Larger-budget continuation of the disagreeing generated pair

The completed local-probe matrix passed all 24 individual fits but only eleven of twelve paired-start comparisons. Preserve that failed qualification. The remaining pair is zero-field globe / nominal intervals / height-only scoring. Its maximum prediction difference is about 1.3 micrometres, but its objective difference 5.56e-8 exceeds the frozen 1e-8 repeatability check.

Continue both saved endpoints under the exact same generated observations, six parameter bounds, native arithmetic, curvature-corrected algorithm and local numerical gate. These are two new continuation identities, not retries or relabeling of completed identities. No tolerance or physical model changes.

Freeze both saved parameter vectors, parent manifest/summary/completion, source/environment and allowance before launch. Allow two starts, each at most 600 tangent calls including fresh checks, plus two diagnostic predictions: 1,204 native generated calls maximum. Use four total numerical threads, a thirty-minute ceiling and no retries. No originals or observed fits.

Require both continuations to pass the same fresh individual checks and their paired cost/prediction criteria. Preserve all endpoint arrays and traces. If they pass, a separate saved-only review may combine these new results with the 22 unchanged passing companions and verify all twelve pairs. It must keep the earlier failed gate visible. Only that complete review can release the prepared full-calibration stage; the failed matrix alone cannot.

If the pair still disagrees, inspect saved traces and derivatives before any new candidate. Do not increase an active allowance, reuse unused calls, relax the paired criterion or advance to recorded-data fitting.