fix: LogUniform NumPy log-prior returns -inf for value<=0 (emcee NaN crash)

[Jammy2211]

Summary

Emcee's stretch move proposes physical values that can go non-positive for a
LogUniform-priored parameter. LogUniformPrior.log_prior_from_value computed
-np.log(value) on the NumPy path, which is NaN for value <= 0; that NaN
propagated into the summed figure-of-merit (fom = log_likelihood + Σ log_prior,
fitness.py) and crashed the search with ValueError: Probability function returned NaN.

This returns -inf (zero density → the proposed move is rejected) for
value <= 0, using the "double where" pattern so no log of a non-positive value
is ever evaluated (avoiding a NumPy RuntimeWarning). Positive values are
unchanged — the NumPy path stays unnormalised and unbounded, returning
-log(value); the JAX path is untouched.

Surfaced by PyAutoHeart #27 release-fidelity validation (run 28784914443); fixes
the autofit_workspace cookbook legs that failed with the emcee NaN crash.

Closes #1328.

API Changes

Behaviour-only: LogUniformPrior.log_prior_from_value(value) now returns -inf
(was NaN/+inf) for value <= 0 on the NumPy path. No signatures, symbols, or
defaults change; positive-value behaviour is identical. The only direct callers are
the autofit_workspace_test prior parity/regression gates
(jax_assertions/priors_xp_dispatch.py, prior_correctness/emcee_gaussian_bias_check.py),
which exercise only positive, in-bounds values and are unaffected.
See full details below.

Test Plan

• python -m pytest test_autofit/mapper/prior/ test_autofit/non_linear/ (918 + 250 pass)
• New regression test__log_prior_from_value__non_positive_returns_neg_inf
• Real emcee over a LogUniform model completes without "Probability function returned NaN"

Full API Changes (for automation & release notes)

Changed Behaviour

• autofit.LogUniformPrior.log_prior_from_value(value) — NumPy path returns -inf for value <= 0 (previously NaN from -log of a non-positive value, +inf at value == 0). Positive values unchanged (-log(value), unnormalised/unbounded). JAX path unchanged.

Migration

• None. Internal behaviour fix. Direct callers (autofit_workspace_test prior parity gates) exercise only positive, in-bounds values and are unaffected.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes an Emcee crash caused by LogUniformPrior.log_prior_from_value returning NaN on the NumPy path when proposed values are non-positive, by safely returning -inf instead (zero density → rejected move) without ever evaluating log on invalid inputs.

Changes:

• Update LogUniformPrior.log_prior_from_value(..., xp=np) to return -inf for value <= 0 using a safe “double where” pattern to avoid NumPy RuntimeWarnings.
• Add a regression test ensuring non-positive values return -inf and that no “invalid value in log” warning is emitted; confirm positive-value behaviour is unchanged.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
autofit/mapper/prior/log_uniform.py	Makes NumPy log-prior robust to non-positive proposals by returning -inf without evaluating invalid logs.
test_autofit/mapper/prior/test_prior.py	Adds regression coverage for the non-positive case and verifies warning-free evaluation plus unchanged positive behaviour.