Add observable command-line Jacobian analysis

[charlesmartin14]

Summary

• replace the opaque Papermill execution path with a normal unbuffered command-line Jacobian runner
• log checkpoint loading, every Jacobian construction, every power-law fit, elapsed time, and rolling ETA
• atomically persist status, partial CSVs, spectrum plots, and alpha-progress plots after each checkpoint
• validate only selected checkpoints as they are consumed, while retaining exact provenance and optional ECS cover ranks
• add resume behavior and full terminal/file traceback reporting

Validation

• python -m py_compile baseline/experiments/mnist_mlp3_tangent_rg/scripts/run_short100_jacobians_cli.py
• bash -n baseline/experiments/mnist_mlp3_tangent_rg/scripts/run_short100_jacobians_cli.sh
• git diff --check
• exercised fatal path: traceback appeared in terminal and jacobians.log, with structured status.json and fatal_error.json

The sandbox lacks PyTorch and pytest, so the synthetic numerical test could not be executed here; it is included for the repository environment.