Status: resolved (2026-11-16). The review found no defect in P21 (ForwardDiff through a relation factor’s flat path with only η dual, and
NonlinearFactorat a dual expansion point). It ran no probe: every Julia command was refused by its approval callback. Its seven probes were run afterwards; their failures were errors of the probes. The gaps it named were probed again by hand and passed, and two of them (the switching factor, two flat inputs) are tests now. One behaviour was found on the way and documented, not changed: a partial of η along null(Λ) is dropped. Related: Feature Backlog (P21, P25), gaussian “Dual inputs on the flat path”.
Sources: a review of the uncommitted P21 change by Mistral (
vibe --agent review,--max-price 3, $1.12 at list price), in the helper worktreeLenticulum.jl-mistralat5eac9e5ewith the diff applied; the brief, the diff, the answer file, its probes with their outputs, and the repair probeprobe_mine.jlin../Lenticulum.jl-Review-2026-11-16/p21/. code:lib/Lenticulum.jl/src/gaussian.jl(_improper_split_dual),lib/Lenticulum.jl/src/nonlinear.jl(_jacobian),lib/LenticulumCore.jl/src/extended_gaussian.jl(_flat_split); tests:lib/Lenticulum.jl/test/forms_crossing.jl.
Problem. The requirement given to the reviewer: with a flagged input of dual η and a precision
without derivative, every relation factor’s message works, equals the value path’s and has its
central differences as partials; a precision with nonzero partials throws an ArgumentError;
NonlinearFactor at a dual expansion point likewise, both methods, with and without a jacobian;
the Float64 path and extended’s refusals are unchanged.
| # | finding or probe | right? | what was done | test |
|---|---|---|---|---|
| R1 | SwitchingConstraintFactor not tested | a gap, not a defect: probe_mine.jl passes (40 instances, flat x to z and to k, 212 checks) | added as a testset | ”P21: the switching factor and two flat inputs” |
| R2 | two flat inputs into one LinearConstraintFactor, several flat directions | a gap: passes (40 instances, 64 checks) | added (same testset) | same |
| R3 | flat input of dimension 1, η exactly 0, an Int Jacobian, Dual tags other than Nothing | gaps: η with value 0 and nonzero partials, the Int Jacobian and a tagged ForwardDiff.derivative pass in probe_mine.jl. Dimension 1: a flat input there is flat everywhere, the _flat_part path, not this change | none | probe_mine.jl only |
| P1–P3 | its switching probes | failed on their own errors: they expect the value path’s source to be a GaussianBelief (it is an ExtendedGaussian) and call a CategoricalBelief constructor that does not exist | — | — |
| P4, P7 | ”zero-partial duals accepted”, “mixed partials”, “extended refusals” | errored with “information vector outside the range”: the probes draw η at random, outside range(Λ), which extended rightly refuses (requirement 5) | — | — |
| P5 | NonlinearFactor derivatives differ from central differences | not reproduced: the probe’s msg(t) draws a new direction randn(rng, nin) on every call, so the four evaluations move along different directions | — | “P21: NonlinearFactor’s message at a dual expansion point” (160 checks) |
| P6 | several flat directions: “both messages non-trivial” | the probe’s expectation: with two flat directions both paths give TrivialBelief() here (agreeing); R2 checks the kinds agree | — | — |
| X1 | found while repairing: a partial of η along null(Λ) is dropped (derivative 0), where the value path refuses the input at every t ≠ 0 | right, as a behaviour; the factors’ own messages keep η in range(Λ) for every t, so only a hand-built input reaches it | documented in gaussian “What it does not do” | not tested |
Leaning / resolution. No code change from the review; two testsets’ worth of coverage added (R1, R2). The review agent’s Julia commands were refused this time although the brief gave the exact command line that worked for P23; why was not looked into.