Status: resolved (2026-11-12). One defect found and fixed, a ForwardDiff crossing the first version of P1 left open. Also: one test that did not build what its comment said, one header that overclaimed (a pre-existing limit, now
@test_broken) and stale note text. One finding is not a defect. Related: Feature Backlog (P1, P21), LP-027 Findings of the review of the I18 plans.
Sources: a review of the uncommitted P1 change by Mistral (
vibe --agent review), run with Julia in the helper worktreeLenticulum.jl-mistral(2026-11-12). Its answer is in../Lenticulum.jl-Review-2026-11-12/P1-from-mistral-2.md(the first run stopped at its turn limit), its scratch scripts are underscratch/p1/in that worktree (probe_req1.jl,probe_req2.jl,probe_stack.jl,probe_nonlinear.jl,probe_jlarrays.jl). code:lib/Mycelium.jl/src/gaussian.jl(combine,belief_distance,can_damp,_dampfor anExtendedGaussianand aGaussianBelief),lib/Lenticulum.jl/src/gaussian.jl(_likelihood_message); tests:lib/Mycelium.jl/test/forms_crossing.jl,lib/Lenticulum.jl/test/forms_crossing.jl; notes:lib/Lenticulum.jl/src/gaussian.md“Where the two forms meet”,lib/Mycelium.jl/src/gaussian.md“Extended Gaussians”.
Problem. P1 asked for tests where the two forms of a flat likelihood meet (a flagged
GaussianBelief, the form under ForwardDiff, and an ExtendedGaussian, the form on values). The change
reviewed fixed two defects those tests found (_likelihood_message built the extended form of a dual η;
damp was a silent no-op on a mixed pair). The brief gave Mistral the requirement, not the reasoning.
| # | finding | right? | what was done | test |
|---|---|---|---|---|
| F1 | A variable that hears a plain extended message (a rank-deficient factor with plain parameters) and a dual flagged one (the factor being differentiated) throws in infer! under ForwardDiff: combine and belief_distance converted the GaussianBelief side with extended, which has no method for duals | right. Reproduced by Mistral (probe_req1.jl, R1b) and by the extended test: against the code before the fix the two-factor inference errs in all 10 instances with MethodError: no method matching svd!(::Matrix{ForwardDiff.Dual...}) | a mixed pair whose GaussianBelief holds non-float numbers is taken in canonical form when the extended side has no hard directions (_canonical); with hard ones combine and belief_distance throw an ArgumentError and can_damp is false. Plain-value pairs are unchanged | Lenticulum test/forms_crossing.jl, “ForwardDiff through an inference with flat backward messages” (two factors on x, derivative against the dense joint); Mycelium test/forms_crossing.jl, “a mixed pair with dual numbers goes through the canonical form” (60 random pairs, 1 080 checks) |
| F2 | The Mycelium test said “a mixed pair with dual numbers” but built Real arrays | right | replaced by the testset of F1 with ForwardDiff duals | as F1 |
| F3 | The Lenticulum test’s header claimed every parameter dual for all three factors; NonlinearFactor with a dual expansion point throws earlier, in _jacobian (Float64(::Dual)) | right; the limit predates P1 | header corrected; the limit recorded as @test_broken and in P21 | Lenticulum test/forms_crossing.jl, “known limit, NonlinearFactor’s backward message at a dual expansion point” |
| F4 | gaussian.md still said the crossings are untested and that _likelihood_message checks “the element type”; the backlog listed P1 as proposed | right for the worktree’s copy, which did not have the note changes; “the element type” was still stale in the main checkout | corrected | — |
| F5 | With JLArrays, the new mixed damp returns a host ExtendedGaussian, where before P1 it returned the device message undamped | right as a description, not a defect: ExtendedGaussian is a host type (Vector, Matrix), and combine of the same pair already returned it on the host before P1 | nothing changed | — |
Leaning / resolution. F1 was a real defect of the same kind as the one P1 fixed in
_likelihood_message: a place that decided by representation and forgot that under ForwardDiff
only some arrays carry duals. F2–F4 corrected tests and notes.
Not checked: the distance near the rank threshold on a run (lib/Mycelium.jl/src/gaussian.md); a
mixed pair where Λ, not only η, holds duals is covered by the same _extendable test but was not
run separately.