ADR-017 — Intel Mac support: build it while it is free, and put the end date in writing¶

Status: Accepted (2026-08-15) Deciders: Mohsen Seyedkazemi Ardebili Context links: issue #264 (the decision this answers), #216 (test on Intel), [[adr-016-dependency-budget]], docs/platform-support.md

Context¶

Issue #264 asks a binary question: "do we pay for an Intel macOS runner, or declare Apple silicon only?" It is well evidenced on the technical side — the .dmg was decompressed on Linux and every Mach-O header read ({'arm64': 122}, zero fat headers), so the artifact really is arm64-only, and universal2 really is unreachable because several runtime wheels ship single-architecture binaries.

Both of its premises turned out to be wrong when checked against the world in August 2026.

Premise 1: "it costs money rather than effort" — false¶

GitHub Actions is free for standard GitHub-hosted runners in public repositories, and that is not a theoretical reading here: test.yml runs macos-latest on every push, and build-macos.yml runs it on every tag. This repository has been consuming GitHub-hosted macOS runner time continuously and #264 does not report a bill, because there is none.

An Intel leg costs runner minutes, not money. The decision was never a spend.

Premise 2: "an Intel runner exists to be paid for" — no longer true either¶

The Intel image the issue implicitly assumes, macos-13, began deprecation on 22 September 2025 and was fully retired on 4 December 2025. Writing runs-on: macos-13 today does not buy an Intel build; it fails.

What exists instead is a migration label, macos-15-intel, and it has a published end of life:

Decision¶

Neither option in the issue. Build Intel now, on macos-15-intel, as an advisory matrix leg — and document that Intel support ends in 2027 because GitHub ends it.

1. Add an Intel leg to build-macos.yml using macos-15-intel, alongside the existing Apple-silicon build.
2. Mark it advisory (continue-on-error), exactly as build-windows.yml already treats its brand-new windows-11-arm leg. This repository's own rule is that a brand-new cross-architecture build must not be able to fail a release the primary build completed fine, and there is no reason for Intel to be the exception.
3. Document it as ⏳, not ✅, until it has produced a working installer. The arm64 snap gap is the standing lesson: a claim nothing has exercised reads exactly like a claim something has.
4. Write the 2027 end date into docs/platform-support.md. An Intel Mac owner deciding whether to adopt YazSes deserves to know the desktop bundle has a dated horizon, and that pipx install yazses — which works on Intel today, since ctranslate2 publishes macosx_11_0_x86_64 wheels — is the path that outlives it.

Consequences¶

Good. The gap closes at zero cost, using a mechanism already proven in this repo. #216 ("test YazSes on Intel") becomes answerable by a contributor with something to download. The honest deadline is published rather than discovered by a user in 2027.

Accepted cost. One more advisory build per tag, and a second .dmg to name, checksum and list in the release notes. The naming must distinguish the two artifacts — the current YazSes-<version>.dmg is architecture-silent, which is itself part of how the arm64-only bundle went unnoticed.

The unavoidable one. This buys roughly two years. In Fall 2027 the Intel .dmg stops being buildable on Actions and the row becomes ❌ unless someone self-hosts an Intel Mac. That is the "declare Apple silicon only" outcome from #264 — it is simply scheduled by GitHub rather than chosen by us, and the right response is to take the two years rather than concede them early.

What this does not decide. Signing and notarisation are untouched; both bundles stay unsigned until the code-signing decision lands separately.

Alternatives considered¶

• Declare Apple silicon only, now. Rejected: it discards two years of free coverage and strands every Intel Mac still in service, on the strength of a cost that does not exist.
• Pay for a third-party Intel runner. Rejected: there is nothing to buy that macos-15-intel does not already give for free until 2027.
• universal2. Rejected on the evidence already in #264 — PyInstaller can only emit a fat binary when every bundled wheel is fat, and several are not.
• Cross-compile x86_64 from an arm64 runner. Rejected: PyInstaller does not cross-build, and target_arch cannot conjure an x86_64 slice out of arm64-only wheels.