[DimaMolod]

Plan item 1 of 3: make AF2's ModelCIF subprocess reliable. ModelCIF stays best effort.

Problem

AF2 post-processing built a shell command containing the output path, ran it with python3, and treated any stderr output as failure. As a result:

• a path with spaces or shell metacharacters broke the command, or ran parts of it as commands;
• python3 could be a different interpreter from the one running AlphaPulldown;
• the converter's ordinary logging was reported as an error, while its real exit code was ignored.

Change

A private helper, _convert_to_modelcif, runs the converter as [sys.executable, <converter path>, "--ap_output", <dir>] without a shell.

• The converter's stdout and stderr are passed on as diagnostics.
• A nonzero exit or a failure to start is reported as a warning.
• Success is logged as "converter finished", not "all PDBs converted".
• The order is unchanged: conversion, then storage cleanup, then fold completion. Conversion stays in its own process.
• The standalone converter, its flags and its exit behaviour are unchanged.

Tests

• Real subprocess, unusual path: a stand-in converter run in fold A+B; touch pwned $(touch pwned2) 'q' & \x``. It writes into exactly that directory, nothing else appears, and informational stderr does not count as failure.
• Failures: a nonzero exit and a launch failure each log a warning, and cleanup and fold completion still run.
• Disabled: with conversion off, no subprocess starts.
• Same output as before: a new integration test converts the TEST fixture through the backend helper inside a directory full of shell metacharacters, and with the standalone script in a plain directory. Every .cif is byte-identical, and the zip archives have identical members; only their timestamps differ. Both images run this test.
• Unit + integration, both release images: identical to main, plus the new tests. AF2 image: 979 passed. AF3 image: 1050 passed.
• GPU: runs on main after this round is merged.

🤖 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.