Reproduce FEGA

[phusroyal]

Summary

This PR adds a notebook-based reproduction of FEGA (Feature-Effect Geometry Analysis) and the generic Murano operations needed to run it.

What changed

Murano

• Added SAEModel.decode() to reconstruct model states from SAE features.
• Added generic access to the underlying model through raw_model.
• Kept these changes independent of FEGA so other Murano workflows can reuse them.

FEGA reproduction

• Added a self-contained FEGA bundle under notebooks/reproductions/fega/.
• Added an executed example that runs one SAE feature from prompts through intervention and geometry analysis.
• Included eight compact effect matrices, one representative for each FEGA geometry family.
• Recomputed metrics, labels, stability decisions, and plots from those matrices.
• Compared the recomputed results and visualizations with the original FEGA outputs.
• Added sphere-surface and 2D comparisons for all eight examples.
• Added a script for rebuilding the compact data from the original FEGA artifacts.
• Added focused tests, documentation, licensing, and source attribution.
• Updated the reproduction website page with clear steps, results, and limitations.

Scope

This reproduces FEGA for selected representative features. It does not rerun the paper’s complete dataset construction, feature search, model and SAE sweeps, aggregate tables, or full atlas.

Validation

• Murano suite: 1,081 passed, 1 skipped
• FEGA tests: 12 passed
• Notebook import tests: 22 passed
• Ruff: all checks passed
• Formatting: all 126 files formatted