[codex] Tighten AF3 trimer pairing summary and coverage

[DimaMolod]

Summary

This PR tightens AlphaFold3 trimer MSA translation reporting and adds regression coverage for three-chain AF2-feature -> AF3 runs.

What changed

• tighten the AF3 backend translation summary so it reports the effective paired depth that survives AF3 multi-chain species pairing, alongside the raw translated paired input depth
• add unit/backend regressions for trimer scenarios, including sparse middle-chain pairing, min-count cropping, and permutation invariance
• add direct native-AF3-vs-translated-AF2 pairing equivalence tests for trimer permutations, including the empty-middle-chain case
• add a cluster AF3 trimer smoke test that runs from regenerated MMseqs AF2 feature pickles
• adjust the trimer cluster assertion to account for AF3 grouping identical chains into a single JSON protein entry with multiple chain IDs

Why

We wanted the AF3 backend path from AF2 pickled features to be trustworthy for trimers, especially around the three-chain pairing behavior. The previous debug summary could hide the difference between translated paired rows and the rows AF3 actually keeps after species pairing, and we were missing trimer-specific regression coverage.

Validation

Local:

• pytest -q test/unit/test_af2_to_af3_msa.py test/unit/test_alphafold3_backend_helpers.py

EMBL cluster:

• test/cluster/check_alphafold3_predictions.py::TestAlphaFold3MmseqsIssue588Inference::test_issue_588_mmseqs_af2_features_enable_af3_species_pairing_trimer_inference
• submitted through test/cluster/run_alphafold3_predictions.py
• final passing Slurm job: 51562990
• log directory: /scratch/dima/af3_trimer_cluster_logs_20260410_final

Impact

This makes the AF3 backend diagnostics much easier to trust for trimers and gives us direct regression coverage for the three-chain pairing cases we were worried about, including permutation handling and missing-chain pairing.