Make tuple-node getindex type-stable; replace @test_broken @inferred with @inferred

[ChrisRackauckas]

🚧 UNREVIEWED — awaiting review by @ChrisRackauckas. Filed by an AI agent running as @ChrisRackauckas; Chris has not seen this.
Harness: Devin CLI 3000.11.92 (diagnosis and fix), Claude Code 2.1.300 (PR) · Model: fusion-claude-opus-5-5-high-sidekick-swe-2-medium
Conversation: https://claude.ai/code/session_01VZjwf7Ux9YtTzwHGHMqnsq (fleet master)

This fixes a failure on the default branch's CI. The diagnosis and the evidence below come from the investigating agent's report; "confirmed by running" and "read in logs" are kept apart there.

Failure 1: test/tuple_nodes.jl:51-52 @test_broken @inferred errors with "Expression evaluated to non-Boolean" on Julia 1.13.2

Error (read in logs)

Core/tuple_nodes.jl: Error During Test at .../test/tuple_nodes.jl:51
  Expression evaluated to non-Boolean
  Expression: @inferred getindex(community, 4)
Core/tuple_nodes.jl: Error During Test at .../test/tuple_nodes.jl:52
  Expression evaluated to non-Boolean
  Expression: @inferred getindex(scenario, 8)
Some tests did not pass: 20 passed, 0 failed, 2 errored, 0 broken.

Two jobs fail with the same error, both on julia version 1.13.2: Core (julia 1, ubuntu-latest) and Core (julia 1, ubuntu-latest, x86).

• PR run (jet-0.12, 622a6ee): https://github.com/SciML/MultiScaleArrays.jl/actions/runs/38812758360 (job https://github.com/SciML/MultiScaleArrays.jl/actions/runs/38812758360/job/116691150425)
• The master push run on 4611c53, which finished after this task was filed, fails the same way: https://github.com/SciML/MultiScaleArrays.jl/actions/runs/38819792025 (jobs https://github.com/SciML/MultiScaleArrays.jl/actions/runs/38819792025/job/116714172009 and https://github.com/SciML/MultiScaleArrays.jl/actions/runs/38819792025/job/116714171940). In that run, Core lts and QA pass. The pre lane passes only because it skips itself when there is no release candidate ("Julia 1.13.2 is not a release candidate ... skipping").
• The last green master Tests run is https://github.com/SciML/MultiScaleArrays.jl/actions/runs/34655459665 (2026-09-11). Its Core julia 1 job ran on julia version 1.13.0 and reported Core/tuple_nodes.jl | 20 pass, 2 broken.

Cause: a Julia release (1.13.1 inference improvement), so the test expectation is now wrong. This is not a dependency change.

getindex(m::AbstractMultiScaleArray, i::Int) calls nodeselect(m.nodes, idx, i) = ns[idx][i] with a runtime idx. m.nodes here is a heterogeneous tuple (two Plants with different settings/organ types). Julia 1.13.1 and later union-split this and infer Float64. Earlier versions infer Any. So @inferred now returns 4.1 instead of throwing, and @test_broken rejects the non-Boolean value. The code path uses only Base (tuples, Vector{Int}, searchsorted), so package versions cannot change the result.

Confirmed by running (scratch/infer.jl against unmodified master 4611c53 copied to scratch/master_src; it prints Base.return_types(getindex, (typeof(x), Int)) and the @inferred result):

Julia	community / scenario return type	@inferred
1.10.12 (lts)	Any	throws
1.12.7	Any	throws
1.13.0 (official tarball)	Any	throws
1.13.1 (official tarball)	Float64	passes
1.13.2 (release)	Float64	passes

The behaviour changes between 1.13.0 and 1.13.1. I did not bisect to the individual Julia commit. The GitHub julia 1 lanes moved from 1.13.0 to 1.13.2 between the 2026-09-11 green run and today.

Existing PRs

None. The only open PRs in SciML/MultiScaleArrays.jl are #74 (ChrisRackauckas, "simple arrayinterface overloads", 2023) and #54 (ChrisRackauckas, "Fix diffeq tag", 2021). Neither touches this. A search for inferred OR tuple_nodes OR test_broken finds only merged QA PRs.

