Remove redundant per-variant normalization in CV notebooks

[jaredgalloway]

Summary

• Both CV notebooks divided total_loss_{training,validation} by a sample-count dict, but PR Normalize functional score loss per variant within each condition #230 already made functional_score_loss a per-variant .mean() — so the division was doubly-normalizing, silently diverging from the "mean Huber loss per variant" axis label.
• Fix: drop the n_samples construction and the division in experiments/simulation/notebooks/cross_validation.ipynb and experiments/scv2-spike/notebooks/cross_validation.ipynb; alias mean_loss = loss so the CSV schema (and manuscript_figures.ipynb, which reads mean_loss) is preserved.
• Delete experiments/scv2-spike/config/config_mean_loss.yaml and its Snakefile profile branch — the default config.yaml already absorbed the V0.4.0-anchored grid, so the profile was a duplicate.
• Update CLAUDE.md's "Key design patterns" to warn that total_loss_* columns are per-variant averages.

Validation

Ran the test profile of both pipelines locally:

• Simulation: 24 CV rows, mean_loss == loss row-wise, loss values 4.93–6.09 (per-variant Huber scale, not sum-scale).
• Spike: 8 CV rows, mean_loss == loss row-wise, loss values 2.53–3.39 (per-variant scale).
• Neither output CSV leaks an n_samples column.

pixi run lint, pixi run fmt-check, and the full pixi run test suite (182 passed) are clean on this branch.

Test plan

• pixi run sim-test completes, CV CSV has mean_loss == loss
• pixi run spike-test completes, CV CSV has mean_loss == loss
• pixi run lint
• pixi run fmt-check
• pixi run test
• CI rerun of all three Python versions

Closes #231

🤖 Generated with Claude Code