Cloud sessions here ship four or more Python interpreters, and they disagree about what is installed. platform/ci/pyenv.py finds one that imports what a test suite needs, installing the gap if none does. When nothing had everything, it ran pip install on the candidates in PATH order, starting with a bare /usr/local/bin/python3 that usually has nothing at all.
A run on 2026-11-10 reported the cost: ten minutes of installs into /usr/local/bin/python3 and then python3.12, while python3.13 was one package short.
The fix is one probe per interpreter. Before installing, the script asks each candidate which of the modules it cannot import, sorts them by how many are missing, and installs only the missing ones into the closest.
On this session's box, for the build-data research suite (-r research/build-data/requirements.txt, seven packages):
| interpreter | missing before install |
|---|---|
/usr/bin/python3.13 | 5 (scipy, scikit-learn, lightgbm, xgboost, pytest) |
/usr/local/bin/python3.13 | 5 |
/usr/local/bin/python3 (old first pick) | 7 |
/usr/bin/python3.12 | 7 |
The probe of all four took 1.1 s. The old order would have started with an interpreter missing every package; the new one starts with one that already has pandas and numpy, and now logs which interpreter it is trying and what for, so a backgrounded call is no longer silent about its choice.