fix: sum PositionsLH penalties instead of doubling the last entry

[Jammy2211]

Overview

Fixes #699. AnalysisLens.log_likelihood_penalty_from overwrote its accumulator with each entry's penalty and then added the variable to itself (lens.py:171-177), so the analysis subtracted 2x the LAST PositionsLH penalty and silently discarded every earlier entry in positions_likelihood_list.

Change

• The penalty is now the sum over all entries, matching the documented 1e8 * (max_separation - threshold) contract (autolens/analysis/analysis/lens.py).
• Docstring aligned to the actual behaviour (array-valued 0.0 when no penalty — kept for the JAX path; the loop change is a plain x = x + y, trivially xp-safe under jnp).

Science impact

• Inside the threshold the penalty is exactly 0 either way — converged posteriors and all positions-free results unchanged.
• Outside the threshold the fence slope halves from the undocumented 2e8/arcsec to the documented 1e8/arcsec: likelihood values in penalized regions shift (release-notes item).
• Multi-plane / double-source-plane penalty stacking is corrected: previously only the last plane's penalty counted.

Tests

• The imaging double-plane test asserted the SAME value (-44097289521.7) for one and for two penalties — the 2x-last signature. Repinned: single-penalty -22048644768.18 (exactly half), double-plane -44140499627.75 (the true sum, distinct from the old value).
• Interferometer positions test repinned the same way (-44097289569.2 → -22048644815.85).
• New regression assertions: the analysis penalty equals the sum of each entry's own log_likelihood_penalty_from, the two per-plane penalties genuinely differ, and the single-entry case equals the entry's value exactly (no 2x).
• python -m pytest test_autolens/ — 538 passed, 0 failed.

Context

CP-1 of the inference-methods programme (planning pass 2026-08-17): every PositionsLH benchmark depends on the penalty matching its documented form. Pending-release: merge is a human act.

🤖 Generated with Claude Code