fix(priors): width-modifier safety + strict sigma>0 (Phase 2 of #1331)

[Jammy2211]

Overview

Phase 2 of the #1331 priors/messages batch — the behaviour-changing pair, split from Phase 1 (#1344, merged via #1345) because it alters prior-passing widths in search chaining (SLaM-style pipelines). Verdicts: bug/priors/06 + 08 on #1330; decisions resolved on #1331 (Decision 5: minimal — abs(mean) + optional configurable absolute floor; Decision 2: reject σ ≤ 0 strictly).

Sequencing constraint (from the Phase-1 impact analysis): the width fix (D5) lands with the strict σ check (D2) in one PR so today's silent σ=0 degeneracy (a mean=0 chained parameter freezes into a delta function) never becomes an unexplained hard crash: where a computed width is still ≤ 0, prior passing raises a clear, parameter-named PriorException pointing at the width-modifier config — not a deep MessageException.

Plan

• RelativeWidthModifier: sigma = value * abs(mean), plus opt-in absolute_floor (YAML-configurable; engages when value*abs(mean) falls below it). Negative means no longer produce negative sigma (today 0.5 * -1.0 = -0.5 flows silently into GaussianPrior and flips value_for).
• Prior passing (gaussian_prior_model_for_arguments width application): if the final width ≤ 0 (e.g. mean=0, no floor), raise PriorException naming the parameter path with remediation guidance (set an absolute width modifier or an absolute_floor).
• NormalMessage.__init__: enforce σ > 0 (numpy path; JAX defers to NaN propagation — the existing assert_sigma_non_negative helper's lax.cond branch is broken by design and is retired). GaussianPrior(mean, sigma<=0) and GaussianPrior.with_limits(5, 5) become loud.
• TruncatedNormalMessage: extend its existing < 0 rejection to <= 0 so the two classes agree (strictly).
• Regression tests: mean=0 chained parameter (clear error without floor; correct width with floor), abs(mean) on negative means, floor engagement + config round-trip, strict ctor rejection on both message classes.
• Full test_autofit/ suite; pre-flight grep for tests that deliberately construct σ ≤ 0 messages (they may encode point-mass behaviour — FixedMessage is the sanctioned path for that).

Detailed implementation plan

Affected Repositories

• PyAutoLabs/PyAutoFit (primary)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main (post-#1345, c0b6c94)	clean

Suggested branch: feature/prior-width-safety

Parallel disjoint PyAutoFit claim alongside ep-graphical-docs (#1334) + ep-diagnostics (#1335), both autofit/graphical — this work is autofit/mapper/prior/width_modifier.py, autofit/mapper/prior_model/abstract.py (width application site), autofit/messages/{normal,truncated_normal}.py. Zero overlap.

Implementation Steps

1. autofit/mapper/prior/width_modifier.py — RelativeWidthModifier.__init__(value, absolute_floor=None); __call__ returns max(value * abs(mean), floor) when floor set, else value * abs(mean); from_dict/dict round-trip the optional key.
2. autofit/mapper/prior_model/abstract.py — after width = width_modifier(mean) (and the explicit r/a paths), raise exc.PriorException naming the prior path if width <= 0.
3. autofit/messages/normal.py — numpy-path strict sigma > 0 enforcement in NormalMessage.__init__; retire the broken assert_sigma_non_negative JAX branch.
4. autofit/messages/truncated_normal.py — < 0 → <= 0.
5. Tests per the plan bullet above (numpy-only).

Behaviour changes (user-visible, disclosed)

• Chained fits with parameters whose posterior median is negative get a positive passed width (previously negative sigma, silently sign-flipping value_for). Sampling of affected chained fits changes — for the better.
• Chained fits with a mean=0 unconfigured parameter now fail loudly with guidance (previously silently froze the parameter via a delta-function prior).
• GaussianPrior/NormalMessage with σ ≤ 0 raise at construction.

Autonomy: supervised (--auto, human instructed live 2026-07-10 after Phase-1 merge; heart-ack 6-reason set recorded in active.md this chain).

🤖 Generated with Claude Code

[Jammy2211]

Supervised-mode disclosure: Decision 2 adjusted on new evidence (strict → permissive)

The #1331 recommendation for Decision 2 was reject σ ≤ 0 strictly. Implementing it broke 16 tests + 4 errors across three subsystems, and the failures are not stale tests — they pin a deliberate, load-bearing design idiom:

1. Latent variables: non_linear/samples/util.py::simple_model_for_kwargs wraps every known latent value as GaussianPrior(mean=value, sigma=0.0) — the σ=0 point-mass carrier is how latent quantities become model parameters (10+ tests).
2. Point-mass projection: message.from_mode(x, covariance=np.zeros(())) is the sanctioned zero-covariance projection (test_correspondence.py::test_from_mode).
3. Pinned API semantics: test_result.py::test_model_relative asserts sigma == 0.0 for a zero-centred parameter through the explicit model_centred_relative(r=...) path.

#1331's own caveat asked exactly this: "the reviewer should confirm whether any current call site depends on σ == 0 constructing" — answer: three subsystems do. So this PR ships the permissive variant (the explicitly-listed alternative on #1331):

• NormalMessage rejects σ < 0 (the actual bug — negative sigma previously constructed silently with deceptive variance = σ² > 0 and sign-flipped value_for); σ = 0 permitted and documented as the point-mass idiom.
• TruncatedNormalMessage keeps its σ < 0 rejection — the two classes now agree, which was the real defect Decision 2 targeted.
• The new prior-passing guard (parameter-named PriorException with absolute_floor guidance) is scoped to the default width-modifier path — the silent-freeze bug from bug/priors/08. Explicit a=/r= widths keep their historical, test-pinned point-mass semantics at mean=0.

Everything else is unchanged from the plan: RelativeWidthModifier → abs(mean) + opt-in absolute_floor (config round-trip included), mean=0 chained regression tests, negative-median chained parameters now get positive widths.

If the strict variant is still wanted, the follow-on work is a first-class point-mass representation for the latent machinery (migrating simple_model_for_kwargs off degenerate Gaussians) — happy to file that as a separate design prompt. Validation state: targeted clusters 67/67 green; full suite re-running now, result on the PR.

[Jammy2211]

Shipped at PR-open — #1348

PR #1348 open (pending-release, --auto validation checklist in body — note the explicit human check on the D2 permissive call).

• Full suite: 1447 passed / 14 skipped (16-fail blast radius under strict D2 fully resolved by the disclosed permissive adjustment; all previously-failing clusters re-verified 67/67).
• Heart: YELLOW, identical 6-reason set to the in-session ack — no new reasons, gate passes.
• Smoke: priors_xp_dispatch PASS; emcee_gaussian_bias_check running, result to follow.
• Mind: library-shipped, awaiting-merge; autonomy-log row appended.

Merge stays human. With this and #1345 merged, the #1331 batch is fully landed except the parked design items (09/11/12/13/14).

[Jammy2211]

Smoke complete: emcee_gaussian_bias_check.py PASS — all four gate legs green for #1348: tests 1447/14 · smoke 2/2 · review CLEAN (D2 amendment disclosed above) · Heart YELLOW within the live ack. Awaiting human merge decision.

[Jammy2211]

Merged — cf0cc4bbb via #1348 (merged-unchanged incl. the disclosed D2 permissive amendment; pending-release). With #1345 + #1348 both in, the #1331 fix batch is fully landed. Closing.