[DimaMolod]

Stacked on #662. I measured coverage of everything changed since 2.10.0 and added a test for each uncovered case a run can actually reach. That surfaced one bug, fixed here.

The bug

write_fold_outputs warned and returned when ranking_scores.csv had no rows, and postprocess went on to mark the fold complete. A later --allow_resume run then skipped a fold that had no model. It now raises: the job is reported failed, the fold record stays incomplete, and a resumed run predicts the fold again. Its test fails on the old code (DID NOT RAISE).

The tests

• AF3 outputs:
  • a seed is still written when its ModelCIF metadata cannot be added;
  • metadata records are kept once (homodimers, copied sidecars);
  • an unreadable sidecar is only a warning.
• AF3 runner:
  • seeds are handed over one by one as they are predicted, which seed-level resume relies on;
  • missing weights stop the fold before anything is written;
  • a kernel record that is not a JSON object is replaced;
  • the model config sets only the fields its AF3 release has.
• AF3 inputs:
  • a JSON input read from its .json.xz copy, as the published database ships them, and a missing one;
  • RNA and DNA JSON chains sliced to regions;
  • regions outside a chain, or on a ligand, rejected;
  • RNA, DNA and ligand chains renamed on a chain-ID clash;
  • an unreadable metadata sidecar leaves the description alone;
  • duplicate-residue normalisation skipped when the sequence doesn't match;
  • template sequences shorter than the query;
  • job names of fold inputs and of nothing.
• AF3 debug:
  • no translation summary without a translation;
  • only rows with a species are paired;
  • the final MSA is skipped without alignments and never fails the prediction.
• Storage:
  • no best sample without a readable ranking;
  • a slim pass on a missing or already-slimmed directory.
• AF2: a JSON file truncated by an older version is ignored on resume.

Coverage of the code changed since 2.10.0

Measured as branch coverage over the unit and integration suites in both release images, combined: 93% → 96.4% of 1,268 changed statements. What's left:

• the model forward pass in ModelRunner, which only the GPU suites run, and all of them passed this cycle;
• defensive except fallbacks and unreachable type guards.

AlphaLink is omitted by .coveragerc (torch). Its change in #661 was signature-only, and test_inference_flags covers it.

Verification

• Unit + integration, both release images: no new failures. AF2 image: 970 passed, 26 skipped (946 before). AF3 image: 1041 passed, 15 skipped (1017 before). The same pre-existing environmental failures as on main.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

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