fix: PositionsLH penalty silently zero when positions contain inf rows

[samlange04]

Summary

Discussion: https://github.com/orgs/PyAutoLabs/discussions/30

The position penalty chained from a previous result via positions_likelihood_from did nothing whenever the positions held (inf, inf) rows, which the point solver's fixed-length output always has. PositionsLH.log_likelihood_penalty_from takes a plain xp.max over the traced separations; an inf position traces to nan, the max is nan, nan > threshold is False, and the penalty is zero for every model. output_positions_info and positions_threshold_from are nan-safe, so positions.info reported the true (violating) separation while the likelihood never saw it. In an HST SLaM fit this let source_pix[1] settle on a demagnified theta_E = 0.37" solution with a 2.38" source-plane separation against a 0.2" threshold.

This PR drops non-finite rows in PositionsLH.__init__ (numpy, at construction, so the JAX path sees a clean static grid) and raises PositionsException when fewer than two finite positions remain, extending the existing length-one check. A comment at the site explains why.

Not changed, open to discussion: positions_likelihood_from could also strip the padding before building the object, and the penalty could use a nan-safe max. A nan-safe max would also hide a real position that traces to nan, so filtering at construction is the narrower fix.

API Changes

• PositionsLH now silently ignores non-finite positions. Previously they disabled the penalty, so no working configuration changes behaviour.
• PositionsException is now raised for fewer than two finite positions, not only for exactly one. The message reports how many non-finite rows were dropped.

Test Plan

• New test__non_finite_positions_are_dropped_and_the_penalty_still_fires: images at +/-1" padded with inf rows; the penalty is 0 at theta_E = 1" and fires at 0.3". Red on main.
• New test__fewer_than_two_finite_positions_raises.
• Every test file referencing PositionsLH / positions_likelihood: 61 passed (macOS, Python 3.13).
• test_autolens/analysis, point, imaging: 381 passed, 1 xfailed, 1 failed. The failure (point/triangles/test_static_lattice_jax.py::test__source_on_a_step_0_vertex_returns_the_two_true_images) fails identically on unmodified main and does not touch PositionsLH.
• Reproduced with the real J0008-0004 positions and fitted mass: max separation nan before, 2.384" after.

🤖 Generated with Claude Code