[PaulHax]

Co-training CI checks no longer match the plotting API and recipes

The plotting test imports an API that the current module does not expose,
stopping collection. Several recipe checks expect settings that the configured
experiments have since changed, and
lint findings prevent the current co-training tree from passing CI.

Exercise the current behavior and retain useful failure output

Test the figure-based plotting CLI through its rendered outputs, assert the
configured recipe differences explicitly, and complete the evaluator's fake
state contract. Resolve the lint and formatting findings while keeping scientific
settings unchanged. Print test names and retain
JUnit artifacts so failed CI runs are easier to diagnose. Run fast co-training
checks in a separate process group so their compiled programs are released
before the other fast checks. Record runner resources, periodic memory readings,
and per-command resource reports. Abrupt runner termination may still prevent
artifact upload; the job log retains periodic readings. The cause of the runner
shutdowns is not yet established.

Related: #23 and #36. Targets cotraining.