[ChrisRackauckas-Claude]

Summary

CI's AD group was taking ~76 minutes while other groups finish in 5–17 minutes. Local LTS timings showed derivative_tests.jl dominated (~105 min wall) with OOP+IIP (~53 min) and BatchIntegralFunction cases (~53 min) nearly equal, so this splits them into AD and ADBatch (both still versions = ["lts"]) without changing any assertions.

Shared setup lives in test/shared/ad_derivative_setup.jl (not a group folder). nested_ad_tests.jl stays in AD.

Verification

Per-file LTS timings (AD env, julia +lts):

• nested_ad_tests.jl: WALL_SECONDS=199.26 (Pass 6)
• derivative_tests.jl (pre-split): WALL_SECONDS=6325.04 (Pass 2005, Broken 1176)

Section timeline (UTC) for the pre-split file supported an OOP+IIP vs Batch split.

GROUP=AD julia +lts --project -e 'using Pkg; Pkg.test()'
# AD/derivative_tests.jl | Pass 1285 Broken 744 Total 2029  62m33.0s
# AD/nested_ad_tests.jl  | Pass 6 Total 6  2m31.6s
# Testing Integrals tests passed
# WALL_SECONDS=4375.77

GROUP=ADBatch julia +lts --project -e 'using Pkg; Pkg.test()'
# ADBatch/derivative_batch_tests.jl | Pass 720 Broken 432 Total 1152  46m02.3s
# Testing Integrals tests passed
# WALL_SECONDS=2843.28

Union check: SciMLTesting discovery runs AD/{derivative_tests,nested_ad_tests}.jl and ADBatch/derivative_batch_tests.jl; @info/@testset labels match master derivative_tests.jl; pass+broken counts 1285+720=2005 and 744+432=1176 match the pre-split file.

Not verified

• Full CI matrix on GitHub Actions
• GROUP=All / Core / QA after this change
• Whether AD SIGSEGV on Julia 1.12+ still reproduces (reported separately in the agent REPORT, not this PR)

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: Low — test layout/CI matrix only; no production code or assertion changes.
• Blast radius: AD CI jobs only (one new matrix row ADBatch on LTS).
• Evidence: Local GROUP=AD and GROUP=ADBatch Pkg.test() both passed; assertion counts union to the original AD file.
• Independent review: pending
• Merge: needs review

Harness: Cursor Agent CLI 2026.09.26-dd393fe; model: auto; transcript: /home/crackauc/sandbox/goals/qa-hygiene/jobs/integrals-ad-split/log.txt

Made with Cursor

Risk assessment

• Risk: low
• Blast radius: Test suite and CI matrix only. No src/ or Project.toml changes, so package users are unaffected. The only CI change is one new LTS-only matrix row, ADBatch. One small maintenance cost: test/ADBatch/Project.toml is a byte-for-byte copy of test/AD/Project.toml, so compat bumps now have to be made in both files.
• Evidence: All 15 checks on head ef9a77ab pass (merge ref 576c36f against base 01dc4d50, which is still master HEAD). There are no failing checks, so nothing is a regression.
  • The last master CI run (2026-09-13) was also all green, with Tests - AD (Julia lts) taking 76 min.
  • On head, Tests - AD (Julia lts) passes in 23m29s with derivative_tests.jl at Pass 1285 / Broken 744 and nested_ad_tests.jl at Pass 6. Tests - ADBatch (Julia lts) passes in 18m38s with Pass 720 / Broken 432.
  • Those add up to 2005 pass and 1176 broken, the same as the original single file per the PR description.
  • I diffed the moved code mechanically. The removed setup block is identical to test/shared/ad_derivative_setup.jl, and the removed batch loops are identical to test/ADBatch/derivative_batch_tests.jl. No assertions, tolerances, or @test_broken lines changed.
  • The Core log shows test/shared/ is not picked up as a test file or group.
  • Core (julia pre) "passed" in 32s because the shared workflow skips the pre lane when the Julia build is not a release candidate (1.13.1 here). The PR has nothing to do with that skip.
• Independent review: Devin CLI (Mac) / fusion-claude-opus-5-5-high-sidekick-swe-2-medium rated it low, high confidence: it is a verbatim, assertion-preserving split of one test file into two LTS-only groups. CI is fully green, test counts add up exactly, and no public API or dependency internals change (Mooncake.rdata_type and Integrals.ReCallVJP/MooncakeVJP were already used in the test code and were only moved).
• Merge: auto-merge candidate. Caveats:
  • The PR is still marked draft and says "ignore until reviewed by @ChrisRackauckas".
  • The description does not mention the duplicated ADBatch/Project.toml.
  • Its "Not verified" list is out of date: CI has since run Core, QA and the full matrix green.
  • Its attribution footer gives the model only as "auto".

---

🤖 Risk assessment posted by an AI agent (fleet master) — harness: Devin CLI (Mac) / fusion-claude-opus-5-5-high-sidekick-swe-2-medium; dispatched by Devin CLI 3000.11.3 (Mac) fleet-master caretaker, model Fusion (claude-opus-5-5 medium + swe-2 medium)
Conversation: local transcript ~/.local/share/devin/cli/transcripts/speckle-whale.json (Chris's Mac)