[Flotapponnier]

Two debts from #2809, now that #2811's counters can measure instead of infer.

The latent miscount

headLagFirst was incremented inside the winners loop, which runs before the providers < 2 return. A trade one feed alone reported therefore handed that feed a first while never entering the race count or the share history: a numerator moving without its denominator, the same defect that published OKX at 95.93% on Solana, and it survived the first fix.

The lag recording stays inside the loop on purpose, since single-feed lags are deliberately kept and only the race is unscored. single_participant now also reaches the voided counter, which it did not before, and race_test.go gains the case.

The number I published in #2809

The bullet said roughly 1,200 trades on the Solana bench pool, read from head_lag_ref_matches_total{outcome="matched"} alone. That counter has a third outcome, ahead_of_reference, where every feed faster than the pool subscription lands on Solana.

What the new counters say directly:

	solana	bnb	base
pool observed	465/h	2,903/h	269/h
races scored	254/h	2,806/h	316/h
voided off_pool	576,530/h	14,893/h	11,707/h

So the filter does bite, hard, and Solana is back inside the 100-900 races/h it ran before the token-scoped feeds arrived (measured with offset). The 25,070/h I reported right after deploying #2809 was a rate() window straddling the container restart: the artefact I warned about in that same PR and then read anyway.

The shares now agree with the latency column

Which is the check that found the bug in the first place:

chain
base	mobula 71.37%	serialized 23.50%	okx 5.13%
bnb	mobula 60.54%	serialized 35.70%	okx 0.29%
solana	serialized 81.03%	mobula 54.61%	okx 0.00%

Solana p50 for the same window: serialized 3 ms, mobula 8 ms, codex 98 ms, okx 249 ms, birdeye 678 ms, geckoterminal 8.1 s. Shares exceed 100% because a tie within 5 ms counts as first for each feed involved, which is documented.

Shares are computed over an in-memory 24 h window that reset at the redeploy, so they will firm up over the next day.

Gates

go build ./..., the full go test ./cmd/script/ suite, pnpm validate (235 specs valid). The corrected bullet went over the 500 character cap on the first attempt and was shortened.

🤖 Generated with Claude Code