open-problem

Status: resolved (2026-11-16). The review of P25 (ForwardDiff through a relation factor’s flat path when the flat directions move, at constant rank) found one defect: nested duals reached _range_complement and failed with a MethodError instead of the ArgumentError the requirement asks for. Fixed, with a regression test. A suspected defect in the covariance was the reviewer’s own wrong identity. Related: Feature Backlog (P25), gaussian “Flat directions that move”.

Sources: a review of the uncommitted P25 change by Mistral (vibe --agent review, --max-price 3, $1.18 at list price), in the helper worktree Lenticulum.jl-mistral at 8a6a9f41 with the diff applied; the brief, the probes and the answer file in ../Lenticulum.jl-Review-2026-11-16/p25/. code: lib/Lenticulum.jl/src/gaussian.jl, src/constraint.jl, src/switching.jl, ext/LenticulumForwardDiffExt.jl; tests: lib/Lenticulum.jl/test/forms_crossing.jl.

Problem. The requirement given to the reviewer: under first-order ForwardDiff, the flat path’s messages are differentiable when a flagged input’s precision is dual or the coefficients are dual, equal to central differences of the value path at constant rank; a tangent that changes the rank at first order, and nested duals, throw an ArgumentError; Float64 values unchanged.

The run stopped at its 100-turn limit with one entry in its answer file. The probes it left were run by the maintainer’s session (probe2_v3.jl, probe4_edge_cases_fixed.jl, probe6_final_check2.jl, probe7_nested_duals.jl, the last repaired as probe7_repaired.jl: its line 44 multiplied a 5×2 by a 2×5 matrix transposed).

#proberight?what was donetest
1_moving_flat_split: constant-rank tangent accepted, rank-changing refused, nested duals refusedpassed—P25 refusal testset
2_range_complement: constant rank, rank change, equal columns separated or kept parallel, log pdet against central differencespassed—P25 refusal testset
4full-rank Λ dual, Λ = 0 with a rank-changing tangent, η outside range(Λ), values equal to _flat_splitpassed—probe4_edge_cases_fixed.jl only (full-rank and η cases)
6”Σ is √Λ⁺”: the probe tested ΣΣᵀ = Λ⁺ next to Σ = Λ⁺; Σ = Λ⁺ to 7e-16not a defect: the second identity is wrong——
7nested duals in _range_complementright: MethodError (no svd for them), not ArgumentErrortyped float method and an ArgumentError fallback; the split’s message for nested duals made explicitP25 refusal testset (failed 20 times before the fix)

Not probed: the plate methods (factor_messages with shared metric) under moving flat directions, and inference through a graph with a dual A upstream of a flat path (only single messages are tested).

Reviewer’s tooling. The command line given to copy ran; the reviewer spent its turns on re-runs of debug variants (26 scripts) and did not append an entry after each probe as asked.

Leaning / resolution. Fixed (finding 7).