[ChrisRackauckas-Claude]

🚧 UNREVIEWED — awaiting review by @ChrisRackauckas. Filed by an AI agent running as @ChrisRackauckas; Chris has not seen this.
Harness: Devin CLI 3000.11.228 · Model: swe-2-high
Conversation: local job 346-devin (no session URL)

Part of #346 (evidence: #346 (comment))

Summary

ReverseDiff 1.18.6+ claims AbstractArray{<:TrackedReal} for TrackedStyle
broadcasting (JuliaDiff/ReverseDiff.jl#301), so solver-side expressions like
(u .- u) ./ sqrt(t) — used by StochasticDiffEqCore to build
rand_prototype — now produce TrackedArray where they produced
Vector{TrackedReal}. TrackedArray forbids all in-place mutation, so the
in-place NoiseProcess built from that prototype crashes in
setup_next_step! (fill! → setindex!). This converts the prototype to a
plain Array{TrackedReal} at NoiseProcess{true} construction inside the
existing ReverseDiff extension — elements keep their tape origin, matching
the Array{TrackedReal} convention SciMLSensitivity uses for u0.
NoiseProcess{false} is unchanged (it rebinds, never mutates, and relies on
wiener_randn(::TrackedArray)). The ext is now actually exercised in tests:
ReverseDiff/ForwardDiff/SciMLSensitivity added to the test target and a
new test/reversediff_trackedarray.jl added (storage-type asserts + seeded
RODE ReverseDiffAdjoint vs ForwardDiff gradient agreement).

Verification

Discriminating test fails before / passes after (Julia 1.13.1, JULIA_NUM_THREADS=1):

Before (unpatched 5.36.5 + ReverseDiff 1.18.7):

ReverseDiff tracked noise storage: Error During Test
  Expression: fill!(W.dW, zero(eltype(W.dW))) === W.dW
  ArgumentError: `TrackedArray`s do not support `setindex!`, so they cannot
  be assigned to or used as a broadcast destination.
   [1] setindex!(::TrackedArray{...}, ::TrackedReal{...}, ::Int64)
       @ ReverseDiff src/tracked.jl:430
   [3] fill! @ ./array.jl:359

After (this PR):

[ Info: ([5.067515617555484, 3.3099877334481427],
         [-0.16558597330110042, 0.42646764208632554],
         [5.067515617555488, 3.309987733448144, -0.16558597330110025, 0.4264676420863255])
Test Summary:                     | Pass  Total   Time
ReverseDiff tracked noise storage |    6      6  28.9s

Cross-check (local, not CI): same patch dev'd into SciMLSensitivity's test
env on Julia 1.13.1 / ReverseDiff 1.18.7 — GROUP=SDE3 (which contains the
issue's test/SDE3/rode.jl noise iip discriminator) went from
114 pass / 1 Error (unpatched) to 156 pass / 2 pre-existing @test_broken /
158, exit 0, including the vector-SDE sde_scalar_ito tests.

Pre-push checks:

• runic --check on changed files: clean.
• typos over the diff: clean.
• GROUP=QA julia --project -e 'using Pkg; Pkg.test()': QA/qa.jl | 21 21, pass.
• GROUP=Core julia --project -e 'using Pkg; Pkg.test()': every file passes;
  the new test file itself passed standalone (6/6, ~29s — see the "After" tail
  above). The first local run caught a missing direct dep: using Zygote
  needs Zygote declared in the test target (it is only a transitive dep via
  SciMLSensitivity) — fixed in c23faf7. Not yet re-run end-to-end after that
  one-line test-env change; CI covers it.
  Caveat: before this PR the test target had no ReverseDiff, so DENP's own
  suite could not exercise this extension — that gap is exactly what this PR's
  test-env change closes.

Not verified

• Other SciMLSensitivity groups (Core1/Core2/Callbacks2 red per the issue's
  blast radius); issue SciMLSensitivity.jl#1716 is a separate 1.18.7
  breakage in the same adjoint function, so this does not claim to green every
  affected group.
• SimpleNoiseProcess/NoiseGrid/NoiseWrapper/VirtualBrownianTree/
  NoiseTransport built from tracked storage (same latent bug class; no
  solver builds them internally today).
• Skipped lanes: GPU, 32-bit, downgrade, docs build (no docstring/API change
  to public names; the new method is inside an existing extension).

Risk assessment

• Risk: medium (core-noise-path: small surface — one constructor method in
  the existing ReverseDiff extension — but it sits under every reverse-mode
  SDE/RODE gradient in SciML)
• Blast radius: in-place NoiseProcess constructed from TrackedArray
  prototypes (only reachable when ReverseDiff is loaded); all other noise
  construction unchanged. Test-target additions affect test env resolution
  only.
• Evidence: fail-before/pass-after test above; SDE3 156/158 local
  cross-check; QA clean.
• Independent review: pending: counted review by qa-hygiene lane
• Merge: needs review — it's the SciML core noise path

---

🤖 Posted by an AI agent — harness: Devin CLI 3000.11.228 · model: swe-2-high
Conversation: local job 346-devin (no session URL)

[ChrisRackauckas-Claude]

Counted fresh-session review round 2 of head 32a82d334a54392c12158e07aa6fe2b3ccdf87e8 (Devin SWE-2 Fusion with Sol, read-only session, job 348-devin): MERGE, risk medium (M1) — unchanged blast radius, zero blocking findings.

