Add reduced Jacobian analysis preset

[charlesmartin14]

Summary

• add a separate reduced command-line preset that fits only the centered log-singular radial Jacobian on every layer and the two ECS/Grassmann covers on fc1
• compress the ECS spectrum from uniform shell-coordinate copies to one physical amplitude per retained-core group
• add explicit method selection and configuration-aware checkpoint resume records to the underlying CLI
• use a new /private/tmp/rg-mnist-mlp3-short100-jacobians-reduced output tree

Uniform ECS replication does not change the empirical CDF, fitted alpha, selected xmin, or KS distance; compression prevents deterministic copies from inflating effective sample size and reduces each ECS fit from tens of thousands of values to roughly k physical amplitudes.

Validation

• Python compilation passed
• shell syntax validation passed
• parser/method-selection check passed
• ECS compression value/count check passed
• git diff --check passed

The sandbox lacks PyTorch and pytest, so the repository numerical test was not executed here.