Test: Phase 1 native T2/T5 tier coverage

[wilke]

Follow-up from PR #22 (tiered-fixture-ladder). Picked up the deferred Phase 1 coverage from docs/TEST_COVERAGE.md.

What's missing

PR #22 implemented T1–T5 tier coverage in Phase 2 (CLI/adapter layer) and Phase 3 (service-script + workspace). Phase 1 (native tool execution, bypassing the predict-structure CLI) was deferred because Phase 2/3 give higher coverage-per-test-time on the layers most likely to drift.

But Phase 1 still has value as a baseline: if Boltz / OpenFold / Chai / AlphaFold native command lines drift in the next container rebuild, Phase 1 catches that without the adapter layer in the way.

What to add

In tests/acceptance/test_phase1_native_tools.py, for each tool that supports it, add:

• test_protein_t2_with_msa — medium fixture (/data/medium_protein.fasta + /data/msa/medium_protein.a3m). Marked tier2. Boltz / Chai / OpenFold need the MSA; ESMFold ignores it; AlphaFold builds its own (skip MSA arg).
• test_protein_t5_large — large fixture (/data/large_protein.fasta + /data/msa/large_protein.a3m). Marked tier5 + slow. Likely Boltz + OpenFold only (Chai gets slow at 434 aa; ESMFold may OOM on H200 without chunk-size tuning).

Reuse the existing per-tool _binds() helper and the input format already established for each tool's native CLI invocation.

References

• tests/acceptance/matrix.py::TIER2, TIER5, msa_args_for
• tests/acceptance/test_phase1_native_tools.py (existing T1 baseline tests)
• Fixtures already committed in PR Tiered fixture ladder (T1-T5) + close test-coverage gaps #22:
  • test_data/medium_protein.fasta + test_data/msa/medium_protein.a3m
  • test_data/large_protein.fasta + test_data/msa/large_protein.a3m