Correct experiment configuration and model selection

[CmsChase]

Summary

This PR corrects the historical evaluation pipeline and publishes the corresponding interpretation note as one reviewable change.

• enforce a strict, versioned experiment config
• record effective configuration fingerprints across generated artifacts
• preserve the historical 70/15/15 split behavior in a dedicated audit config
• select models and feature settings only on validation MAE
• attach test metrics only after the candidate is fixed
• remove test-selected exposure winner claims
• add disagreement fixtures where validation and test favor different candidates
• publish a calm public erratum while retaining historical papers and tags
• consolidate the two-minute project introduction in QUICK_OVERVIEW.md
• move error analysis under docs/

Root cause

The previous loader did not consume the nested split keys in the checked-in config and fell back to 70/15/15 for both strategies. Several downstream summaries also selected a reported winner after observing test MAE. Training code used training rows, but repeated test-informed selection means the historical test outcomes are retrospective rather than unseen confirmatory evidence.

Reproducibility impact

• configs/experiment.yaml is the corrected current pipeline.
• configs/experiment_legacy_v6.yaml records the effective historical ratios for audit.
• The legacy config alone does not recreate a paper; exact historical reproduction also requires the matching Git tag and original local data snapshot.
• Historical generated tables and PDFs are intentionally not overwritten.

Validation

• python -m pytest -q ? 131 passed
• python -m compileall -q src
• git diff --check origin/main...HEAD
• scanned the proposed diff for prohibited public naming and prospective files

Scope boundary

Prospective input capture, frozen prediction, protocol, and ledger work are intentionally excluded and will be reviewed separately after this PR.