CI: serialize precompile workers for python-using groups

[ChrisRackauckas-Claude]

Summary

Replaces #1180 (closed) with a substantive non-retry fix.

The OptimizationSciPy (and occasionally OptimizationPyCMA) jobs have been observed to fail at the precompile stage with

InitError: UndefRefError: access to undefined reference
 [1] _pyjl_new(t::Ptr{...}, ::Ptr{...}, ::Ptr{...})
     @ PythonCall.JlWrap.Cjl ~/.julia/packages/PythonCall/.../JlWrap/C.jl:24
 ...
[11] __init__()
     @ PythonCall.JlWrap ~/.julia/packages/PythonCall/.../JlWrap/JlWrap.jl:70
during initialization of module JlWrap
in expression starting at .../OptimizationSciPy/src/OptimizationSciPy.jl:2

with the same PythonCall v0.9.31 on which other CI runs succeed — OptimizationSciPy, lts passes on master run 24144818227 but the same job fails on PR #1169 run 23171032225.

Root cause

The intermittency + identical package versions + failure happening inside PythonCall's own wrapper-type registration in JlWrap.__init__ all point at a parallel precompile race. JlWrap.__init__ (.../src/JlWrap/JlWrap.jl:54-75) does

function __init__()
    init_base()
    init_raw()
    init_any()
    ...
    init_module()
    ...
    jl = pyjuliacallmodule
    jl.Core = Base.Core      # ← line 70: pysetattr → _pyjl_new
    jl.Base = Base
    jl.Main = Main
    ...
end

When multiple precompile workers spin up wrapper-type registration concurrently, one observes a not-yet-populated Python type slot inside _pyjl_new and throws UndefRefError. This explains the every-N-th-run flake without any other moving parts.

Fix

Pkg.precompile() reads JULIA_NUM_PRECOMPILE_TASKS (Julia base/precompilation.jl:437) to size its parallel-task semaphore. Setting it to 1 serializes precompile workers and removes the race.

- if: ${{ matrix.group == 'OptimizationSciPy' || matrix.group == 'OptimizationPyCMA' }}
  run: echo "JULIA_NUM_PRECOMPILE_TASKS=1" >> $GITHUB_ENV

Targeted to OptimizationSciPy and OptimizationPyCMA only. Other groups keep their default parallel precompile, so there's no global slowdown. The cost on these small python-wrapping subpackages is a few seconds.

Why not retry

A retry-on-failure approach (#1180) would have masked the race instead of removing it. Subsequent failures of unrelated bugs in those groups would also have been silently retried. Per maintainer feedback, JULIA_NUM_PRECOMPILE_TASKS=1 is the more honest, targeted fix.

Test plan

• JULIA_NUM_PRECOMPILE_TASKS=1 confirmed to be read by Julia 1.11 (base/precompilation.jl:437)
• YAML validates with yaml.safe_load
• Conditional if: block scoped to the two python-using groups only
• Live CI on this PR

🤖 Generated with Claude Code