chore(ci): move clippy + test from PR-time CI to release pipeline (#21)

The `clippy + test (macOS ARM64)` job in ci.yml was burning ~15 minutes
of the single shared self-hosted Apple Silicon runner on every PR push.
That cost was hard to justify given that:

- Local developers now have `make verify` / `make verify-clean` which
  reproduce the exact same `cargo clippy --features metal,accelerate
  -- -D warnings` + `cargo test --release --features metal,accelerate`
  invocation, so the runner round-trip wasn't adding signal that wasn't
  reachable locally.
- The self-hosted runner is shared with the release pipeline, and PRs
  routinely starved release builds for runner time.

This commit:

1. Removes the `clippy-and-test` job (and the now-obsolete header
   comment block) from `.github/workflows/ci.yml`. PR-time CI now runs
   only the lightweight cargo-deny + cargo-fmt gates on ubuntu.
2. Adds the same fmt / clippy / test steps to `release.yml`'s
   `build-macos` job, in front of the signed-build step. A clippy or
   test regression in release builds will now fail the release before
   any signing or packaging work happens.

`release.yml::build-macos` already uses the same self-hosted runner with
a persistent cargo target cache, so the lint/test steps benefit from
incremental compilation and don't redo the full build for the artifact
step that follows.

Author: inureyes

.github/workflows/ci.yml

@@ -1,8 +1,14 @@
-# PR / push CI for lightweight quality gates. Builds and tests are covered by
-# release.yml (signed artifacts) and pipeline-parallel-ci.yml (distributed
-# runs); this workflow runs cargo-deny (license / advisory), cargo-fmt, and
-# cargo-clippy + cargo-test checks on every touched-Rust change so formatting
-# drift, license issues, and lint/test regressions are caught at PR time.
+# PR / push CI for lightweight quality gates. Runs cargo-deny (license /
+# advisory) and cargo-fmt on every touched-Rust change so formatting drift and
+# license issues are caught at PR time.
+#
+# Clippy and cargo-test used to run here as `clippy + test (macOS ARM64)` on
+# the self-hosted Apple Silicon runner, but each run consumed ~15 minutes of
+# that single shared runner per PR push. Those gates were moved to the release
+# workflow (`release.yml::build-macos`), which now runs fmt/clippy/test before
+# producing signed artifacts. Local developers should run `make verify` (or
+# `make verify-clean` for a cache-clear pass) before pushing to catch
+# regressions that release-time CI will otherwise surface.
 #
 # Note on CUDA gating: CUDA verification stays exclusive to release.yml because
 # it requires a Linux self-hosted runner (currently only the GB10 node used for
@@ -76,29 +82,3 @@ jobs:
           components: rustfmt
       - run: cargo fmt --all -- --check
 
-  clippy-and-test:
-    name: clippy + test (macOS ARM64)
-    needs: changes
-    # Only run when Rust files changed; block fork-PR access to the self-hosted
-    # runner pool (mirrors the guard in release.yml and pipeline-parallel-ci.yml).
-    if: >-
-      needs.changes.outputs.rust == 'true' &&
-      github.repository == 'lablup/mlxcel' &&
-      (github.event_name != 'pull_request' ||
-        github.event.pull_request.head.repo.full_name == github.repository)
-    runs-on: self-hosted-macos-26-arm64
-    permissions:
-      contents: read
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          persist-credentials: false
-      - uses: dtolnay/rust-toolchain@stable
-        with:
-          components: clippy
-      - name: Build (verify MLX FFI compiles)
-        run: cargo build --release --features metal,accelerate
-      - name: Clippy
-        run: cargo clippy --all-targets --features metal,accelerate -- -D warnings
-      - name: Test
-        run: cargo test --release --features metal,accelerate

.github/workflows/release.yml

@@ -96,6 +96,21 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: aarch64-apple-darwin
+          components: clippy, rustfmt
+
+      # Quality gate — fmt / clippy / test. Used to live in ci.yml's
+      # `clippy + test (macOS ARM64)` job, but consumed ~15 min of the shared
+      # self-hosted runner on every PR push for limited PR-time signal. Moved
+      # here so the cost is paid once per release. Local devs should run
+      # `make verify` before pushing to catch regressions earlier.
+      - name: Format check
+        run: cargo fmt --all -- --check
+
+      - name: Clippy
+        run: cargo clippy --all-targets --features metal,accelerate -- -D warnings
+
+      - name: Test
+        run: cargo test --release --features metal,accelerate
 
       - name: Build release binaries
         run: cargo build --release --target aarch64-apple-darwin --locked