open-problem

Status: resolved (2026-11-12). One defect: round_trip threw on a precision vector of element type Any, which implicit_infer accepts; fixed, with a regression test. The other finding was a list of untested edge cases, all of which Mistral ran and found working; three of them are now tests. Related: Feature Backlog (F3), LP-029 Findings of the review of P17 and P18.

Sources: a review of the uncommitted round_trip (F3) by Mistral (vibe --agent review), run with Julia in the helper worktree Lenticulum.jl-mistral (2026-11-12). Its answer is in ../Lenticulum.jl-Review-2026-11-12/REVIEW-f3-from-mistral.md (the first run failed on a server error of the API before its first turn); its scripts are scratch/f3/any_rho.jl, scratch/f3/probe.jl, scratch/f3/probe2.jl and scratch/f3/probe3.jl in that worktree. code: lib/VariationalDiffusion.jl/src/answers.jl (round_trip); test: lib/VariationalDiffusion.jl/test/runtests.jl, “round trip: answers on the relation stay, projections walk back”.

Problem. round_trip(m, sol, z₀, ρ, ps, st) reverses a query (frees the hard clamps, clamps the rest at the answer, solves back from the answer) and returns how far the clamps move, Inf when the reversed solve is not a converged stable root. It must accept every query implicit_infer accepts.

Mistral ran the testset (10 of 10 at the time) and the whole VariationalDiffusion suite (899 of 899, Aqua 11 of 11). It probed element types (Float16, Float32, Float64, Int, Real, Number, a union, Any), views, an all-clamped query, soft evidence, a ProductRelation, a one-dimensional relation, two hard coordinates, a non-energy MLP relation on the Newton path, and kwargs passed through.

#findingright?what was donetest
F1round_trip built the reversed precision with zero(eltype(ρ)), a MethodError for ρ::Vector{Any} and views of oneright: reproduced (the extended testset failed with that MethodError before the fix)the reversed precision takes the float type of the answer’s stateAny[Inf, 0.0], a view and Float32 in the testset
F2The testset had no case for an all-clamped query, soft evidence, a ProductRelation, Float32, integer ρ, viewsright as a gap; Mistral found the code working on eachsoft evidence, Any, a view and Float32 added. Checking soft evidence showed that the evidence pulls the answer off the relation, and the round trip measures that pull (0.011, 0.12, 0.23 for precisions 0.1, 0.5, 2 at ); the reversed answer lies on the mode circle. This is now asserted and stated in the docstringin the testset
F3(checked, no defect) the Inf case of the testset does reach the stability branch: the reversed solve from the centre converges at iteration 0 with stable = false—nothing—

Leaning / resolution. F1 fixed, F2 partly turned into tests. Not tested in the suite: the ProductRelation and the Krylov path through round_trip (Mistral ran them in its scratch scripts only), and device arrays (no device-specific code in round_trip).