Stop dist:'exponential' condemning itself on a scale it never estimates

[jrosskopf]

Recorded as a known limitation rather than fixed, on the grounds that the fix was a contract decision. Taking the decision — and the narrower of the two options.

The defect

Under dist: 'exponential' the AFT scale is held at 1 by the definition of the distribution. Its posterior sd is exactly zero, so ess returns 0.0, and the gate condemned the whole fit as degenerate — while every coefficient was estimated correctly and the other three distributions converged on the same data.

An agent reading anofox_bayes_is_actionable was told to ignore a fit that was fine. Same class as the two refusal-contract bugs fixed in #1: a verdict that doesn't match reality.

Measured through SQL, before → after: degenerate / false → converged / true.

Why it cannot be fixed by looking at the draws

ess returning zero for a parameter that never moves is deliberate — it is what stops a genuinely stuck sampler sailing through an ess >= 400 gate. And a stuck chain is indistinguishable from a constant one by its values alone. Inferring "structurally fixed" from the draws would hand the stuck chain the same exemption, trading a false refusal for a false pass. That is the worse trade.

So the family declares it

CompiledModel::fixed_params names the parameters a model reports but does not estimate. Default is empty — a family that declares nothing gets no exemption. censored_aft returns sigma under exponential and nothing under the other three, so a scale that genuinely fails to mix still condemns the fit it belongs to.

Nothing moves in the draws contract

sigma is still published, still exactly 1. A consumer joining on param = 'sigma' need not branch on dist — which is precisely why the alternative (omitting the row for this dist) was rejected: it would have changed a join surface to fix a grading bug.

The scope is asserted, not described

declaring_a_parameter_fixed_exempts_that_parameter_and_no_other grades one posterior with two dead parameters, three ways:

• exempt neither → degenerate
• exempt sigma → still degenerate for beta, and only beta is named
• exempt both → converged

The interesting case is telling two identical-looking parameters apart, which is exactly what the declaration exists to do.

Verification

cargo test 393 pass · clippy clean · make test 681 assertions · make test_scenario 57 · end-to-end SQL check of the before/after verdict.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}