[Flotapponnier]

Nova on bench 295 and the perp-cost tool, which I had said was impossible.

The correction first

I concluded Nova could not go on either because it "lists neither BTC nor ETH". That was wrong, and wrong for a reason worth recording: I had read app.nova.markets/api/markets, which is the Hyperliquid-routed frontend list. The gTrade deployment is a different thing, and its pair 0 is BTC/USD, pair 1 ETH/USD.

All five selectors the leverage source uses answer on Nova's diamond, so the reader now takes the venue and the diamond as arguments rather than hardcoding Gains' — the same shape as the volume source in #3150, instead of a second copy of 150 lines of ABI decoding.

What the two deployments actually are

Read on chain 2026-10-22, both venues, both markets:

product	bounds	threshold	Gains fee/leg	Nova fee/leg
standard pair	1.1x–200x	90/75 between 25x and 60x	0.035%	0.050%
x500 degen	500x fixed	80% flat	0.020%	0.020%

One contract priced twice: identical thresholds and bounds, a dearer standard pair. The fee sits inside the liquidation trigger, so Nova leaves strictly less room at every leverage — 8.90% against 8.93% at 10x, 0.275% against 0.305% at 200x.

The near miss

perp-cost-model.ts carries a spreadBps per row and Gains' is 1.0. Copying it would have looked entirely right beside the identical thresholds, and been wrong.

Word 7 of pairs() is the pair spread at 1e10. It reads 100000000 on Gains and 0 on Nova. So Nova is 5.0 bps a leg, not 5.5, and the "43% dearer" I quoted earlier from the taker fee alone is really 25% once the spread is counted on both sides:

• Gains: 3.5 taker + half of a 1.0 spread = 4.0 bps
• Nova: 5.0 taker + no spread = 5.0 bps

That file's header says two venues were left out rather than guessed at. This is the reason it says that, so nothing here was copied from the neighbouring row.

Nova's figures were read a day after everything else in that file, so rows can now carry their own measuredAt, shown beside the provenance tag, and the footer no longer claims one date for figures gathered on two.

Two pre-existing problems found while editing bench 295

The methodology described a method the harness no longer uses. It said the bench reads "pairs 300 and 313, the x500 cohort rather than the standard pairs which cap at 200x". It reads all four: gainsPairs is {BTC: [0, 300], ETH: [1, 313]} and the publisher picks whichever product can be opened at each leverage tab. Gains' own formula field said the same thing. Both corrected, and the bullet split in two because the parameters and the pair list are separate facts and one bullet over the length limit was the hint.

A factual ordering claim that Nova falsifies. The copy said "Papertrade leaves the most room and Gains the least". With Nova on the board that is false — Nova is the floor, on the same thresholds and a higher fee. Fixed, along with "four venues" in six other places. A count in prose is exactly what goes stale when a cohort grows.

Cache keys: all three this time

Bench 295 declares dimensions (kind and bucket), so bench-filters matters here where it did not for 041 and 116, and a provider missing from it is invisible on every tab. v94 / v87 / v36.

My own cache-key test from #3150 did not know that third key existed. It does now: defined once, environment-scoped, and never pinned to a per-bench version that has not happened. The conditional part of the rule cannot be checked statically, so the test says so in words instead of pretending otherwise.

Tests

In the harness: the two deployments must agree on leverage and disagree on fee at the standard pair, the venues and diamonds cannot collide, and the pair indices are checked against each venue's own registry rather than trusted to stay aligned — a wrong index would publish another asset's terms under this venue's name, plausibly, since pair 7 also answers with a real leverage and a real threshold. Both live tests behind OCB_LIVE=1, both run: 493 pairs each, indices verified.

In the model: identical bounds and maintenance curve, the standard pair as the only difference with Nova the dearer, Nova's spread zero and not Gains', the degen rows identical as their label claims, Nova leaving less room at every shared leverage, and every gTrade row marked chain.

Each verified against a mutation, including the copied spread and the dropped date.

Gate: validate 253 specs, typecheck, lint at its 9 pre-existing warnings, test 711 pass, go build / go vet / go test green.

After merge

perp-cohort-stats needs redeploying for the Nova rows on 295 to carry numbers. The perp-cost tool is static data and is live with the build.