Correct maximin profiled Gaussian score diagnostics

[FlorianPfaff]

Scope

Corrects the v0.7 maximin planning diagnostic to match the candidate-specific residual variances used by held-out recovery.

• replaces the fixed-variance linear rate with the population-optimal/asymptotic profiled Gaussian held-out log-score gap 0.5 * log1p(a^2 R / sigma^2);
• renames exported score/target fields away from Bayes-factor terminology while retaining deprecated Python and CLI aliases;
• leaves the residual-maximin objective and allocation unchanged because the corrected gap is monotone in R;
• states and tests the identification condition
  R_{k|l} = min_{b,c} E_w[(x_k - b - c x_l)^2]: for a nonconstant alternative, R=0 exactly when the generator is affine in that alternative on the positive-weight support; constant-alternative and constant-generator cases are handled explicitly;
• updates default unit-SNR target guidance to 45 / 93 / 1,113 trials for optimized / uniform / coupled-novelty designs.

Recovery rates and allocation evidence must be regenerated on the prespecified five seeds before manuscript import. The planning gap is a matched-candidate population target; it does not itself establish open-set/mixture robustness or quantify finite-training plug-in score uncertainty.