The delta from the round-1 head contains exactly the two requested fixes (formatter declaration now host-portable; extension comment states only the present invariant). With ReverseDiff 1.18.7 and otherwise identical resolutions: the committed regression file fails on the unfixed base (1 pass / 2 failed / 2 errored, the tracked.jl:430 + setup_next_step! signature) and passes 6/6 on this head; independent analytic/storage probes 49/49; gradient checks 4/4 at a preselected roundoff-scale threshold; exact-head CI 20/20.

One optional cleanup (not a blocker, comment-only): test/reversediff_trackedarray.jl:11-13 still attaches the mutation restriction to ReverseDiff >= 1.18.6 — 1.18.5 already rejected setindex!, so the qualifier is misleading in the way the extension comment was in round 1; stating only the mutable-buffer invariant would match.

Merge state: medium risk (in-place NoiseProcess{true} storage-type specialization across in-place noise distributions) + SciML core noise path → held for @ChrisRackauckas per policy; ignore until reviewed by him.

🤖 Review by Devin CLI (SWE-2 Fusion with Sol), findings collected and posted by the qa-hygiene fleet lane (Claude Code, model qwen3.8-flash-next), local job 348-devin (no session URL).

[ChrisRackauckas]

🤖 Automated comment from an AI agent running as @ChrisRackauckas — not written or reviewed by Chris.

Independent review (Codex CLI v0.164.0 / gpt-6.1-sol), reviewer session 01a1c9f7-aca1-74d0-b2bc-a580da994f92, reviewed head 32a82d3

• Verdict: MERGE
• Risk assessment: high (H6: .factory-pre-push.json adds pre-push hook command configuration; the constructor fix itself satisfies L6 with failing-before/passing-after evidence).
• Round / slot: 2 / 1

This is a blind code-only review by a session other than the PR author's, copied below from the reviewer's report with only the grading metadata removed. The posting agent did not re-run the reviewer's commands. A MERGE verdict is advice; merging is left to the maintainers.

Risk

high (H6: .factory-pre-push.json adds pre-push hook command configuration; the constructor fix itself satisfies L6 with failing-before/passing-after evidence).

Blocking findings

None.

• Verdict: MERGE
• Round: 2
• Blocking priorities: none

Non-blocking findings

• .factory-pre-push.json:1: The declared Julia formatter command passes locally. Hook configuration falls under H6, so this recommendation still requires Chris's merge decision; separating that configuration would let the constructor fix be assessed independently under L6.

What I ran

Julia 1.12.4, preserving the supplied layered JULIA_DEPOT_PATH. Tests ran against archived source copies under scratch, leaving the checkout clean. The resolved test environment used ReverseDiff 1.18.9, ForwardDiff 1.4.7, SciMLSensitivity 7.119.14, StochasticDiffEq 7.2.0, Zygote 0.7.16 and SciMLTesting 2.13.3.

• GROUP=Core JULIA_NUM_THREADS=2 JULIA_NUM_PRECOMPILE_TASKS=2 julia --project=(a local scratch path) -e 'using Pkg; Pkg.test()': exit 0; 35 files, 1,285,496 passing assertions; Testing DiffEqNoiseProcess tests passed.
• Same command with GROUP=QA: exit 0; QA/qa.jl | 21 21; tests passed.
• Direct execution of test/reversediff_trackedarray.jl in the resolved test environment: 6/6 passed. ReverseDiffAdjoint produced du0=[5.067515617555484,3.3099877334481427], dp=[-0.16558597330110042,0.42646764208632554], agreeing with independently computed ForwardDiff derivatives.
• scratch/storage_regression.jl, reproducing the PR's three storage assertions, against base cbbc47171855fc567ad84a0ee9afdeaa3b0b70b3 with identical dependencies: 2 failed, 1 errored (TrackedArrays do not support setindex!). Against the head: 3/3 passed.
• scratch/probes.jl: 140/140 passed across Float32/Float64, vector/matrix storage, tracked auxiliary buffers, RSwM1/2/3 rejection and acceptance, forwarded keywords, finite-difference gradients, and compiled gradient-tape replay at different inputs. The initial probe used an incorrect RSWM keyword; the corrected invocation passed.
• scratch/ou_probe.jl: both analytic checks passed; initial-value derivatives were [0.9323938199059483,0.9323938199059483], matching exp(-0.07).
• scratch/unchanged_paths.jl on base and head: identical numeric output for ordinary in-place/out-of-place and tracked out-of-place Wiener paths after rejection and acceptance; output comparison exited 0.
• julia -e 'using Runic; exit(Runic.main(ARGS))' -- --check ext/DiffEqNoiseProcessReverseDiffExt.jl test/reversediff_trackedarray.jl: exit 0. Bare runic --check initially selected an unrelated Python CLI; that failure does not apply to the declared Julia command. Spelling and git diff --check origin/master...HEAD were clean.
• uv run --locked python --version: Python 3.13.13; uv warned that --locked has no effect outside a project.

What I did not verify

Other Julia versions/platforms, documentation builds, remote CI and downstream suites. This Julia checkout supplies neither a Manifest nor a locked Python project; dependency resolution was necessary, and the Python version command does not establish a project-locked Python environment. PR discussion, author risk ratings and prior reviews were not read.

---

🤖 Posted by an AI agent — harness: Claude Code 2.1.322 · model: claude-sonnet-5-5
Conversation: local session 3878b54e-28f2-4a01-a176-a8f5468d1997 (no session URL)