Priors & messages: 9 confirmed bugs — guidance wanted on 5 decisions

[Jammy2211]

An AI-assisted audit of autofit/mapper/prior/ and autofit/messages/ (follow-up to the LogUniform fix, #1329) found 9 real bugs plus refactor ideas. Every finding has now been independently re-verified on main @ 0f26ff2d8 — reproductions and full detail in #1330. This issue is the decision hub: maintainer guidance + contributor input wanted, especially from anyone with exponential-family / EP background (@rhayes777).

Will fix as one batch — no decision needed

• LogGaussianPrior.with_limits crashes on first call (TypeError: passes lower_limit/upper_limit kwargs the ctor doesn't accept); _new_for_base_message also broken. Fix: drop the kwargs — a log-Gaussian has support (0, ∞).
• UniformPrior.logpdf fails on array input (scalar-only boundary snap, acknowledged # TODO). Fix: vectorise with np.where.
• TruncatedNormalMessage.log_partition is missing the Gaussian term — via the generic exponential-family interface the pdf integrates to 2.27, not 1.0 (error factor exactly σ·exp(μ²/2σ²)). Sampling and log_prior_from_value are unaffected; the wrong path is the one the EP machinery consumes. Fix: add the Gaussian log-partition (as NormalMessage already does).
• FixedMessage.logpdf_cache is an unbounded class-level dict that also returns an aliased mutable array (mutating a result corrupts later calls). Fix: np.zeros_like(x), no cache.

Decisions wanted (my recommendation in bold)

1. Beta projection clamp is a no-op (inv_beta_suffstats clamps into a local that's immediately overwritten, so negative α, β escape with only a warning). When Newton–Raphson goes negative: raise (a failed projection shouldn't be silently patched to 0.5), or make the clamp work as originally intended?
2. NormalMessage accepts negative sigma (the assert is commented out; TruncatedNormalMessage rejects it, so the two disagree). Enforce reject σ ≤ 0 strictly, or reject only σ < 0 and permit a σ = 0 point-mass?
3. GammaMessage.from_mode(mode, V) builds the wrong variance — inversely wrong: request var 0.25, get 2.0; request 4.0, get 0.235 (α = 1 + m²·V has V upside-down). Which invariant should from_mode preserve: match mean + variance (α = m²/V, β = m/V — consistent with the Normal family), the α ≥ 1-guaranteed variant (α = 1 + m²/V), or solve the quadratic to match mode + variance exactly?
4. log_prior_from_value normalisation is inconsistent: Uniform/LogUniform/Gaussian/LogGaussian drop their additive constants; TruncatedNormal returns the fully normalised density. Posterior shape is unaffected either way, but evidence/Bayes-factor arithmetic and logpdf users see silently different conventions. Option A: drop constants everywhere, document it as the contract, add an optional per-prior log-normaliser hook — or Option B: fully normalised density everywhere (scipy convention, touches every prior + JAX paths)?
5. RelativeWidthModifier (the default prior-passing width, 0.5·mean) yields σ = 0 for means at zero and σ < 0 for negative means — silently, until (2) lands. Scope: minimal — use abs(mean) with an optional configurable absolute floor, mandatory floor with config migration, or a full width-modifier redesign?

Parked (design work, sequenced behind the EP framework review)

Property-based prior tests (after the fixes land), TransformedMessage semantics documentation, the single-source-of-truth density refactor, the Prior/Message hierarchy collapse, and bijector-library migration. These will be shaped by the EP statistics review now underway and get their own design issues later.

🤖 Generated with Claude Code

[Jammy2211]

Status update — fixes are landing.

• Phase 1 MERGED (c0b6c94b8 via fix(priors): correctness batch — crashes, log-partition, projections #1345, tracker fix(priors): correctness batch — crashes, log-partition, projections #1344 closed): the 4 no-decision fixes plus Decisions 1 (Beta clamp → raise), 3 (Gamma from_mode matches mean+variance), and 4 (Option A — drop-constants contract documented everywhere, TruncatedNormal brought into line, new Prior.log_normalisation() hook recovers the constant for evidence users). Statistics independently re-verified on Fable; full suite 1437/14; Emcee bias-check smoke confirms posteriors unchanged.
• Phase 2 filed: fix(priors): width-modifier safety + strict sigma>0 (Phase 2 of #1331) #1346 — Decisions 5 + 2 as one PR (the width-modifier pair): RelativeWidthModifier → abs(mean) + opt-in absolute_floor, strict σ ≤ 0 rejection on both message classes, and a parameter-named PriorException when a chained width still comes out ≤ 0. This is the only leg that changes sampling behaviour in chained fits (for the better); split out and sequenced deliberately.
• Still parked here (design work, post-EP-review as planned): property-based prior tests (09), TransformedMessage semantics doc (11 — code half shipped in Phase 1's docstrings), single-source density refactor (12), Prior/Message collapse (13), bijector migration (14).

[Jammy2211]

Resolved — all 9 bugs fixed, all 5 decisions taken (2026-07-10)

Recording the outcome so this hub reflects reality. Everything below is merged and verified live on main (2026-07-14).

Batch fixes (no decision) — PR #1345

• LogGaussianPrior.with_limits crash — fixed (bad ctor kwargs dropped; _new_for_base_message repaired)
• UniformPrior.logpdf array input — vectorised with np.where
• TruncatedNormalMessage.log_partition (pdf integrated to 2.27) — Gaussian cumulant added → integrates to 1.0; EP log_norm telescope identity holds
• FixedMessage.logpdf_cache aliasing — cache removed, returns fresh np.zeros_like(x)

Decisions — PRs #1345 (1, 3, 4) and #1348 (2, 5)

1. Beta projection clamp → raises ValueError on negative Newton–Raphson projection (no silent escape).
2. NormalMessage negative σ → σ < 0 rejected, σ = 0 permitted. Evidence-based deviation from the strict recommendation: strict σ ≤ 0 broke 20 tests across the deliberate σ=0 point-mass idiom (latent simple_model_for_kwargs, from_mode(covariance=0), model_centred_relative). The two message classes now agree. Strict would require a first-class point-mass representation first (separate design prompt on request).
3. Gamma from_mode → match mean+variance (α = m²/V, β = m/V), consistent with the Normal family. α < 1 now reachable when V > m² (accepted trade).
4. log_prior_from_value normalisation → Option A: drop additive constants everywhere as the documented contract, plus an opt-in Prior.log_normalisation() hook recovering the constant for evidence/Bayes-factor users.
5. RelativeWidthModifier → abs(mean) + optional absolute_floor; the default prior-passing path now raises a parameter-named PriorException on a non-positive computed width instead of silently freezing the parameter.

Regression coverage: test_priors_messages_fixes_1331.py (12) + test_prior_width_safety.py (11); full suite green (1447 passed / 14 skipped).

Leaving this issue open only for the parked design items (property-based prior tests, TransformedMessage semantics docs, single-source density refactor, Prior/Message hierarchy collapse, bijector migration), which are sequenced behind the EP framework review (#1332).