diff --git a/.github/actions/install-rust/action.yml b/.github/actions/install-rust/action.yml
index 04bee8af..f4d9168e 100644
--- a/.github/actions/install-rust/action.yml
+++ b/.github/actions/install-rust/action.yml
@@ -10,9 +10,14 @@ runs:
   steps:
     - name: Install rust-toolchain.toml
       shell: bash
-      run: rustup show
+      run: rustup toolchain install
     - name: Install additional rustup components
       if: inputs.components != ''
       shell: bash
       run: echo "${{ inputs.components }}" | tr ',' '\n' | xargs rustup component add
     - uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2
+      with:
+        # Avoid caching ~/.cargo/bin on macOS — restoring it on a runner with a
+        # different rustup version overwrites the cargo proxy with a stale
+        # rustup-init, breaking `cargo fmt`/`clippy`/etc.
+        cache-bin: ${{ runner.os != 'macOS' }}
diff --git a/skills/codspeed-setup-harness/SKILL.md b/skills/codspeed-setup-harness/SKILL.md
index 1fe35771..d8f7fe1b 100644
--- a/skills/codspeed-setup-harness/SKILL.md
+++ b/skills/codspeed-setup-harness/SKILL.md
@@ -287,7 +287,7 @@ Good benchmarks are representative, isolated, and stable. Here are guidelines:
 
 - **Avoid benchmarking setup**: Use the framework's setup/teardown mechanisms to exclude initialization from measurements.
 
-- **Prevent dead code elimination**: Use `black_box()` (Rust), `benchmark.pedantic()` (Python), or equivalent to ensure the compiler/runtime doesn't optimize away the work you're measuring.
+- **Prevent dead code elimination**: Use `black_box()` (Rust), `benchmark::DoNotOptimize` (C++), or `Blackhole.consume` (JMH) so the compiler doesn't optimize away unused results.
 
 - **Cover the critical path**: Benchmark the functions that matter most to your users — the ones called frequently or on the hot path.