Add post-freeze abstention stress for prospective designs

[FlorianPfaff]

Scope

This is a separate, post-freeze sensitivity benchmark stacked on the corrected profiled-Gaussian design evidence. It does not alter or replace the immutable five-seed matched-generator baseline.

It adds:

• matched pure-generator recovery at the fixed N=60 benchmark allocation, with Wilson intervals;
• disjoint deterministic threshold-calibration, calibration-audit, and evaluation streams, all hash-bound in the artifact;
• a familywise null threshold over the best of six pure labels, a separate winner/runner ambiguity threshold, and an all-six ridge adequacy gate whose penalty is selected inside training data only;
• null evaluation, all 15 equal 50/50 in-span mixtures, and one explicitly bounded full-grid-orthogonalized tanh(surprise) residual probe;
• explicit abstention and false-confident-pure-call rates rather than forced closed-set classification;
• checksum-verified ingestion of the chronologically earlier heuristic N=60 allocation used by the accepted five-seed benchmark;
• clean-Git enforcement, a canonical configuration digest, payload hashes, CLI, tests, and scientific-scope documentation.

The primary allocation is locked by SHA-256 a823be49faf6c6cbebf60b11d4b5ca895cf7734d6e9c577ee98f97a5907b69b2 and source commit 1b2028929ac6ebc1cce0882f0c22af9918044342. It was fixed before either the exact certificate or stress results. Certified N=45/N=60 allocations are not substituted after observing their downstream recovery.

The mixture and nonlinear probes are diagnostics, not an open-set robustness certificate. The simulation remains conditional on effectively independent Gaussian observations, the fixed finite grid, globally standardized candidates, and independently instantiated cells. It is not an executable ordered-history or animal protocol. N=60 is a benchmark count, while the analytic N_eff index remains a population observation-equivalent quantity.

Verification state

At head 0a0a5fba03ca47c04963a636b9041fd8c5f1df08, the sole prior CI issue was corrected import ordering; the new CI and an independent server gate are running. No claim-bearing stress artifact has yet been committed.