fix: OptimisationState limits guard, VariableData.any, redundant limits overrides

[Jammy2211]

Overview

Three tidy-ups PyAutoFit#1527 left behind, plus one live bug found while doing the first.

VariableData.any reduced through var_all, so it meant "is there a variable whose elements are all True" rather than "is any element True". OptimisationState.valid asks (parameters < lower_limit).any() — so a parameter vector with some components outside their limits was reported valid, and only a variable violating on every component was caught. MeanField's valid.any() under-reported the same way.

That bug, not the truthiness guard the #1527 follow-up list named, is why the Laplace limits check under-enforced. It surfaced only because rewriting the guard needed a test, and OptimisationState.valid had no test coverage at all — the process lesson from #1477, third time it has paid out in this lineage.

Plan

• Fix VariableData.any to reduce through var_any, restoring the limits check in OptimisationState.valid and the validity count in MeanField.
• Rewrite OptimisationState.valid's guards from truthiness to is not None, so "was a limit supplied?" is stated rather than implied by VariableData happening to be a dict.
• Delete the three limits overrides fix: LogGaussianPrior declares its own (0, inf) support #1527 made exact duplicates of the base, once the float() coercion difference is measured rather than assumed.
• Cover both OptimisationState.valid and VariableData.any, which had no tests, and verify the new tests by inversion.

Detailed implementation plan

Work Classification

Library — PyAutoFit only. No workspace changes.

Affected Repositories

• PyAutoFit (primary)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	claude/loggaussian-prior-support-ngh59x (from main @ 34d6dff)	clean

Branch: claude/loggaussian-prior-support-ngh59x

Worktree root: none — this ran web-github against a direct clone at /home/user/pyautofit.

Implementation Steps

1. autofit/mapper/variable.py — VariableData.any dispatches to var_any, not var_all. The other four .any() call sites in the library (utils.py:539, multi_start_gradient/search.py:468, latent.py:274, and the numpy paths) are on numpy arrays and are untouched; the only VariableData consumers are line_search.py and mean_field.py:501, both of which want a real any.
2. autofit/graphical/laplace/line_search.py — valid's two guards become is not None, with a docstring recording why truthiness worked only by accident.
3. autofit/mapper/prior/{uniform,log_uniform,truncated_gaussian}.py — delete the limits overrides and the now-unused Tuple imports.
4. test_autofit/graphical/test_optimisation_state_valid.py — new; 8 tests over valid (no limits, empty limits, inside, below, above, on-the-bound, a 0.0 lower limit) and VariableData.any / .all().

Key Files

• autofit/mapper/variable.py:396 — the any/all reduction.
• autofit/graphical/laplace/line_search.py:106-127 — the limits guard.
• autofit/mapper/prior/abstract.py:360 — the base Prior.limits fix: LogGaussianPrior declares its own (0, inf) support #1527 made canonical.

Measured, not assumed

• The limits deletion is not a type change. The base coerces with float() and the overrides did not, which looked like a downstream risk via TruncatedGaussianPrior(mean, sigma, *limits) in prior passing. All three priors already store Python floats, so the coercion is a no-op; and under jax.jit a prior never reaches limits at all, because tree_unflatten → __init__ calls float() on the tracer and raises ConcretizationTypeError first — with or without this change.
• The guard rewrite is behaviour-preserving. The only case whose guard differs is the empty VariableData: falsy under the old form, so the comparison was skipped; run under the new form, where an empty comparison's .any() is False. valid returns True either way.
• The tests fail without the fix. Reverting var_any to var_all fails 3 of the 8.

Validation

• Full suite: NUMBA_CACHE_DIR=… MPLCONFIGDIR=… python3 -m pytest -x -q test_autofit/ — 2186 passed, 36 skipped (baseline on main: 2178 / 36).
• Readiness gate: PyAutoHeart is not present in this session, so the documented fallback applies — per-repo pytest -x as the gate, GREEN.

Scope note

This covers two PyAutoMind prompts, against the usual one-prompt-one-PR rule, because both are the same cleanup left by #1527 and the second is three line deletions:

• draft/refactor/autofit/optimisation_state_limit_guard_truthiness.md
• draft/refactor/autofit/redundant_prior_limits_overrides.md

It also widens past both to fix VariableData.any, without which the guard being tidied does not actually work. Happy to split if a reviewer prefers.

Not done here

• The check_limits EP question — the prior declares (0, inf) while its message stays at ±inf. Measured: EP is not wrong today, because the message's own density returns a clean -inf at negative values rather than NaN. Tracked in draft/research/graphical_ep/transformed_message_declares_support.md.
• message.logpdf(0.0) is -1.798e308 where the prior says -inf. Noted, unverified, out of scope.

Origin

Follow-ups 1 and 2 from complete/2026/08/loggaussian-prior-declares-own-support.md (this repo's #1526 / #1527).

[Jammy2211]

Shipped.

• PR fix: OptimisationState limits guard, VariableData.any, redundant limits overrides #1532 MERGED as 6e2d8c8 (merge commit), head 4c0f79b, +167/-20 over 6 files.
• CI green on every leg: unittest (3.12) 4m04s, unittest (3.13) 3m15s, unittest-nojax 2m20s, docs / docs-build 1m36s.
• Full suite 2186 passed / 36 skipped (baseline on main: 2178 / 36).
• Label pending-release — not released; the merged PR sits in the pending-release queue, so anything on a PyPI wheel does not have this.

What landed

File	Change
autofit/mapper/variable.py	VariableData.any reduces through var_any, not var_all
autofit/graphical/laplace/line_search.py	valid's guards → is not None
autofit/mapper/prior/{uniform,log_uniform,truncated_gaussian}.py	three duplicate limits overrides + unused Tuple imports removed
test_autofit/graphical/test_optimisation_state_valid.py	new, 8 tests

The load-bearing one is VariableData.any. It is a behaviour change, not a tidy-up: OptimisationState.valid now rejects a parameter vector with some components outside their limits, where before only an all-components violation was caught, and MeanField's validity count now reports partial validity. Both are corrections, and both were previously untested.

Follow-ups from this work

Neither is filed as a prompt yet:

• message.logpdf(0.0) is -1.798e308 (negative float max) where LogGaussianPrior.log_prior_from_value(0.0) is -inf. Finite-but-astronomically-negative behaves differently from -inf under any isfinite test. Unverified, cheap to check.
• The check_limits EP question — the prior declares (0, inf) while its message stays at ±inf, so MeanField.lower_limit hands OptimisationState a -inf for a strictly positive parameter. Measured during this work: EP is not producing wrong results today, because the message's own density returns a clean -inf at negative values rather than NaN — the limits check is redundant for that prior, not load-bearing. Tracked in PyAutoMind draft/research/graphical_ep/transformed_message_declares_support.md.

---

Generated by Claude Code