fix: stabilize border relocation for degenerate PCA axes

[Jammy2211]

Summary

• detect when the border covariance's PCA eigenvalues are equal within sqrt(machine epsilon) relative scale
• select a deterministic axis-aligned frame only in that near-isotropic case
• retain the ordinary PCA major axis for non-degenerate borders
• add regression coverage for the degenerate branch and its anisotropic control

PCA eigenvectors are mathematically undefined for equal eigenvalues. NumPy and JAX can choose different valid orientations; deriving maximum ellipse extents along those arbitrary axes made border relocation backend-dependent. The deterministic branch removes that ambiguity without changing the public interface.

Evidence

The downstream full-FitImaging reproducer measured, before this fix:

• relocated source-grid error: 2.835e-1
• mapping-matrix error: 1.492e-1
• curvature-system error: 8.690e-3
• reconstruction error: 8.989e-3

With this patch applied locally at the worst one-ULP point:

• curvature-system error: 1.583e-16
• data-vector error: 1.331e-16
• reconstruction error: 8.202e-11
• figure-of-merit error: exactly zero
• NumPy/JAX support masks: identical

Validation

• 2 focused regression tests passed
• direct JAX jit and vmap round-trips passed with deterministic phi=0
• downstream NumPy/JAX full-likelihood parity reproducer passed
• GitHub's Python 3.12/3.13 full test matrix is the merge gate

API Changes

None. No new configuration or public symbol.

Numerical Behavior

Only near-isotropic border covariances enter the deterministic-axis branch. Non-degenerate PCA behavior and all NNLS settings/algorithms are unchanged.

RED Gate Authorization

The user explicitly instructed this session to continue the work and merge each change once its CI is green. This PR will not be merged unless its own required checks succeed. No release is authorized or performed.

Closes #445