Reject two T740 axis-search candidates on exact operator replay

[Gonzalo-Mier]

Result

This PR now records exact operator replays for two frozen T740 axis-search candidates. Our global CP-SAT rank-only model gives a hypothetical κ of about 0.001734807085759196… under PR #3200's verifier, but its owner-0 / port-0 operator differs at coordinate character 2^53.

A fresh overlap refresh found the higher open conditional headline in PR #3251: 0.001734815770385098563484. We pinned its BINS, MOVES and RESULT artifacts to head d7d6254416f0b3ffe659079cca42fd152df63a36. The inherited PR #3200 numerical verifier accepts its profile and root, but the source-bound exact eight-stream replay finds a mismatch for its owner-0 / port-0 moves at character 2^52. This one group is enough to reject that frozen allocation as operator-preserving. Its headline is about 8.6846259e-9 above our hypothetical; neither candidate supplies a new admissible κ here.

The exact replay reports three changed Gaussian-integer matrix entries, and the registered package reproduces both witnesses. The selected record remains unchanged.

Scope and overlap

The global search spans 2,328 (port, child-usage) patterns and 82,880 groups; its false positive uses partially hosted donors (children 14 and 15). Related work in #3199, #3204 and #3220 is acknowledged. This PR adds candidate-specific exact counterexamples, not a proof that every exchange or repair fails, a full audit of either supplier stack, or a selected-frontier change.

Reproduction

python3 scripts/research_package.py --package research/t740-joint-axis-optimization --check-only
python3.12 -B scripts/research_package.py --package research/t740-joint-axis-optimization
python3 scripts/check_research_commit.py --package research/t740-joint-axis-optimization --ref HEAD

The registered replay authenticates both source stacks, verifies both conditional rank/profile results with the inherited PR #3200 auditor, positive-controls the exact baseline, reproduces the mismatches at 2^53 and 2^52, and rejects two mutation controls. The optional OR-Tools discovery run is FEASIBLE, not OPTIMAL, and its parallel incumbent is nondeterministic.