Unify experiment comparison protocol + real-data grid figure pipeline

[johmathe]

Summary

Addresses the figure feedback and unifies the comparison protocol across all three experiments (PCam, OrganMNIST3D, Spherical MNIST). No experiments were run in this PR — sweeps are meant to run on separate GPU machines (see experiments/README.md for the run matrix, tmux commands, and rsync instructions).

Unified protocol

• Non-equivariant CNN baselines are augmentation-trained (train_mode R) and labeled "Aug. CNN" everywhere (feedback Update general API and add cg coefficients #3).
• Invariant models are canonical-trained (train_mode C); rotation (OOD) evaluation is always on.
• All line plots report the rotated test metric ("Rotated test AUC/accuracy") — no more "NR/R" jargon (feedback Add pre commit and actions #1).
• Every bar method appears in both line plots of its row, as a single point when it has no capacity axis (feedback Fix indices in the bsp #4 generalized).
• One power spectrum only: power_spectrum_matched removed; the capacity sweep now extends the regular power-spectrum curve to the bispectrum's max budget (feedback Fix indices in the bsp #4).
• Cohen S²CNN stays a dashed published-reference line with a legend entry and a caption justification (feedbacks Add code #2, Revert "Fix indices in the bsp" #5): rows 1–2 already contain G-equivariant architectures; S²CNN appears in row 3 only as the canonical published spherical-image architecture (cited, not re-run).

Code changes

• PCam: new norm_pool model variant — equivariant DenseNet with the previously-unused GroupNormPool as final invariant pool (matched-architecture incomplete-invariant ablation, analog of Organ3D's norm_pool / SMNIST's power spectrum). Wired into train.py and both sweep scripts.
• Spherical MNIST: new run_capacity_sweep.sh (2 models × 5 widths × 3 seeds, C-trained); PS widths reach ~550K params to cover the bispectrum budget (~529K). Matched-run concept removed from analyze_results.py.
• Organ3D: all 4 models (incl. norm_pool and R-trained standard) now covered in the data-efficiency and capacity sweeps, 3 seeds each; redundant run_tier1_sweep.sh deleted; skip-check dir names fixed to match train.py output naming.
• Grid figure: experiments/make_grid_figure.py replaces make_mock_grid_panels.py — same render path fed by results.json loaders, --mock flag for layout previews (watermarked), emits caption.txt with the S²CNN justification.
• Docs: experiments/README.md with the full run matrix for external GPU machines.

Sanity checks run (no training)

• norm_pool model builds, forward/backward, exact C8 invariance (gap = 0.0), param count matches fourier_elu (109,815).
• make_grid_figure.py --mock renders; real-data path verified end-to-end against a synthetic results tree exercising every runner's output layout.
• ruff check clean; bash -n on all runner scripts.

Test plan

• Run the 8 sweeps on the GPU machines per experiments/README.md
• rsync results back and run python experiments/make_grid_figure.py
• Check panel coverage warnings (loader prints a WARNING for any missing method/panel)

Made with Cursor

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.46%. Comparing base (88ae4df) to head (a9dd551).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #50   +/-   ##
=======================================
  Coverage   95.46%   95.46%           
=======================================
  Files          11       11           
  Lines        2227     2227           
=======================================
  Hits         2126     2126           
  Misses        101      101           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.