Status: resolved (2026-11-13). No wrong result, crash or missing check: Mistral’s probes recomputed every field of
ridge_residualindependently (a 3-D Gaussian, the ring, an MLP with a non-symmetric Jacobian, a product relation, Float32, device arrays) and found them equal. Four findings were missing tests: mutants that drop the symmetrisation, test positivity on every eigenvalue, or never flag anInfdistance passed the testsets. Each now has a test that fails on its mutant. Related: Feature Backlog (F3), LP-031 Findings of the review of the query report, LP-030 Findings of the review of round_trip.
Sources: a review of the uncommitted change by Mistral (
vibe --agent review), run with Julia in the helper worktreeLenticulum.jl-mistral(2026-11-12), its answer written probe by probe as the brief asked (../Lenticulum.jl-Review-2026-11-12/ridge-from-mistral.md; probes inscratch/ridge/of that worktree). The worktree was at7277fe4f; I6, the commit between it and the reviewed base, changed neither VariationalDiffusion nor LenticulumCore. code:lib/VariationalDiffusion.jl/src/answers.jl(ridge_residual),src/report.jl; tests:lib/VariationalDiffusion.jl/test/runtests.jl, “ridge residual: …” and “query report: …“.
Problem. ridge_residual is the first-order distance from an answer to the ridge of the
smoothed density, from the field and the symmetric part of its Jacobian at the answer
(implicit §5); query_report reports it and flags :off_ridge.
| # | finding | right? | what was done | test |
|---|---|---|---|---|
| F1 | the symmetrisation is never exercised: the fixtures’ Jacobians are symmetric to , and a mutant taking eigen(Symmetric(J)) passed all 54 assertions | right (the code was correct on an MLP with a non-symmetric Jacobian) | a linear test field with antisymmetric; the expected values come from | the mutant fails 6 assertions |
| F2 | positivity is checked on the normal curvatures only, and no fixture has a mixed-sign Jacobian: a mutant checking every eigenvalue passed | right | the same field with , a saddle: finite at codim 1, Inf at codim 2 | the mutant fails 1 |
| F3 | :off_ridge on an Inf distance is reached (the ring’s centre given as the answer) but never asserted | right | asserted, also with ridge_tol = 1e9 | a mutant flagging only by tolerance fails 1 |
| F4 | the default codimension of the ImplicitSolution method is only exercised with one free and one clamped coordinate | right | a 3-D field with 0, 1, 2 and 3 free coordinates (precisions as Vector{Any} too), against the explicit codim | — |
| F5 | notes stale: report.md still said “no ridge residual”, no note on ridge_residual, Feature Backlog rows | true of the diff it saw; the notes were written after the diff was sent (implicit.md §5, report.md, the backlog) | — | — |
Observations, not defects: gap = -Inf at codim = 0 (a convention); at tied eigenvalues the
distance depends on the eigenbasis (documented: a small gap means the distance is not determined);
the returned fields mix element types for a Float32 state (distance is promoted).
Leaning / resolution. F1–F4 closed by tests that fail on their mutants. The review did not run
the whole suite nor examples/calibration.jl (the maintainer’s session ran both).