Change beta0_ridge to standard ridge on β0 magnitude

[jaredgalloway]

Summary

• Rewrite _beta_ridge_penalty as a standard ridge: sum β0**2 across all conditions (reference included) times beta0_ridge. The old form penalized non-reference β0 difference from the reference — asymmetric under relabeling, and left the reference unanchored.
• Extract _beta_ridge_penalty to a module-level function so it's unit-testable directly (was a closure inside fit).
• Add three new correctness tests: gradient matches 2 * r * β0[d] for every d; penalty is invariant under condition relabeling; penalty is 0 iff every β0 is 0, with the formula checked against the model-under-test.
• Rewrite test_beta0_ridge_penalty — the old assertion (differences shrink) was literally the old invariant, so it had to change.
• Fix tuple(fit_config["beta_clip_range"]) crash when YAML sets beta_clip_range: null. Shared fix across three experiment helpers (scv2-spike, simulation, loss-normalization): a _clip_range helper returns None untouched, tuple-wraps otherwise.
• Update production spike config (config.yaml): tol 1e-4 → 1e-5, maxiter 50 → 75 (top-level + ge_kwargs + cal_kwargs); l2reg 0 → 1e-7; beta0_ridge 0 → 1e-3 (new semantics); beta_clip_range [-10, 10] → null (disables clipping, replaced by smooth regularization on the intercept).
• Reset beta0_ridge in config_experimental.yaml and config_test.yaml to 0.0 — the existing 1e-4 values were calibrated for the old difference-from-reference semantics and are not directly transferable.
• Update CLAUDE.md with the new beta0_ridge semantics and the correct YAML encoding for disabling beta_clip_range (single null, not [null, null]).

Test plan

• pixi run lint — clean
• pixi run fmt-check — clean
• pixi run test (pytest --doctest-modules) — 185 passed, including 4 new correctness tests
• Snakemake dry-run on updated production config.yaml — DAG constructs cleanly with beta_clip_range: null
• Production scv2-spike pipeline executed on orca04 — 5/5 rules complete in ~30 min, no errors, no TypeError from clip handling

Production run results

Executed pixi run remote-pipeline -- spike prod host=orca04. Pipeline completed cleanly; results preserved under experiments/scv2-spike/results-prod-237-beta0-ridge-magnitude/ in the main clone (gitignored per repo convention, matching how results-prod-232-* and results-prod-235-* are handled).

Training loss at fusionreg=0.0, rep_1 (first row of cross_validation_loss.csv):

Run	loss
#237 (this PR)	0.6147
#232 l2reg-1e-7 (closest baseline)	0.6192
#232 tight-tol	0.5968
#235 times-seen-threshold	0.6166

The new run lands squarely within the baseline range — a touch lower than the most directly comparable baseline (#232 l2reg-1e-7), consistent with tighter tol=1e-5 and longer maxiter=75 producing slightly better convergence. All per-condition losses finite, CV completed across the full fusionreg grid.

Closes #237

🤖 Generated with Claude Code

[jaredgalloway]

Closing without merging — failed experiment. Keeping the branch (237-beta0-ridge-magnitude) and the linked issue around for reference.