Harden MNIST trace-log diagnostics

[charlesmartin14]

Purpose

Harden the existing MNIST/MLP3 matrix trace-log diagnostic without changing its scientific definition or replacing WeightWatcher's detX calculation.

Changes

• require the retrieved ESD to contain exactly min(weight.shape) eigenvalues for strict baseline measurements
• reject non-finite, zero, or negative eigenvalues instead of silently filtering them and renormalizing a reduced positive-rank spectrum
• require each WeightWatcher layer to map unambiguously to a two-dimensional model parameter
• audit that WeightWatcher's rescaled eigenvalues sum to the full spectral dimension
• reject out-of-range detX_num and num_pl_spikes values instead of clipping them into a valid range
• preserve the existing midpoint rule and trace-log formula exactly
• add analytic regression coverage for total trace-log, per-eigenvalue trace-log, and geometric mean
• add failure tests for rank deficiency, incomplete ESDs, invalid boundaries, inconsistent ERG gaps, and incorrect normalization

Deliberately unchanged

• WeightWatcher remains authoritative for detX_num, num_pl_spikes, and ERG_gap
• no independent cumulative-log detX boundary is introduced
• the midpoint retained rank remains floor((detX_num + num_pl_spikes) / 2)
• optimizer mathematics, hyperparameters, notebook source, output notebooks, and committed experiment results are unchanged

Validation

• git diff --check
• full source and test compilation
• 7 focused trace-log tests passed
• complete baseline suite: 67 tests passed
• pip check: no broken requirements

The final PR diff contains only:

• baseline/rg_baselines/diagnostics.py
• baseline/tests/test_diagnostics.py