Stabilize Euclidean MTT distances at extreme scales - #4968

Summary

• compute Euclidean MTT pairwise norms with numpy.hypot.reduce instead of numpy.linalg.norm
• preserve finite localization distances when coordinate magnitudes are close to the float64 limit
• add a focused regression that runs with floating-point overflow configured to raise

Bug

_euclidean_mtt_distance() formed finite displacement vectors and evaluated them with numpy.linalg.norm. For a displacement such as [1e308, 1e308], the mathematically correct norm is approximately 1.4142e308 and is still representable, but numpy.linalg.norm squares the components internally and overflows to infinity.

With a cutoff of 1.5e308, the overflow caused the pairwise cost to be clipped to 1.5e308, overstating the localization error. Under numpy.errstate(over="raise"), the same valid input raised FloatingPointError.

Fix

Use numpy.hypot.reduce(..., axis=2), which accumulates Euclidean norms with scale-aware arithmetic and returns the correct finite value without an overflowing square.

Validation

• reproduced the old result: 1.5e308 after erroneous cutoff clipping
• verified the patched calculation returns 1.4142135623730951e308
• verified the patched path succeeds under numpy.errstate(over="raise", invalid="raise")
• added a regression to tests/evaluation/test_euclidean_mtt_distance_orientation.py
• branch is 2 commits ahead of and 0 behind current main; the diff contains only the implementation line and focused test

Full repository tests were not run locally because this environment could not resolve GitHub for a checkout; GitHub Actions should exercise the in-tree regression and full matrix.