protein_binding: switch s4 to Boltz-2, remove hardcoded paths, pin GPUs

[drawadiagram]

Splits the protein-binding workflow changes out of #59. Second of four.

Based on #60 (needs find_gpus and IMPRESS_SESSION_DIR). GitHub will retarget this to main when #60 merges — review only the commit on this branch.

Structure prediction: AlphaFold multimer → Boltz-2

• s4_boltz.sh resolves its interpreter from BOLTZ_VENV/VIRTUAL_ENV instead of sourcing a hardcoded /anvil venv, and makes --use_msa_server opt-in via BOLTZ_USE_MSA_SERVER=1 because compute nodes have no internet.
• CCD cache extraction race fixed. This killed 13 of 16 pipelines in a production run. boltz's own download_boltz2() decides to skip extraction based on whether the mols/ directory exists, but tarfile.extractall() creates that directory entry immediately — so a second concurrent task saw mols/ present, skipped extraction, read a half-populated cache, and died with CCD component not found. The lock is now held across a completeness check of mols/ against mols.tar, with a .mols_complete marker so later runs skip the O(45k) recount once warmed.
• s3 embeds pre-computed MSA paths in the FASTA header, falling back to single-sequence mode when the cache is cold.
• af2_multimer_reduced.sh and scripts/s4_alphafold.sh are deleted, along with the commented-out s4_alphafold task block that was their only remaining call site. That also resolves @AymenFJA's "remove any Dragon reference" comment on s4_alphafold.sh:10 — the file is gone.

⚠️ GPU scheduling — the change most worth reviewing

task_description={"gpus_per_rank": 1} is removed from the s1 and s4 signatures. GPU assignment now happens by pinning: a gpu_id kwarg is threaded into s4_boltz.sh as CUDA_VISIBLE_DEVICES, propagated to child pipelines, with the runner round-robining find_gpus() across pipelines. A module-level per-GPU asyncio.Semaphore(2) caps concurrent Boltz launches on each GPU across every pipeline pinned to it.

@AymenFJA — this is what your protein_binding.py:203/:212 comments were pointing at. To be precise about one thing: no task decorator changed. s4_post_exec was already local_task=True on main (I diffed them), and s4 already had capture_stdio=True. The real change is the resource-request removal above, and it deserves your eye.

Portability

MPNN_PATH comes from the environment and fails loudly when unset; base_path splits into input_base_path/output_base_path (IMPRESS_BASE_DIR / IMPRESS_OUTPUT_DIR / IMPRESS_SCRIPTS_DIR); every scripts/*.sh re-activates $VIRTUAL_ENV instead of a hardcoded cluster path; plddt_extract_pipeline.py writes its CSV under --path rather than cwd.

Adds delta_env_setup.sh and delta_gpu_run.sh (with IMPRESS_BACKEND=dragon|local and IMPRESS_TEST_MODE switches). The leaked --mail-user address @AymenFJA flagged is now a <your e-mail> placeholder.

Robustness

s1, s4 and s5 each check for their real output before re-raising, because the execution backend intermittently reports failure for a task that completed. run() now aborts the pass if every s4 failed instead of proceeding to pLDDT extraction on nothing.

Beyond a straight extraction

Deleting s4_alphafold.sh left dangling references in CLAUDE.md and README.md; both now describe Boltz-2 as the only s4 path.

Still open from #59

@AymenFJA's comment on delta_env_setup.sh:105 (the Dragon/Delta comment being misleading) is carried over unchanged — I'd rather you tell me what it should say than guess. On installing Dragon twice (:107): it's installed exactly once, via rhapsody-py[dragon,telemetry], which is what your comment asks for.

Replaces part of #59.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lo8DwSbyvdWZRkkkka6gA2