fix: keep AF2 chain IDs at A, B, C under padding and out-of-order asym_id

[DimaMolod]

Two chain-labelling bugs in _normalize_asym_id, both in the AF2 asym_id → Protein.chain_index handoff that feeds protein.to_pdb.

Ranking used the wrong permutation. appearance_order[inverse_indices] applies the first-appearance ordering instead of its inverse. Three or more chains whose asym_id are not already in first-appearance order come out shuffled — 3,1,2 produced B, C, A.

Batch padding outranked the real chains. pad_input_features zero-pads asym_id to the batch-wide residue bound, so --desired_num_res runs reach the remap as [1…1, 2…2, 0…0]. The padding then took rank 0 and a dimer came out labelled C and A. Trailing zeros are now treated as padding and ranked after every real chain; a leading zero still reads as a genuine 0-based encoding.

Verified against the real pad_input_features inside alphafold2_2.5.0.sif: a padded dimer goes from C/A to A/B.

Four regression tests in test/unit/test_alphafold2_backend_helpers.py. The two covering these bugs fail before the change; the other two pin behaviour that already worked — stability across the five models that share one feature dict in multimer mode, and 0-based encodings. test/unit is 500 passed, 11 skipped (test_diagnostics.py fails to collect on main too, missing af2plots).

🤖 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.