[KartikP]

Changes from a new user's review of UMI v2.

• RandomSelection.matched_to(selection, seed=0): the matched random control in one call. It copies the layer, unit count, population and layout from a resolved functional selection.
• Readout warning: the sklearn multi_class deprecation notice is silenced at fit only. The argument is still honoured below 1.8, and dropping it would move binary probabilities; environment_check already guards the version.
• Plugin installs:
  • Installing a plugin's requirements now holds the shared pins (numpy, torch, transformers, ...) at their installed versions. A plugin that needs other versions raises PluginDependencyError naming the conflict ("the plugin asks for transformers==4.30.2; this environment holds transformers==4.57.6") and how to give it its own environment. Packages with no build for the platform, or that fail to build, get their own message.
  • Requirements lines that name Python's own modules (imp, collections) are skipped. They made 16 vision models (BiT-S, mobilenet_v2_*, vonegrcnn) fail to install anywhere.
  • In newenv mode, the plugin may still set any versions it needs.

Dry-run of all 630 vision and language plugins against the pinned environment:

• 535 install without changing anything; 67 have no requirements.
• 8 would move a pin (7 CvT models: transformers 4.57 to 4.30; lm1b: numpy 2). These are now refused with the message above.
• 1 other moves only OpenCV (cv_18_dagger_408_pretrained) and still installs.
• The other 19 cannot install at all: 16 were the built-in module names, now skipped; 3 have packages with no Mac build.

Note: the existing newenv runner (conda_score.sh) cannot work on v2 today. It installs vision into a fresh environment, and vision pins brainscore-core==2.4.0rc1, which is not on PyPI. The conflict message therefore suggests cloning the environment instead.

CI runs unified at KartikP/brainscore-unified feat/newuser-ergonomics, whose dependency test checks the new install command.

Tests: core suite passes locally, except tests that need conda on PATH or a checkout path without spaces (they fail the same way on base).

https://claude.ai/code/session_01WU1w1HjLtiEMf1v2LHdoKs