Benchmarking & evaluation harness for perturbation prediction (baseline-aware)

Context

As we add perturbation-prediction models, we need a standardized, honest way to evaluate them.
There is strong evidence in the field that trivial baselines (control mean, additive model) often match or beat elaborate models, so baselines must be first-class, not an afterthought.
PerturbationComparison is a start but is limited.

What's missing

• Standard train/test splits for perturbation prediction: held-out perturbations, held-out combinations, held-out cell types.
• Built-in baselines: control mean, additive/linear, nearest-perturbation (reuse DistanceSpace.nearest_perturbations).
• Standard metrics: per-DEG delta / logFC correlation, E-distance to ground truth, direction (sign) accuracy, top-k DEG overlap.
• A simple leaderboard-style summary over models × metrics × splits.

Proposal / API

A pertpy.tools evaluator that takes ground-truth and predicted AnnData (as returned by the prediction models) and returns a tidy results DataFrame, reusing the existing Distance metrics.

Why it matters

Gives users — and reviewers — an honest, reproducible answer to "is this model actually better than doing nothing clever?".
This is exactly the rigor scverse is trusted for, and it is the natural companion to a broader prediction module.

Related

Prior discussion in #173 (benchmarking tool speed) is about runtime, not predictive accuracy — this is complementary.

Start by reviewing PerturbationComparison and the existing Distance metrics, then inspect DistanceSpace.nearest_perturbations. The proposed entry point is a pertpy.tools evaluator accepting ground-truth and predicted AnnData and returning a tidy results DataFrame. Done means the requested splits, baselines, metrics, and leaderboard-style summary are standardized and covered by appropriate tests.