[Jammy2211]

Summary

Closes #732. The Euclid initial_lens_model/vis_lp stage (two-basis MGE lens light, SIE + shear, MGE source) of RAL job 342398 wrote no latent output on all ten euclid_dr1_prelim tiles. LatentLens.variables built its per-sample instance with model.instance_from_vector(vector=parameters), whose default assertion check applies a Python not to a traced boolean. The latent engine evaluates variables under a per-sample jax.jit, so a model carrying any assertion (the ordered MGE bases added to vis_lp on 2026-09-08) raised TracerBoolConversionError on every sample; the engine swallowed each raise into a NaN row, the global mask found nothing, and the search logged only "no finite latent samples remained".

This adds latent_instance_from(model, parameters, xp): the NumPy path is unchanged (assertions checked, a violation raises FitException and is dropped as a NaN row); under JAX it skips assertions and threads xp, exactly as Fitness does on its JAX path. LatentLens.variables uses it, and the class docstring tells subclasses that build their own instance to do the same. The companion PyAutoFit PR makes the engine report the exceptions it swallows; the euclid pipeline's own two trace failures are filed in PyAutoMind as draft/bug/euclid/vis_lp_latent_euclid_jit_trace.md.

API Changes

Added one public helper, autolens.analysis.latent.latent_instance_from. No existing symbol changed; NumPy-path behaviour of LatentLens.variables is unchanged, the JAX path now succeeds where it silently produced NaN rows.
See full details below.

Test Plan

• test_latent_instance_from_numpy_checks_assertions — violation raises FitException, satisfied builds the instance
• test_latent_instance_from_jax_skips_assertions_under_jit — the default path raises TracerBoolConversionError under jax.jit, the helper traces
• test_latent_lens_variables_traces_under_jit_with_model_assertion — end to end on the 7x7 fixture with use_jax=True, every latent finite and equal to the NumPy evaluation
• Full suite: 621 passed, 1 skipped, 1 xfailed
• Reproduction on the euclid pipeline's simulated dataset: with this and the pipeline follow-up, jax.jit(LatentEuclid.variables) returns all 12 latents finite and equal to the eager values

Full API Changes (for automation & release notes)

Added

• autolens.analysis.latent.latent_instance_from(model, parameters, xp=np) — builds the instance a latent function is evaluated on; NumPy checks assertions, JAX skips them with ignore_assertions=True, xp=xp

Changed Behaviour

• LatentLens.variables under JAX: previously raised inside the engine's per-sample jax.jit for any model with an assertion (silently masked to NaN rows); now evaluates

Migration

• A LatentLens subclass that builds its own instance: replace model.instance_from_vector(vector=parameters) with latent_instance_from(model=model, parameters=parameters, xp=analysis._xp)

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FFSFgEBiZc7kTRutfK1RsN

---

Generated by Claude Code