Audit Prop 4 model math and publish PM-CAST-D results

[Harman6139]

What changed

• repairs and strengthens the Prop 4 CP-REPO-OPS-HG mathematical-fidelity audits
• adds the PM-native discriminating validation and the PM-CAST-D practitioner model
• publishes cleaned run summaries, diagnostics, figures, and reproducibility metadata
• removes current-tree credential literals and documents environment-variable setup
• records the verified D: archive for local raw/downloaded artifacts and prevents those scratch copies from being recommitted

Why

This promotes the validated model implementation and review evidence from repo 1
into the consolidated repo-2 branch while keeping downloaded Actions bundles,
duplicate full prediction panels, smoke probes, and temporary outputs outside Git.

Research impact

• Corrected CP_REPO_FRESH remains the strongest always-on SMAPE benchmark in the Prop 4 ladder.
• PM-native geometry has conditional value but does not pass the full PM-specificity success bar.
• PM-CAST-D improves pooled SMAPE and MAE, but not MSE or RMSE, so it is retained as a practitioner/diagnostic model rather than an all-fronts winner.

Validation

• python -m pytest projects/harv-extension/tests/test_prop4_model_math.py projects/harv-extension/tests/test_pm_cast_d_math.py -q
• result: 16 passed
• notebook JSON parse and order-book Python syntax checks passed
• current-tree and staged-snapshot credential scans found no remaining literal credentials
• archive verification: 1,046 files / 2,582,051,591 bytes / zero SHA-256 mismatches

Security note

The current tree removes legacy credential literals. The exposed historical API
credentials and database password should be rotated/revoked by their owners;
this PR intentionally does not rewrite shared Git history.