[P0] Intel-Mac first launch is broken: uv.lock pins torch 2.8.0, which has no macOS x86_64 wheels #889

What breaks

Every Intel-Mac user's first launch fails at the dependency bootstrap. The shipped x64.dmg installs fine, but ensure_venv_ready → uv sync --frozen --no-dev dies with:

error: Distribution torch==2.8.0 … doesn't have a source distribution or wheel for the current platform
hint: You're on macOS (macosx_15_0_x86_64), but torch (v2.8.0) only has wheels for macosx_11_0_arm64

PyTorch dropped Intel-Mac wheels after 2.2.x; uv.lock pins 2.8.0. The app can never reach a working backend on that platform — a violation of the platform-parity hard rule (default features must work on every supported platform, else move the platform behind explicit opt-in/unsupported status).

Evidence

Found by the new cross-platform install-test CI (run 28577632544, branch ci/test-install): macOS ARM / Windows / Linux all complete a full first-run smoke (real 3.1 GB model download + verified synthesis WAV); the macos-15-intel leg fails at uv sync --frozen exactly as an end user would. Present in v0.3.8 (and earlier) — not a recent regression.

Options

1. Platform-marker torch pin (e.g. torch==2.2.2 on sys_platform == 'darwin' and platform_machine == 'x86_64') — but transformers 5.x and other deps may require torch ≥ 2.4+, so this needs a compat audit; likely a downgraded/limited engine set on Intel.
2. Declare Intel macOS unsupported: stop shipping the x64 dmg (or gate it with a clear message), update README's platform matrix + docs. Honest, cheap, and matches PyTorch's upstream direction.
3. Hybrid: keep the x64 build for the frontend but pin a reduced CPU-only backend profile for Intel.

The install-test workflow on ci/test-install reproduces this in ~10 minutes and will verify any fix (gh workflow run "Test Install (no release)" --ref ci/test-install).

🤖 Generated with Claude Code