[julia-script]

Milestone (iii) needs macOS arm64 in CI. This adds it without a macOS seed asset or a seed schema change.

• silkc build accepts --target <triple>. It feeds the build profile's existing target input, so every target fact comes from the triple and none from the host. A non-host target emits only LLVM IR (--emit llvm-ir). An executable build for a non-host target is an invalid build profile, since the driver links only for its host, and there is no host fallback. The TypeScript CLI already has --target.
• New job macos-ir (Linux, needs native-n1): N1 cross-emits the compiler for aarch64-apple-darwin from the checkout.
• New job macos (macos-15, needs macos-ir) works in three steps:
  • It links that IR with a pinned LLVM 22.1.8 macOS release (sha256-verified, cached) and SDKROOT from xcrun. It passes the same clang arguments silkc's own link uses.
  • The result M1 emits the compiler from its own checkout, and the bytes must equal the cross-emitted IR. That is the fixpoint, since M2 would be clang(same IR). On a mismatch the first differing lines are printed.
  • M1 runs the native corpus (corpus-full).
• Both jobs run on selfhost pushes and on manual workflow_dispatch only, not on PRs. A change to target or ABI code can dispatch the workflow on its branch. Both jobs are non-required: neither feeds Focused Linux Silk or publish-seed.
• Focused Linux Silk accepts skipped bootstrap cases on any non-push event (!= push), so on a dispatch as on a PR. Without that, a dispatch would post a red required check on the PR's head. The step keeps Fail the required aggregate when any shard fails #1460's one-test-per-line shard checks.
• compiler/scripts/test-native-build.ts covers cross-target emission (it emits IR without invoking clang, and that IR differs from host IR). It also covers rejecting an executable for a non-host target, an unknown triple, and a repeated --target. It passes locally with this branch's N1.
• compiler/README.md documents the flag. PAPERCUTS gains two entries.

None of the touched files is a guarded input, so this lands on selfhost directly.

Validation

Local bootstrap compile on head 94e4fb1 (selfhost 11525ab, including #1459, #1458 and #1461, merged in), as required for compiler-changing PRs: node packages/cli/dist/bin.js check --manifest-path compiler/silk.toml, using a CLI built from current main, reports 1 succeeded, 0 failed (14m21s, 10.8 GB peak RSS). It checks the compiler root src/main.silk, which carries the --target change. An earlier run on adf5fd3 also passed.

Dispatch run 38184329312 on head 23e2b36: no fixed path, each leg builds from its own checkout, which #1450 made possible.

step	time
macos-ir job, Linux	about 2m
link M1	51s
M1 re-emits the compiler, byte-identical to the Linux cross-emission	1m49s
native corpus with M1	failed on the four P6 programs, as Linux did then

#1459 has since fixed those four on selfhost.

Earlier PR run 38176701398, before P6, with the then-temporary fixed path: the macOS leg was fully green, corpus pass=414 fail=0 unsupported=5. Timings: macos job 18m57s (LLVM cache restore 59s, link 64s, re-emit 3m57s, corpus 12m31s), macos-ir 2m07s.

Cost per run is about 21 runner-minutes (19 on macOS). It now applies only to selfhost pushes and dispatches, so PR latency is unchanged.

For #567 ("Current status")

macOS arm64: built and self-built in CI on selfhost pushes (non-required; workflow_dispatch for PRs touching target or ABI code). The Linux N1 cross-emits the compiler for aarch64-apple-darwin (silkc build --target), macOS links it as M1, and M1 re-emits byte-identical compiler IR. The native corpus runs on M1 (414 pass, 0 fail, 5 unsupported).

🤖 Generated with Claude Code

https://claude.ai/code/session_01SJak9tcQcwEZgNXjANPTiA

[julia-script]

"Selfhost native corpus" fails on this head (run 38184331627) with four pinned programs: nominal-union-operator-provider, operator-interface-contract, vector-sort-search-stability and vector-move-only-sort-and-cleanup.

This failure isn't caused by this PR. This PR touches only the workflow, the --target CLI option and docs. The same four failures happen on the selfhost push of 3ca78ea (run 38183198621), which predates this branch's latest base merge. No fix exists yet, so nothing is ported here.

The aggregate check passed on that push despite the red corpus. That gate hole is fixed separately in #1460.

---

Generated by Claude Code