perfect record. I read the worker's bench skip: missing p50,p90,p99 against queries that answered when asked directly, and concluded the skip was in the publish path.

It was not. increase() over a window reads 0 on a counter that has not incremented twice inside it, and the counters were hours old. The delayed tab crossed that line first and looked correct; the same-block tab had not and looked broken. One cause, two appearances, and I diagnosed the appearance.

Fifteen hours later both tabs resolve through the real loader, with every row carrying ~106 token-questions against the 120 this bench calls a day of them (healthy, not the 7 of last night):

[delayed]    trnch 41.5  lifi 44.3  paraswap 61.0  sushi 92.5  kyberswap 95.8   n≈106
[same_block] trnch 41.9  lifi 44.2  paraswap 61.6  sushi 93.3  kyberswap 95.2   n≈105

Trnch first and LI.FI second is the result the method predicted and the only one worth publishing: both screen the token and decline — one with its own agent before polling any router, one with a deny list. A quote board cannot see that difference; this one is built to.

The two tiers agreeing within 1.5 points is also the expected shape, and it says the anti-bot cooldown question the selector exists for is not what separates these routers.

Nothing changes but the status and the keys. The queries, the sample gating and the counters are as merged yesterday in #3517 / #3519; what changed is that there is now enough measurement to stand behind.

Keys bench-unfiltered-v165, all-benchmarks-v158. Provider set unchanged.

Gates: pnpm validate 261 specs, tsc --noEmit clean, 910 tests / 0 fail.

🤖 Generated with Claude Code