Fix

The test is not version-gated. Instead, the code is made type-stable on every supported Julia version, so the test can assert inference unconditionally.

• src/indexing.jl: add Tuple methods for nodeselect. They recurse over the tuple, so a runtime index into heterogeneous tuple nodes infers as the union of each element's result (concrete when the leaves share an eltype):

  nodeselect(ns::Tuple, i, I...) = i == 1 ? first(ns)[I...] : nodeselect(Base.tail(ns), i - 1, I...)
  nodeselect(ns::Tuple{}, i, I...) = throw(BoundsError(ns, i))

  Vector nodes keep using the generic ns[i][I...] method.
• test/tuple_nodes.jl:51-52: @test_broken @inferred getindex(...) becomes @inferred getindex(...), matching lines 49-50. The old expectation ("this is not inferable") is false on Julia ≥ 1.13.1, and with the fix it is false on every supported version. The test is now strictly stronger: it fails if inference regresses on any version.

With the fix, return_types gives [Float64] for both community and scenario on 1.10.12, 1.12.7 and 1.13.2 (confirmed by running).

Test before and after (confirmed by running; logs in scratch/):

• Before, master test on 1.13.2 (master_release_nonboolean.log): Error During Test ... Expression evaluated to non-Boolean ... Value: 4.1, exit 1. This reproduces the CI failure.
• Before, new test with the src fix stashed, on 1.12.7 (failbefore_1.12.log): ERROR: LoadError: return type Float64 does not match inferred return type Any at tuple_nodes.jl:51, exit 1.
• After, include("test/tuple_nodes.jl") against the dev'd branch (tuple_release.log, tuple_1.12.log, tuple_lts.log): exit 0 on 1.13.2, 1.12.7 and 1.10.12.
• After, GROUP=Core Pkg.test() on 1.13.2 (core_release.log): all pass, including Core/tuple_nodes.jl | 20 20 and indexing_and_creation_tests.jl | 134 134.
• After, GROUP=QA Pkg.test() on 1.13.2 (qa_release.log): QA/qa.jl | 20 20, passed.
• Runic --inplace on both files: no changes. typos on both files: clean.

Not run locally: the 32-bit Core lane (x86), and the full Core group on lts and 1.12. On lts and 1.12 only tuple_nodes.jl was run.

---

🤖 Posted by an AI agent — harness: Devin CLI 3000.11.92 (fix), Claude Code 2.1.300 (PR) · model: fusion-claude-opus-5-5-high-sidekick-swe-2-medium / claude-opus-5-5
Conversation: https://claude.ai/code/session_01VZjwf7Ux9YtTzwHGHMqnsq (fleet master)

🤖 Generated with Claude Code

Independent review (added by the fleet master)

Risk assessment

• Risk: low
• Blast radius: Tuple-backed node selection, including linear and hierarchical indexing; vector-backed nodes retain the generic method. No public name, dependency, or compatibility bound changes.
• Evidence: Independently tested the actual indexing/shape-construction source and the tuple fixture on Julia 1.10.12 and 1.13.2: community/scenario inference, all 12 values, and lower/upper node bounds pass. With original indexing source and the new assertions on LTS, observed return type Float64 does not match inferred return type Any. Read the default-branch log independently: both release architectures report non-Boolean results at the two broken-inference assertions. Also ran the full tuple_nodes.jl on Julia 1.10.12, including its ODE solve: exit 0. PR CI remains pending.
• Independent review: codex gpt-6.1-sol rated it low, high confidence: tuple recursion exposes each concrete node's indexing result to LTS inference while retaining valid integer-index results.
• Merge: auto-merge candidate — after Core (including LTS/x86), QA and downgrade CI pass.

---

🤖 Section added by an AI agent — harness: Claude Code 2.1.300 · model: claude-opus-5-5 (review by codex-cli 0.162.1, gpt-6.1-sol)
Conversation: https://claude.ai/code/session_01VZjwf7Ux9YtTzwHGHMqnsq