ci: cross-OS solx-tester suite on free macOS/Windows runners (ci:cross-os label + weekly cron)

[nebasuke]

What

A label-gated + weekly-cron workflow running the full solx-tester suite on free macOS (arm64 + Intel) and Windows runners, plus the run-tester-suite composite action it is built from. The workflow consumes the build-toolchain composite (landed separately in #581) for its standard-config LLVM/solc prologue, so it adds no toolchain-build steps of its own.

• Triggers: PR label ci:cross-os (on-demand, for platform-sensitive changes like process/IPC work), workflow_dispatch, and a Monday 06:00 UTC cron on main as a drift net. Advisory only — must never become a required check.
• Scope: solx-tester only. Cross-OS legs test platform behavior (process management, IPC, paths, toolchain quirks); output quality stays on the Linux integration suite, and Hardhat/Foundry are deliberately not replicated.
• Cost: $0 — macos-15, macos-15-intel, and windows-2025 are free for public repos; LLVM/solc come from the finished-install caches saved on main, so nothing heavier than the Rust workspace compiles.

Depends on

#581 (merged) — extracted the standard-config toolchain prologue (submodules → SFW → msys → macOS-disk → LLVM → solc) into the shared build-toolchain composite. This workflow is its sixth consumer; the only change to the composite here is adding cross-os-tests.yaml to its consumer-list description. #581 also folded in #550's fix option 2, so the bundled static libstdc++.a drop now happens inside build-toolchain for every standard-config Windows build — this workflow inherits the working shared-libstdc++ link with no step of its own.

Why

One week of evidence that this layer catches what Linux cannot:

• the perf(core): reuse persistent worker subprocesses across translation units #524 worker-pool hang on macOS (present at an intermediate tip, fixed before merge — confirmed by this harness's watchdog samples, PR comment);
• the rustc 1.96.1 enum-niche miscompile killing solx-tester on Intel macs (solx-tester SIGSEGV on macOS x86-64: rustc 1.95/1.96 enum-niche miscompilation (rust-lang/rust#159035) — fix by bumping to 1.97.1 #547, since fixed by the 1.97.1 bump in chore: bump Rust toolchain to 1.97.1 #556 — the Intel leg here is what keeps verifying it, since test.yaml's Intel leg never runs solx-tester);
• the Windows libstdc++ link fragility (Windows: fresh llvm-sys builds cannot link against the bundled static libstdc++.a under MSYS2 GCC 16 — CI is one cache-fingerprint change from breaking #550), found the day this workflow first ran solx-tester on Windows — which had zero prior executions there; it passed 74,605 cases once the libstdc++ drop (now in build-toolchain) and the corpus-dir retention below were in place.

How it works

run-tester-suite (composite action): builds solx + solx-tester (symbols kept), smoke test, then the suite under a parametrized watchdog. A run that neither crashes nor finishes gets diagnosed before the kill — on macOS, sample stacks of the tester, the top-level solx parents, and the pooled workers, plus new crash reports from DiagnosticReports; on Windows, a tasklist snapshot. Signal deaths are reported by name (killed by signal 11 (SIGSEGV)) instead of the solx-dev wrapper's "subprocess failed without exit code". Stage outcomes are aggregated at the end so diagnostics always run.

Windows specifics:

• the bundled static libstdc++.a drop (Windows: fresh llvm-sys builds cannot link against the bundled static libstdc++.a under MSYS2 GCC 16 — CI is one cache-fingerprint change from breaking #550) is inherited from build-toolchain, applied to all standard-config Windows test builds — not a step this workflow adds;
• keep solx-solidity/test/ through the disk cleanup (the Ethereum corpus is read from it at run time) — this workflow's own Free disk space (Windows) step.

Verification

Run 29429860239: macos-15 and windows-2025 legs green, full suites — ~22 min and ~37 min job time respectively. That run predates the rebase onto build-toolchain, but the composite encodes byte-identical LLVM/solc params, so the tester results carry over; a fresh ci:cross-os run on the rebased branch confirms the composite path.

The macos-15-intel leg has no post-1.97.1 timing yet (pre-fix runs exceeded 75 min without finishing), so its watchdog (110 min) and job timeout (240 min) are a probe budget rather than the ~2x margin the other legs get. First runs establish the baseline; then tighten, or scope down to --path tests/solidity if the full suite can't finish on the free Intel minis.

Follow-ups (not in this PR)

• Tune the Intel leg's watchdog/timeout once real runs establish a full-suite baseline (see Verification).
• Windows: fresh llvm-sys builds cannot link against the bundled static libstdc++.a under MSYS2 GCC 16 — CI is one cache-fingerprint change from breaking #550's durable fix (option 1): stop bundling libstdc++.a into the LLVM tree in solx-dev and make release packaging link it statically on its own. That's a solx-dev + release.yaml change whose verification surface is a release build, so it stays out of CI-only work like this. When it lands, build-toolchain's drop step gets deleted — and its plain rm fails loudly if the bundling disappears first, so the workaround can't linger silently.
• The mac-tester-debug* / win-tester-spike branches and ci: debug solx-tester on macOS runners (main baseline for #524 mac validation) #546 get retired once this lands.

[github-actions]

Coverage Summary

Crate	Line Coverage	Function Coverage
solx	🟢 94.7%	🟡 68.0%
solx-benchmark-converter	🔴 0.0%	🔴 0.0%
solx-codegen-evm	🟢 88.9%	🟢 83.4%
solx-compiler-downloader	🔴 0.0%	🔴 0.0%
solx-core	🟢 96.6%	🟢 91.0%
solx-dev	🔴 2.4%	🔴 3.0%
solx-evm-assembly	🟡 70.6%	🟡 66.2%
solx-solc-test-adapter	🔴 1.7%	🔴 2.1%
solx-standard-json	🟢 95.4%	🟢 91.5%
solx-tester	🔴 38.2%	🔴 35.9%
solx-utils	🟢 83.8%	🟢 83.6%
solx-yul	🟡 78.9%	🟡 71.3%
Total	🟡 53.9%	🔴 47.0%

Codecov Report | HTML Report | Workflow Run