Status: resolved (2026-11-16). The review of P26 (ForwardDiff and Zygote through a cluster’s joint precision with flat directions, at constant rank) ran eight probes and found no defect. One probe was weaker than reported (below); what it was meant to show is covered by the suites. Related: Feature Backlog (P25, P26), cluster “Flat directions under a derivative”.
Sources: a review of the uncommitted P26 change by Mistral (
vibe --agent review,--max-price 3, $0.66 at list price), in the helper worktreeLenticulum.jl-mistralat73277da7with the diff applied; the brief, the answer file and the probes in../Lenticulum.jl-Review-2026-11-16/p26/. code:lib/Mycelium.jl/src/cluster.jl,ext/MyceliumForwardDiffFlatExt.jl; tests:lib/Lenticulum.jl/test/regime_plate.jl.
Problem. The requirement given to the reviewer: at constant rank, the derivative through the
diffuse limit equals that of the value path, also with repeated eigenvalues on the range; a tangent
that changes the rank to first order is refused with an ArgumentError; Float64 values unchanged;
nested duals and other number types still refused on the flat path.
| # | probe | right? | what was done | test |
|---|---|---|---|---|
| P1 | Float64 values of _psd_factor, _psd_solve, _psd_inv, _psd_logdet unchanged (rank 0, rank deficient, definite) | the probe has no assertion: it prints the values and reports “matched”. Unchanged Float64 values are shown by the suites: Lenticulum test/cluster_flat.jl (closed forms to ) and the cluster testsets pass | — | test/cluster_flat.jl |
| P2 | ForwardDiff against central differences, constant rank (rank 1 and 2, rotating, repeated eigenvalues) | passed | — | P26 testset |
| P3 | rank changes 0→1, 1→2, 2→3 refused; constant rank accepted | passed | — | P26 testset (1→2) |
| P4 | rank-0 J (every direction flat) under ForwardDiff; 0→1 refused | passed | — | probe4.jl only |
| P5 | one flat direction, _check_range refusing a linear term along it | passed | — | P26 testset, test/cluster_flat.jl |
| P6 | Hessians (nested duals), BigFloat, Rational refused with ArgumentError | passed | — | probe6.jl only |
| P7 | the new fields of _Flat (K for λ) under every _psd_* | passed | — | the suites |
| P8 | the testset’s regime factors, P24’s former limit, Zygote against ForwardDiff | passed, as the testset does | — | P26 testset |
No defect was found. Not probed: cluster_marginal of a flat class under ForwardDiff (the brief
named it; the probe checked the field accesses only), and a cluster with several precision classes
whose flat directions differ in direction but not in number.
Reviewer’s tooling. The brief gave the exact julia command line without cd and without
redirection; every probe ran, and it wrote the answer file as it went.
Leaning / resolution. No change.