ci: add GPU E2E job on self-hosted runner (app-dev)

Runs all 20 cucumber scenarios on the app-dev-gpu self-hosted runner
with real AMD GPU hardware. Fails the build on test failures — no
continue-on-error. Expected-failure scenarios will fail until the
corresponding bugs are fixed.

Signed-off-by: fredespi <fredrik.espinoza@gmail.com>

Author: fredespi

.github/workflows/ci.yml

@@ -414,12 +414,10 @@ jobs:
   # ── E2E tests (cucumber-rs) ───────────────────────────────────────────────
   #
   # BDD scenarios in Gherkin (.feature files) backed by Rust step functions.
-  # Runs on hosted runners — mock-tier tests use an in-process mock server,
-  # GPU-tier tests are tagged @expected-failure and skipped until hardware
-  # runners are available.
+  # Two tiers: hosted (mock server, no GPU) and self-hosted (real GPU hardware).
 
   e2e:
-    name: E2E tests (cucumber-rs)
+    name: E2E tests
     runs-on: ubuntu-latest
     needs: [changes, build-and-test]
     if: needs.changes.outputs.heavy == 'true'
@@ -445,7 +443,25 @@ jobs:
           name: e2e-report
           path: tests/e2e-cucumber/results/
 
-  # Hardware (real AMD GPU / WSL) smoke tests on dedicated self-hosted runners
-  # are planned beyond the E2E tests above. See docs/ci-hardware-testing.md
-  # for the full design covering per-engine GPU validation, WSL testing, and
-  # Windows native testing.
+  e2e-gpu:
+    name: E2E tests (GPU)
+    runs-on: [self-hosted, linux, amd-gpu]
+    needs: [changes, build-and-test]
+    if: needs.changes.outputs.heavy == 'true'
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: actions-rust-lang/setup-rust-toolchain@v1
+
+      - name: Run E2E tests on GPU hardware
+        run: |
+          cargo build --release -p rocm
+          ROCM_CLI_BINARY=${{ github.workspace }}/target/release/rocm \
+            cargo test -p e2e-cucumber --test e2e
+
+      - name: Upload E2E report
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: e2e-gpu-report
+          path: tests/e2e-cucumber/results/