Two research runs on 2026-10-25 reported the same thing: platform/ci/pyenv.py torch took over 15 minutes, and a second call started while the first was running began a duplicate pip install.
The cause was in pip_install: every pip run had a fixed 300-second timeout. A torch wheel does not download in 300 s on the cloud proxy, so pip was killed, the script went on to the next interpreter and then to the private venv, and each restarted the download from nothing.
The fix in platform/ci/pyenv.py: torch-class packages (torch, torchvision, torchaudio, tensorflow, jax) get 1800 s per pip run, PYENV_PIP_TIMEOUT overrides it, and a lock file next to the venv makes a second concurrent call wait and then find the first call's install importable. The "can take over 120 s" notice now says torch-class wheels can take 15 minutes or more. A unit test pins the timeout rule; the lock was checked only by the existing tests still passing.