Fix scoring provenance, input validation and report integrity

[DimaMolod]

AlphaJudge could reuse stale scores, reject native AlphaFold-Multimer output, choose structures nondeterministically, and omit PDF table rows. This change makes input handling and cache reuse explicit and preserves all report rows.

• Validate cached CSVs against scoring options, input fingerprints, software/calibration and the CSV checksum. Default input checks use size, mtime, ctime and file identity; --cache_validation content enables full hashes. Metadata can miss same-size edits with coarse or preserved timestamps. Old manifests and incomplete scoring runs are recomputed; files are published atomically.
• Separate PAE rendering from scoring. PNG options and rendering failures do not invalidate CSVs. Missing or stale PNGs can be regenerated on a cache hit without constructing interfaces or rerunning biophysical scores.
• Accept native AF2 Multimer rankings and recover individual confidences/full PAE from result pickles, including gzip/xz. Preserve missing individual scores when only combined confidence exists.
• Record the parser backend and selected structure. Explicit backend provenance prevents Boltz filenames from selecting an AF2 calibration. AF2 selection is deterministic; --af2_structure unrelaxed overrides the default relaxed-first preference. Historical calibration predates per-row structure provenance, so its relaxation policy cannot be reconstructed reliably; the frozen distributions are retained.
• Reject corrupt or unalignable PAE with file context and per-model tracebacks; discard partial rows from failed models. Missing AF3 full-confidence files raise FileNotFoundError even when the summary contains chain-pair minima; combined exports carrying actual full PAE remain supported.
• Preserve Bio.PDB residue hierarchy, paginate interface and aggregate ranking tables, preserve the caller's Matplotlib backend/settings, and reject invalid aggregate-report CLI options before scoring.
• Record cutoffs, calibration identity and the actual contributing features/count. Reports explain recalculation and qualify custom/unknown cutoffs. Remove the unsupported interface-area percentile registry entry while retaining raw area.

pDockQ2 coefficients and its existing pairwise formula are unchanged. A regression preserves agreement with the larger directional score from IPSAE at 6174cf9. Frozen quantile ladders and percentile tie policy are unchanged. README is shorter than on main, with a plain cache note and changelog link. New CSV field definitions are in the changelog.

Validation:

• Regressions reproduced the failures before the fixes. The focused suite now passes 128 tests after the pre-release cleanup, including PNG recovery without rescoring, stat/content cache policies, explicit coordinate preference, AF3 error distinctions and pDockQ2 compatibility.
• Cache identity benchmark on 32 fixture files (59,719,098 input bytes), three warm measurements: old median 149.5 ms, default metadata mode 34.5 ms with zero input payload bytes read, explicit content mode 128.9 ms. Software and CSV checksums remain enabled.
• Rendered PDF review: all 36 interface rows and all 42 aggregate ranking rows remain visible across repeated-header pages.
• Final broader integration suite after cleanup: 182 passed, 6 skipped on Python 3.12 in 11m15s. CI runs all six slow CCP4/SC references across Python 3.10–3.13. The 1.4.4 wheel and source archive also pass strict metadata checks; an isolated wheel installation passes CLI scoring and warm-cache reuse.

Pre-release cleanup: shared parser options, typed cache settings, separate scoring/publication steps, shared PAE and manifest helpers, unified report labels/page sizes, and shared test fixtures. Version is 1.4.4. Input discovery remains conservative so newly available preferred structures invalidate the cache. The final broader suite and all pre-merge CI checks passed, including Python 3.10–3.13 and Docker. Released as 1.4.4.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.