open-problem

Status: resolved (2026-11-12). No defect was found. One claim of a compile-time regression was not reproduced, a gap in the tests (concurrency, several stores) was filled, and one note was not yet written when the review ran. Related: Feature Backlog (I18), LP-026 Findings of the review of I18.

Sources: a review of the uncommitted second part of I18 (the planned factor → variable messages) by Mistral (vibe --agent review), run with Julia in the helper worktree Lenticulum.jl-mistral (2026-11-12). Its answer is in ../Lenticulum.jl-Review-2026-11-12/REVIEW-i18b-from-mistral-2.md (the first run stopped at its turn limit), and its scratch tests are scratch/i18b_equiv.jl, scratch/i18b_equiv2.jl, scratch/i18b_race.jl and scratch/i18b_compile.jl in that worktree. code: lib/Mycelium.jl/src/passing.jl (_EmitPlan, _emit_plan, _emit), lib/Mycelium.jl/src/graph.jl (emit_plans); note: lib/Mycelium.jl/src/polarity_resolution.md, difficulty 3, “Planned per graph”.

Problem. The change was required to give the same results, errors and warnings as the code it replaced, for every graph, store and rule. Mistral re-implemented the old branch from the diff and compared it with the new one, task by task and sweep by sweep: random graphs with belief-propagation, expectation-propagation and mixture-reduction rules, detached edges, pins, damping, plate graphs with block rows, factors of 63, 64, 65 and 70 edges, and reset!. It also ran 16 threads on one graph with the same keys. It found the same beliefs, residuals and errors in each case, and no race.

#findingright?what was donetest
F1Each new set of informative inputs compiles a new _emit → _rule_message → factor_message chain, while the old code “compiled factor_message once per factor type”; emit_plans is not boundednot reproduced as a regression. The old code passed the inputs as a NamedTuple{chans} too, so it also compiled once per set. A scratch comparison on a 20-channel factor (30 new sets on each path, not kept in the repository) took 2.09 s on the old path and 2.10 s on the new one. emit_plans grows like the polarities cache, which has one entry per (target, available channels)nothing changed; recorded in polarity_resolution.md—
F2no test of concurrent tasks on one graph, nor of several stores sharing one graph’s plansright (a gap; Mistral’s scratch test found the property holds)test addedMycelium test/emit_plan.jl, “plans under concurrent tasks and several stores”: 8 tasks with their own stores, 20 rounds, compared with the old path, then every stored plan with a fresh one
F3no test that the warning of the TrivialBelief() case firesright, but the warning is unchanged code with maxlog = 1, which a test sees only on the first call of a sessionnot added; the outcome (TrivialBelief()) is tested—
F4polarity_resolution.md still said the dispatch “remains”right when the review ran; the note was written after the diff was takenupdated with the change—

Leaning / resolution. Only F2 called for code, and it got a test. F1 is recorded as not reproduced.

Not checked: the compile time per new input set on a schedule that sees many different sets (a residual schedule on a loopy graph); it is the same on both paths by the argument above but was measured only on one factor.