Fix flaky zero-allocation tests on Julia 1.12+ via AllocCheck.check_allocs

[ChrisRackauckas-Claude]

Problem

The Core test group is red on main across every platform (macOS, Ubuntu, Windows) and on both Julia 1 (1.12.6) and pre (1.13.0-rc1). The failures are all in test/alloc_tests.jl:

DisjointRange - Zero Allocations: Test Failed at test/alloc_tests.jl:21
  Evaluated: 739144 == 0
DisjointRange - Zero Allocations: Test Failed at test/alloc_tests.jl:45
  Evaluated: 704 == 0
getindex/setindex! - Zero Allocations: Test Failed at test/alloc_tests.jl:100
  Evaluated: 16 == 0

Root cause

These are @allocated measurement artifacts on Julia 1.12+, not real allocations in the package. @allocated measures the allocations of a single runtime call, which conflates first-call compilation and global-scope boxing with genuine heap allocation:

• Line 21 (dr[8]) — the second-range branch of DisjointRange getindex is never warmed up before this @allocated (only dr[3], the first-range branch, runs at line 20), so it measures ~700 KB of first-call compilation.
• Line 45 (iteration) — the accumulator s lives in the global scope of the @allocated begin ... end block, so it is boxed (704 bytes). That is an artifact of measuring outside a function, not of the iterate methods.
• Line 100 (A[50, 50]) — even after one warmup call, the first @allocated captured residual specialization allocations (16 bytes).

The package targeted Julia 1.10/1.11 CI when these tests were written; they passed there and only began failing as CI moved to Julia 1.12/1.13.

Fix

Rewrite the tests to use AllocCheck.check_allocs (a test dependency that was already declared but unused). check_allocs statically proves the absence of any allocating code path for the given argument types — it is immune to compilation and global-scope boxing noise and is a strictly stronger guarantee than @allocated == 0. This is not a loosening: the same intended invariant (the operation never allocates) is verified more rigorously.

Local verification

check_allocs reports 0 allocation sites for every operation, confirming the package code is genuinely allocation-free; only the test methodology was flaky. Running test/alloc_tests.jl after the change:

Julia	Result (before)	Result (after)
1.10.11	16/16 pass	11/11 pass
1.12.6	13 pass, 3 fail	11/11 pass
1.13.0-rc1	13 pass, 3 fail	11/11 pass

Full Core group via the real SciMLTesting.run_tests() harness on 1.12.6:

Test Summary:       | Pass  Total   Time
Core/alloc_tests.jl |   11     11  23.0s
Test Summary:      | Pass  Broken  Total  Time
Core/core_tests.jl |   28       2     30  9.6s   (2 pre-existing @test_broken, unchanged)

Runic format check on the changed file passes (exit 0).

---

Please ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code