zeekayandHanzo Dev 904874cce2 consensus: derive the export signer floor from the set, and stop configuring it ...

QuorumCertEngine took an alpha and used it as the Quasar distinct-voter floor.
That made the number of independent parties export finality reports a value the
operator picks, and at alpha = 1 it reports one: a validator holding two thirds
of the stake could mint an export-grade certificate on a single signature. A
security parameter chosen by the party it constrains is not a floor.

signer_floor(Quasar) is now two_thirds_count(n) = floor(2n/3)+1 — the same
supermajority the stake floor asks for, read in seats, derived from
two_thirds_stake_floor over unit weights so the two cannot drift. Go's
config.TwoThirdsCount and Rust's two_thirds_count are the same number, and all
three now decide the whale case the same way.

Quasar was alpha's only use, so the constructor parameter is gone and Party no
longer carries one. Both tiers' floors come from the validator set and nothing
else. alpha survives only as the wave's per-round sampling ratio in
alpha_threshold(k, alpha) — a different quantity that shares a letter.

equal_stake_supermajority is renamed two_thirds_count for the same reason the
floor exists: the value is the two-thirds count of n at every stake
distribution, and "equal stake" named one of its two uses rather than the value.

The conformance harness used to build the engine at alpha = 1, reasoning that
the smallest legal value could never bind, so C++ would ask exactly Go's
question. It was right about Go and wrong about what a corpus is for: Go's
export rung had no count floor, C++ had whichever one the deployment
configured, and a harness at alpha = 1 could not see the difference — so the
corpus reported agreement between two implementations that decided the whale
case differently on every real deployment. It now builds the engine from the set
and checks the floor the engine derives against the one Go recorded.

decision.json's signer_floor is the RUNG's, so a Quasar row no longer reports
Nova's, and carries the two new whale-at-export rows.

The skewed row in toy_stub_killer isolates the stake gate with four light
voters, not three: three no longer clears the count floor on five seats, so the
row would have stopped isolating anything.

Co-authored-by: Hanzo Dev <[email protected]>

2026-09-01 17:40:17 -07:00

..

bench_trilang.cpp

consensus: derive the export signer floor from the set, and stop configuring it

2026-09-01 17:40:17 -07:00