Status: resolved (2026-11-12). One defect found and fixed with a test; one suspicion not reproduced. Related: Feature Backlog (I5), LP-024 Findings of the review of 2026-11-09.
Sources: a review of the uncommitted I5 change (the parameter and state trees read by position) by Mistral, run with Julia in the helper worktree
Lenticulum.jl-mistral(2026-11-12). It stopped twice at its turn limit before writing a final list, so the findings below come from its scratch scripts there (scratch/i5_review.jl,scratch/i5_ad.jl,scratch/i5_plate_scaling.jl), which the maintainer’s session ran again. code:lib/Mycelium.jl/src/passing.jl,lib/MyceliumLearning.jl/src/gradient.jl; note:lib/Mycelium.jl/src/passing.md, “The trees by position”.
Problem. Mistral checked the I5 change against its requirement: every schedule returns a
NamedTuple, and the very tree it was given when there is no state; missing keys are appended in
the order of their first write; the tied rule (the last write wins); plates with block rows and a
stateful factor; residual probes do not write; step! on a NamedTuple; Zygote and ForwardDiff
through the free energy against the reference; parameter_step and free_energy_gradient
against finite differences. Rerun by the maintainer’s session: 195 checks pass in
i5_review.jl and 26 in i5_ad.jl. The one error in i5_review.jl is a defect of the scratch
test itself (MaybeCount has no parameterlength), which i5_ad.jl corrects.
| # | finding | right? | what was done | test |
|---|---|---|---|---|
| 1 | parameter_step with an optimiser state given as a Dict throws MethodError: no method matching _stored(::Dict…); the old code read it with haskey and getindex, so a Dict worked | right, reproduced (i5_ad.jl). The docstring asks for a NamedTuple, but the old code accepted a Dict and nothing announced a change | _stored(d::AbstractDict, name) = get(d, name, nothing) | MyceliumLearning test/gradient_couplings.jl, “parameter_step: an optimiser state given as a Dict”: one step with a Dict equals one with the NamedTuple; it fails without the method |
| 2 | _row_items scans all energy rows for each key in _envelope_gradient: O(keys × rows) | not reproduced: Mistral’s i5_plate_scaling.jl found the cost linear at the sizes it tried. The rows are few (one per plate and channel set), so it would matter only for graphs with many plates | nothing changed | — |
Leaning / resolution. Fixed with a test where reproduced; the second is recorded only.