diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 72eec96..11a14be 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -97,6 +97,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
       - uses: actions/setup-node@v4
         with:
           node-version: '22'
@@ -122,13 +124,26 @@ jobs:
             libxkbcommon0 \
             libxrandr2 \
             libasound2t64
-      # No compiler build needed — the `.nes` files are committed
-      # alongside their `.ne` sources, and the `examples` CI job
-      # fails on any stale ROM, so the emulator job can trust them.
+      - name: Build compiler
+        run: cargo build --release
+      - name: Compile all .ne examples
+        # Overwrites the committed examples/*.nes in the CI workspace
+        # (the workspace is ephemeral; the committed ROMs aren't
+        # touched). This is load-bearing: the goldens must be diffed
+        # against what *this* compiler produces, not against the ROMs
+        # an earlier commit happened to ship, otherwise a broken
+        # compiler change can still pass the emulator job as long as
+        # the contributor forgot to rebuild the ROMs. The `examples`
+        # job catches that second case via its reproducibility diff.
+        run: |
+          for f in examples/*.ne; do
+            echo "==> Compiling $f"
+            ./target/release/nescript build "$f"
+          done
       - name: Install emulator harness dependencies
         working-directory: tests/emulator
         run: npm ci
-      - name: Diff each ROM's framebuffer against its committed golden
+      - name: Diff each ROM's framebuffer against its golden
         working-directory: tests/emulator
         run: node run_examples.mjs
       - name: Upload actual + diff PNGs on failure
diff --git a/CLAUDE.md b/CLAUDE.md
index 3be1741..4df3038 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -70,29 +70,31 @@ tests/emulator/
 The harness is **separate** from `cargo test`. You have to run it by hand:
 
 ```bash
-# 1. Install node deps (once per worktree; node_modules/ is gitignored).
+# 1. Rebuild every example with the current compiler. The harness
+#    reads whatever sits under examples/*.nes — if you want to test
+#    your working copy you have to rebuild them first.
+cargo build --release
+for f in examples/*.ne; do ./target/release/nescript build "$f"; done
+
+# 2. Install node deps (once per worktree; node_modules/ is gitignored).
 cd tests/emulator
 npm install          # or `npm ci` in CI
 
-# 2. Verify every committed ROM still matches its golden.
+# 3. Verify every ROM still matches its golden.
 node run_examples.mjs
 # → "22/22 ROMs match their goldens" on success
 # → FAIL / MISS lines + `actual/<name>.png`, `actual/<name>.diff.png`,
 #   `actual/<name>.wav` written for any ROM that mismatched
 ```
 
-The harness reads the committed `examples/*.nes` files directly —
-since those are rebuilt and diffed by the `examples` CI job, the
-emulator job doesn't need a compiler toolchain at all. If you're
-iterating on the compiler locally, rebuild the example ROM(s) you
-care about before re-running the harness:
-
-```bash
-cargo build --release
-./target/release/nescript build examples/hello_sprite.ne
-# then:
-(cd tests/emulator && node run_examples.mjs)
-```
+The harness always runs against whatever sits in `examples/*.nes`,
+so iterating on the compiler means rebuilding the example first.
+CI's `emulator` job does this too — it builds the compiler, compiles
+every `.ne` into the workspace (overwriting the committed ROMs,
+which are ephemeral in the CI checkout), and then runs the harness.
+The committed ROMs are a PR-review convenience and a "did this
+change affect codegen" tripwire via the `examples` job's
+reproducibility diff; they are **not** what the emulator job tests.
 
 ### Updating goldens