c0f59769

By: Elijah Cavan <eli_cavan@live.ca>

oneloop benchmarks: cross-check masters against oneloopmaster

Adds a second, independent master evaluator alongside avh_olo. Every numeric
validation in this crate ultimately traced back to avh_olo; this breaks that
single point of trust.

1593/1593 masters agree on the full Laurent series -- A0 4.9e-17, B0 5.1e-16,
C0 6.3e-14, D0 1.6e-14. Two conventions had to be reconciled and both are
verified by executed code, not asserted: the coefficient tuple order is
reversed between the two engines, and oneloopmaster takes mu SQUARED where
avh_olo's setter takes the scale (probed at mu^2 = 4 and 9; the wrong call
reproduces the square bit-for-bit).

Degrades gracefully -- with the module absent the harness runs exactly as
before, verified by blocking the import from outside the file.

Also fixes a pre-existing hole this would otherwise have inherited: the
harness printed FAIL and exited 0, so a master disagreement could not fail a
run. Now exits non-zero, fault-tested across seven scenarios including a
forced disagreement and a stubbed-out engine.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lgs5exR1ZjgEDpd9w3xVqs

Total

1h15m37s

Timeline