fix: enforce prior bounds in the NumPy-path objective (UniformPrior, LogUniformPrior)

[Jammy2211]

Summary

Restores the prior-support enforcement lost in release 2025.10.16.1, when assert_within_limits / PriorLimitException were removed (#1155, part of the JAX jit-compat cleanup) without a replacement on the NumPy path. Since then UniformPrior.log_prior_from_value returned 0.0 unconditionally and LogUniformPrior returned a finite -log(value) above its upper_limit, so the searches that form a log posterior on the NumPy path (Emcee, Zeus, Drawer, LBFGS) never penalised an out-of-support sample — an Emcee walker on a poorly-constrained parameter under UniformPrior(-0.1, 0.1) escaped to |value| ≈ 1e14, with 100% of accepted samples outside the box. Full reproduction, git archaeology, and the fix-shape decision are on #1489.

This is fix shape A from the issue: strict bounds in the priors themselves, exactly mirroring the JAX path (made strict in 2e3540771, which created the asymmetry). Shape C (defaulting LBFGS to ClipperPriorBox) is deliberately not included — orthogonal follow-up.

API Changes

No symbols added or removed — a behaviour restoration:

• UniformPrior.log_prior_from_value (NumPy path): 0.0 inside [lower_limit, upper_limit], -inf outside (was: always 0.0). Limits inclusive, matching JAX.
• LogUniformPrior.log_prior_from_value (NumPy path): -inf everywhere outside [lower_limit, upper_limit] (was: -inf only for value <= 0; finite -log(value) above upper_limit and below lower_limit).
• Fitness.log_likelihood_from: skips the log-prior subtraction when the prior sum is non-finite, so a rejected point inverts to -inf rather than NaN.
• Behavioural consequence: Emcee/Zeus reject out-of-box proposals (walker stays put); an unclipped LBFGS step out of the box now sees an infinite objective on the NumPy branch, as it always did on JAX.

See full details below.

Test Plan

• New test_autofit/mapper/prior/test_prior_bounds_1489.py: scalar + array bounds for both priors (inclusive limits, no RuntimeWarning), -inf figure of merit for an out-of-box vector, NaN-free log_likelihood_from inversion, and an end-to-end Emcee containment test (guaranteed deterministic: walkers start in-box and every out-of-box proposal is rejected).
• test_prior.py updated where it pinned the unbounded behaviour; test_clipper.py LBFGS end-to-end updated (its foil asserted the clipper-less fit escapes the box — that escape was the bug).
• Full suite: 1734 passed, 8 skipped; 2 failures (test_messages.py::test_beta, test_nautilus.py::test__single_core_builds_no_pool) are missing-optional-dependency environment failures that reproduce identically on untouched main.

Measured before/after (same script, same seed)

	before (main @ fe9f813)	after (this PR)	autofit 2025.5.10.1 (pre-regression)
Well-constrained 3-param Gaussian, posterior median (c, n, σ)	50.016, 24.562, 4.832	50.018, 24.580, 4.839	—
… posterior std	0.117, 0.487, 0.115	0.116, 0.487, 0.115	—
Unconstrained param under UniformPrior(-0.1, 0.1): samples outside box	630 / 630	0 / 540	0 / 630
… sampled min/max	−5.5e14 / 6.0e14	−0.0997 / 0.0999	−0.0992 / 0.0999

The well-constrained posterior is statistically unchanged (median shifts ~0.02 against σ ≈ 0.1–0.5); the containment matches the pre-regression release.

Full API Changes (for automation & release notes)

Removed

• (none)

Added

• (none)

Changed behaviour

• autofit.mapper.prior.uniform.UniformPrior.log_prior_from_value(value, xp=np) — NumPy path now returns -inf for value outside [lower_limit, upper_limit] (previously 0.0 unconditionally). JAX path unchanged.
• autofit.mapper.prior.log_uniform.LogUniformPrior.log_prior_from_value(value, xp=np) — NumPy path now returns -inf for any value outside [lower_limit, upper_limit] (previously only value <= 0). JAX path unchanged.
• autofit.non_linear.fitness.Fitness.log_likelihood_from — returns -inf (not NaN) for a figure of merit produced by an out-of-support vector.

Migration

• Emcee / Zeus / Drawer / LBFGS results on models whose posterior mass genuinely pressed against a UniformPrior/LogUniformPrior bound may change: chains now respect the declared box. This restores pre-2025.10.16.1 behaviour; no code changes required.
• A fit that relied on samples leaving a declared box should widen the prior instead.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DP4nPgsLuB2czRjrV4qcSk

---

Generated by Claude Code