test: run the precompile workload testset (and: the Windows julia-1 failure is upstream)

[ChrisRackauckas]

Summary

Two separate things were investigated here; only one of them is an actual repo bug.

1. The failing Core (julia 1, windows-latest) lane is NOT caused by anything in this repo

The lane fails while precompiling Reactant (a test/Project.toml dependency loaded by
test/shared_testsetup.jl), not while running NeuralOperators code:

lld: error: undefined symbol: jl_cstr_to_string
>>> referenced by jl_50E6.tmp(text#0.o):(julia_call_with_reactant_54807u54810)
...
LoadError: failed process: Process(setenv(`...\julia\1.12.7\x64\libexec\julia\lld.exe' -flavor gnu -m i386pep ...
  -o 'C:\Users\runneradmin\.julia\compiled\v1.12\Reactant\jl_A381.tmp' ...`), ProcessExited(1))
in expression starting at ...\test\shared_testsetup.jl:1
in expression starting at ...\test\models\fno_tests.jl:3

Evidence that this is an environment/upstream regression rather than a regression from
57cbb6c/eb240e7/1927dfb:

• Last green Windows/julia 1 job (run 31109559340, commit 57cbb6c, 2026-08-06) used
  Julia 1.12.6 and Reactant v0.2.278. First red one (run 32005859786, commit eb240e7,
  2026-08-17) used Julia 1.12.7 and Reactant v0.2.279. Nothing in eb240e7 touches
  Reactant, precompilation, or the test harness; it only adds a SciMLTesting = "2.10"
  compat entry.
• NeuralOperators itself precompiles fine on that lane at master HEAD
  (24043.3 ms ✓ NeuralOperators), including the new @compile_workload. The failure happens
  later, inside Reactant's own precompilation.
• The isolating experiment already exists. The codex/ci-probe-clean-main-neuraloperators
  probe (run 31919259681, 2026-08-16, sha 4e52d06) is 57cbb6c plus a comment-only line in
  Project.toml — functionally the same tree as the last green run — and its
  Core (julia 1, windows-latest) job failed with the identical
  lld: error: undefined symbol: jl_cstr_to_string, on Julia 1.12.7 with Reactant v0.2.279.
  Same code, newer toolchain, red.
• The failure is deterministic, not a flake or a timeout: lld: error: undefined symbol appears
  three times in the same job (Julia's precompile retries) and in every run since 2026-08-16.
• The julia lts Windows lane (Julia 1.10) is green at HEAD, matching a Julia-1.12.7-specific
  codegen/linker problem.

So the Windows lane should not, by itself, block the release: the released code fails the same
way in the current environment. The fix belongs upstream (Reactant/Julia on Windows).

2. Real bug from #160: the new precompile-workload test never ran

test/runtests.jl gained the "Precompile workload" @safetestset after
return @time @safetestset "Transform Interface" ... end, so it was unreachable — the
regression test for the new @compile_workload was dead code in every group. This PR moves the
return to the last testset.

Verification

• julia --project=. -e 'using Pkg; Pkg.instantiate(); using NeuralOperators' — the
  @compile_workload runs clean (38715.0 ms ✓ NeuralOperators on Julia 1.11.8, macOS aarch64).
• julia --project=. -e 'include("test/precompile_workload.jl")' — 3 passed, so the newly
  reachable testset is green.
• Runic --check test/runtests.jl — clean.

No version bump here; the release pass handles that.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Rmh6B9eoW3N8oCbuuVPVxJ