Download a working Crisp ASR build on Intel Macs

[niksedk]

Fixes #13559.

Upstream's crispasr-macos.tar.gz is arm64-only — its build job runs on macos-latest, which has been Apple Silicon since macos-13 was retired, and it sets no CMAKE_OSX_ARCHITECTURES. I checked every release from v0.1.0 through the pinned v0.8.29: all of them are arm64. An Intel Mac downloads 15 MB and gets Bad CPU type in executable, or crashes inside ggml-metal when the backend is forced. Rosetta can't bridge it — it translates x86_64 → arm64, never the reverse.

So Subtitle Edit now builds the Intel slice itself, from the same pinned tag:

• Build: build-crispasr-macos-x64-release.yml in SubtitleEdit/support-files, cross-compiling x86_64 on the arm64 runner.
• Release: crispasr-0829-macos-x64 → crispasr-macos-x86_64.tar.gz (17 MB).
• This PR: CrispAsrDownloadService picks that URL on non-arm64 macOS, and the Intel archive gets its own hash keys.

How the Intel build differs from upstream's macOS recipe

	why
-DGGML_METAL=OFF	ggml's Metal kernels target Apple-family GPUs; on the AMD GPUs in Intel Macs they crash — exactly the failure in the report. CPU + Accelerate BLAS instead.
-DGGML_NATIVE=OFF	otherwise the runner's -mcpu=apple-mN flags land in an x86 compile. ggml then picks its portable baseline: SSE4.2 + AVX + AVX2 + FMA + F16C + BMI2 — Haswell and up, i.e. every Intel Mac that runs macOS 12.
-DCMAKE_OSX_DEPLOYMENT_TARGET=12.0	left unset, clang targets the runner's own macOS and dyld refuses to launch the binary on anything older.
c2pa prebuilt fetched explicitly	-DCRISPASR_C2PA_FETCH=ON derives its triple from CMAKE_SYSTEM_PROCESSOR, which on a cross-arch Apple build is still the host — it would drop an aarch64 libc2pa_c.dylib into an x86_64 build.

The archive's inner folder is crispasr-macos, same as upstream's, so the unpack path needs no arch-specific case. The workflow's Verify step gates on all of it: x86_64-only slices, no arm64 leftovers, the expected minos, and no links outside /System and /usr/lib.

Verification

The published artifact, run under Rosetta on an M-series Mac:

  arch          : x86_64
  ggml backends : cpu,blas
system_info: ... AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | ACCELERATE = 1
[00:00:00.000 --> 00:00:03.800]   Subtitly added now, will crisp AS alpha Intel Max.

(the wording is tiny.en being tiny.en — the point is that it runs and decodes.)

Note for a follow-up

Upstream's arm64 binary is stamped minos 26.0 from v0.8.20 onwards (v0.7.0 was 15.0), because their runner moved to macOS 26. dyld refuses to launch a binary whose minimum is newer than the OS, so the pinned v0.8.29 likely won't start on Apple Silicon Macs running macOS 15 or older either. Not addressed here — worth an upstream issue asking them to pin the deployment target, or the same treatment as this Intel build.

🤖 Generated with Claude Code