fix(priors): correctness batch — crashes, log-partition, projections

[Jammy2211]

Overview

Phase 1 of the priors/messages correctness batch surfaced by the EP framework review and re-verified on main @ 0f26ff2d8 (reproductions in #1330, decisions resolved on the hub #1331). This is the safe batch: crash fixes, a message log-partition correction, projection guards, and a normalisation-contract cleanup — all EP-isolated or numerically inert for standard non-linear searches. The behaviour-changing width-modifier pair (Decisions 2 + 5) is deliberately split into a separate Phase 2 follow-up because it changes prior-passing widths in search chaining and needs careful ordering + a regression test.

Resolves the awaiting-input state on #1330/#1331 and unblocks the #1332 EP fix batch.

Plan

• Overwriting a prior over-writes the upper / lower limits and resets them to +/- inf #1 LogGaussianPrior.with_limits crashes on first call — drop the invalid lower_limit/upper_limit kwargs (log-Gaussian support is (0, ∞)); also fixes _new_for_base_message.
• model_mapper does not return PpriorLimitException, but Raises it #2 UniformPrior.logpdf fails on array input — vectorise with xp.where, keeping the scalar result bit-identical.
• Backup results in a non-sim link directory. #4 TruncatedNormalMessage.log_partition is missing the Gaussian term — via the generic exponential-family interface the pdf integrates to 2.27, not 1.0. Add the Gaussian log-partition (as NormalMessage already does). This is the path the EP machinery consumes; sampling / log_prior_from_value are unaffected.
• Output results of classic grid search #10 FixedMessage.logpdf_cache is an unbounded class-level dict that returns an aliased mutable array — drop the cache, return xp.zeros_like(x).
• Decision 1 — Beta projection clamp is a no-op: raise on negative α/β from inv_beta_suffstats rather than silently patching to 0.5.
• Decision 3 — GammaMessage.from_mode(mode, V) builds the wrong variance: preserve mean + variance (α = m²/V, β = m/V), consistent with the Normal family.
• Decision 4 (Option A) — log_prior_from_value normalisation: document the drop-additive-constants contract, bring TruncatedGaussianPrior into line (it currently returns the fully-normalised density), and add an optional per-prior log-normaliser hook for evidence/Bayes-factor users. No posterior-shape or nested-sampling change (constants cancel in the MCMC ratio; nested samplers use the unit-cube ppf).

Not in this issue — Phase 2 follow-up: Decision 5 (RelativeWidthModifier → abs(mean) + configurable floor) landed with/before Decision 2 (NormalMessage strict σ > 0), plus a mean = 0 chained-parameter regression test. Filed separately as this issue nears shipping.

Detailed implementation plan

Affected Repositories

• PyAutoLabs/PyAutoFit (primary)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main	clean

Suggested branch: bug/priors-messages-fixes

Parallel disjoint PyAutoFit claim alongside ep-graphical-docs (#1334, autofit/graphical, awaiting-merge) and ep-diagnostics (#1335, autofit/graphical). This batch lives in autofit/mapper/prior/ + autofit/messages/ — zero file overlap with the graphical subtree.

Implementation Steps

1. autofit/mapper/prior/log_gaussian.py — with_limits / _new_for_base_message: drop lower_limit/upper_limit kwargs.
2. autofit/mapper/prior/uniform.py (+ autofit/messages/normal.py UniformNormalMessage) — vectorise logpdf with xp.where; assert scalar path unchanged.
3. autofit/messages/truncated_normal.py — add Gaussian log-partition term; verify generic exp-family pdf integrates to 1.0.
4. autofit/messages/fixed.py — FixedMessage: remove class-level logpdf_cache, return xp.zeros_like(x).
5. autofit/messages/ (inv_beta_suffstats) — raise on negative α/β.
6. autofit/messages/gamma.py — GammaMessage.from_mode: α = m²/V, β = m/V.
7. autofit/mapper/prior/*.py — Decision 4: docstring contract (constants dropped), TruncatedGaussianPrior.log_prior_from_value drops constants, add optional log-normaliser hook.

Key Files

• autofit/mapper/prior/log_gaussian.py, uniform.py, truncated_gaussian.py
• autofit/messages/truncated_normal.py, normal.py, fixed.py, gamma.py, Beta suff-stats

Testing

Full test_autofit/ suite (numpy-only unit tests) plus per-bug unit tests: with_limits round-trip, uniform array logpdf, truncated log_partition → pdf 1.0 integration, fixed-cache aliasing, beta-projection raise, gamma from_mode mean/var, log_prior drop-constants contract. No public API removals. Downstream workspace impact: none expected.

Original Prompt

Click to expand starting prompt

Batch from decision hub #1331 (EP framework review Phase 0/1 follow-up). Per-bug detail: PyAutoMind/bug/priors/01-14; Fable verdicts on #1330; maintainer decisions resolved on #1331. Autonomy: supervised (--auto).

[Jammy2211]

Shipped at PR-open — #1345

All 7 Phase-1 fixes are implemented, verified, and up as PR #1345 (pending-release).

Verification story:

• Full test_autofit/ suite: 1437 passed, 14 skipped (was 1422/14 on main — the delta is the new regression file + contract-updated truncated tests).
• Implemented on Opus, then independently re-verified on Fable (model switched mid-session at James's request): the truncated-normal exp-family logpdf now equals scipy.truncnorm.logpdf pointwise to 1e-9 (including the half-bounded prior-passing geometry), the EP product log_norm telescope identity holds against direct quadrature (the incomplete cumulant was silently breaking it), GammaMessage.from_mode reproduces requested mean+variance exactly, and all five log_normalisation hooks recover the scipy-normalised densities.
• Targeted smoke: autofit_workspace_test/scripts/jax_assertions/priors_xp_dispatch.py PASS (directly exercises the edited xp-dispatch paths); prior_correctness/emcee_gaussian_bias_check.py running now — result to follow.
• Heart: YELLOW on the same 6 pre-existing organism-wide reasons as recent ships (none touch this diff); acknowledged live in-session.

Two discoveries during implementation (both disclosed in the PR body):

1. LogGaussianPrior._new_for_base_message was doubly broken — beyond the invalid kwargs, it called a nonexistent self.instance().id. It also has zero callers (dead code). Repaired fully rather than half-fixed.
2. inv_beta_suffstats's internal np.linalg.solve is shape-fragile for multi-element batch inputs — pre-existing and orthogonal to the Decision-1 guard; left untouched, noted for the EP fix-batch follow-up.

Next: merge is human (#1345 carries the --auto validation checklist). Phase 2 — the width-modifier pair (Decision 5 abs(mean)+floor landed with/before Decision 2 strict σ>0, plus a mean=0 chained-parameter regression test) — gets filed as its own issue now that Phase 1 is at PR-open, per the sequencing analysis (it's the only leg with real sampling-behaviour impact).

[Jammy2211]

Smoke complete: emcee_gaussian_bias_check.py PASS as well — a real Emcee fit through log_prior_from_value with unbiased posterior recovery, which empirically confirms the D4 claim that the dropped constants cancel in the Metropolis ratio. All four ship-gate legs green: tests 1437/14 · smoke 2/2 · review CLEAN · Heart YELLOW-acked.

[Jammy2211]

Merged — c0b6c94b8 via #1345 (merged-unchanged; pending-release). Phase 2 (Decisions 5+2, the width-modifier pair) is being filed as its own issue now — link to follow on decision hub #1331. Closing this Phase-1 tracker.