Implement final adversarial statistical audit

[rickyjreyes]

Goal

Implement the strict final adversarial statistical audit for the NIST atomic-line analysis without changing the canonical scanner.

What changed

• Adds R/run_final_adversarial_audit.R, a fail-closed one-command final runner with explicit full-resolution budgets and no hidden 500/1,000-trial caps.
• Adds R/run_alternative_nulls.R with five scan-global null stress tests:
  • canonical fixed-baseline Poisson,
  • baseline-refit Poisson,
  • conditional multinomial,
  • overdispersed negative binomial,
  • block-residual/baseline-misspecification diagnostic.
• Expands R/global_multiple_testing.R from the old limited family to the complete declared Fe preprocessing multiverse plus source and neighbouring-ion controls.
• Corrects dependence language: Holm/Bonferroni do not require independence; adds Benjamini-Yekutieli FDR for arbitrary dependence while retaining BH and the approximate family-max calibration.
• Hardens frozen holdouts: k remains train-only and locked, while test-null calibration is now computed both with the observed smoothing baseline fixed and with the smoothing baseline re-estimated in every null replicate. The final holdout verdict uses the larger p-value.
• Adds R/build_final_adversarial_summary.R to emit conservative global evidence, strict claim-by-claim verdicts, and failed_claims.csv containing every non-passing claim.
• Adds a dedicated Quarto report and run documentation.
• Adds fast R tests for null helpers, holdout/refit behavior, failure-preserving verdict logic, and source/parse safety of the final entry points.

Final default budgets

• bootstrap: 5,000
• primary scan/null: 5,000
• observed/Ritz bootstrap: 2,000
• specification null per specification: 1,000
• frozen holdout null per design/calibration: 5,000
• full-family max null: 5,000
• outer null calibration: 10,000
• injection/recovery: 2,000 per amplitude/frequency
• injection decision null: 5,000
• alternative null: 5,000 per model

Failure preservation

The final claim matrix deliberately does not soften negative results. For example:

• bin stability below the audit-declared 80% threshold is fail, not silently promoted;
• partial frozen-holdout success is mixed;
• negative held-out M1-vs-M0 predictive transfer is fail even when the in-sample fit is very strong;
• independent experimental confirmation and a WCT physical mechanism remain not established.

largest_reported_global_p is only a conservative diagnostic maximum across reported global/multiplicity/alternative-null p-values. It is explicitly not a newly combined formal p-value.

Resolution-mode follow-up

The branch also contains a standalone exploratory follow-up in R/run_resolution_mode_diagnostics.R, documented by RESOLUTION_MODE_DIAGNOSTICS.md.

This follow-up does not run inside the final audit and cannot overwrite its verdicts. It was added to diagnose the observed resolution-dependent winner by:

• comparing the original fixed sigma=6-bins baseline with sigma(B)=6*B/160, which approximately fixes smoothing width in ell=log(wavenumber) space;
• tracking both the canonical Fe primary mode and the previously frozen GWTC k=9.602325620315224 at every dense bin resolution;
• storing the top five distinct local peaks so a winner swap is distinguished from disappearance of a mode;
• running conservative fixed-frequency tests for the exact frozen GWTC target across Fe wavenumber/observed/Ritz and Cr/Mn/Co/Ni/Ti ion-II data, with both fixed-baseline and baseline-refit nulls.

The GWTC value predates this NIST follow-up, but noticing the NIST/GWTC match is post-hoc. These diagnostics are therefore explicitly exploratory and do not rescue a failed 80% bin-stability claim.

Reproduction

Final audit:

Rscript R/run_final_adversarial_audit.R --parallel true --render-report true

Resolution follow-up, after preserving the final audit outputs:

Rscript R/run_resolution_mode_diagnostics.R --parallel true --fixed-null-n 5000

The heavy full-resolution audit is intentionally separate from CI. CI sources/parses the new R modules, runs the unit tests, and checks the canonical Python scanner/container. Final Monte Carlo outputs are produced by the commands above.

[rickyjreyes]

Release gate update (2026-09-11):

• Full-family multiplicity rerun at 5,000 simulations: family-max p=0.0002; Holm FWER p=0.0118; BY FDR p=0.0009485.
• Five adversarial null models x 5,000 replicates: worst-case p=0.0005999 (block_residual_refit).
• Bootstrap peak stability: 5000/5000 within the 2% reference region.
• Specification multiverse: 89.4% in the reference region.
• Observed/Ritz/wavenumber representations all select k=31.32653 at the Monte Carlo floor.
• Frozen holdouts: 8/8 direction-consistent; median conservative fixed-k p=0.001.
• Null calibration is compatible at 0.10, 0.05, and 0.01.
• Injection recovery is calibrated at A=0 and reaches high correct-detection power by A=0.075-0.10.
• Original fixed-sigma bin-grid failure is preserved (70% in region). Exploratory matched-ell smoothing selects the primary Fe branch across 60-240 bins, diagnosing the resolution/smoothing confound without rewriting the original verdict.

Release blocker fixed in this PR: model_comparison.R previously recomputed polynomial centering/scaling on the held-out test block while applying training coefficients. The branch now freezes the training transform for test prediction and includes a regression test. The previous held-out gain (-36.836) must therefore not be used in the final release; rerun only Rscript R/model_comparison.R, then rebuild R/build_final_adversarial_summary.R before freezing/tagging.

Also fixed dependency metadata: viridis is now required because several plotting modules call it directly.