Capitalization-Efficiency Monitor MVP

Run with:

/home/maxwelhelp/main/bin/python3 experiment.py

experiment.py performs two checks with fixed seed 2206:

1. An exact quadratic toy model with DeltaV = a*gamma - b*gamma^2/2 and Gaussian posterior KL gamma^2/(2*s2). It tests KL quadratic scaling, value linear scaling for small updates, efficiency inverse scaling, the predicted aligned-value zero crossing 2a/b, and nuisance updates with constant negative efficiency.
2. A small Adam regression experiment with one useful feature and 24 random nuisance features. The monitor uses posterior-to-prior diagonal-Gaussian KL approximated by squared parameter displacement and treats negative/low held-out value-per-KL as a trust-region signal by undoing the update. results.json contains all outputs.

This is an MVP: optimizer dissipation Sigma is omitted, the posterior variance is fixed, and held-out value is measured directly rather than through paired deletion-counterfactual tasks.