Hyperliquid has published half a day since 22 September and we summed it as a whole one

[Flotapponnier]

Reported as "our stats don't match CoinMarketMan, sometimes +50% diff". They don't, and the cause is upstream.

What is wrong

Hyperliquid's per-builder fills export stops partway through each day. Trust Wallet (0x5af1…5f71), last fill in the published file:

day	fills	last fill
10 Sep	20,946	23:59:57	whole
21 Sep	33,681	23:59:54	whole
22 Sep	14,890	12:27:11	truncated
25 Sep	13,906	12:13:31	truncated
30 Sep	9,985	12:11:42	truncated
2 Oct	14,365	12:10:58	truncated

Every day since 22 September ends within three minutes of 12:11 UTC. Re-downloading gives an identical md5, so this is the published artefact, not a partial fetch.

Proof it is a subset, not a different universe

Checked against the live userFillsByTime API for one user on 2 October:

• CSV: 953 fills · API: 1,299 carrying a builder fee
• Exact match on (second, coin, px, sz): 953 matched, 0 CSV fills unmatched, 346 API fills with no CSV counterpart
• Implied fee rate 0.0950% on both sides — Trust Wallet's own rate, so the extra fills are its own and not another builder's
• By hour: 04–11 kept 100%, hour 12 kept 47%, hours 13–15 kept nothing

So our numbers were a faithful read of an upstream that had quietly halved. Fees and volume were understated together, the implied rates stayed correct, and nothing in the pipeline could tell. That also explains the shape of the report: users were only ~12% low (morning traders are all present) while volume was ~39% low, and the ratio varied per builder with their users' trading hours — which is why it flips the top two places rather than scaling everything.

What this adds

The cohort's furthest fill per day is collected inside the pass that already walks those days, so the check costs no extra parsing. One builder going quiet in the evening proves nothing; the whole cohort going quiet at the same minute is the feed being cut short.

• hl_frontend_day_coverage_hours_v2 — hours the headline day reached
• hl_frontend_day_truncated_v2 — 1 when under 23
• hl_frontend_truncated_days_window_v2 — such days inside the 30-day window

Threshold 23 h leaves an hour of slack for a genuinely quiet evening across all 104 builders, which has never happened in the mirror, while catching the widest real truncation (12.45 h).

Figures stay published. The answer to a truncated upstream is to say so, not to drop the day and let the window silently shorten.

Not an option: the node

It is running on SGP, but it has not advanced since 3 October 09:10, looping on could not read abci state: early eof and abci_stream timed out. It holds 10 of 24 hours for that day and nothing for the 4th. Resyncing still needs the requester-pays S3 snapshot.

Follow-ups this does not do

• Report it to Hyperliquid. Only they can fix the source, and it affects every consumer of that feed.
• Backfill the missing halves via userFillsByTime, seeded from the CSV user list. Proven to work above; recovers the afternoon fills of users we already see, though not users who traded only after the cut. Sizeable: ~1,400 users for one builder for one day.
• Surface the coverage on the page once the metric has data.

Verified

go build, go vet, go test ./..., gofmt clean; 3 new tests pinned to the measured days above; pnpm validate 235 specs.