Testing & CI hardening phase 3: union × protocol conformance cells, coverage tooling (#48)

.github/workflows/ci.yml

@@ -146,6 +146,39 @@ jobs:
           (cd codegen && gradle generateFixtures generateProtocolTests)
           git diff --exit-code -- examples protocol-tests
 
+  # Coverage measurement for the runtime (issue #48): the combined lcov
+  # report + rendered HTML land as a build artifact and the per-module
+  # summary prints in the log, so srcs<->tests gaps stop needing a manual
+  # audit. Measurement only — no threshold gate yet; add one once the
+  # baseline has soaked (same posture as benchmarks below).
+  coverage:
+    name: coverage (runtime, lcov)
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v7
+      - uses: bazel-contrib/setup-bazel@0.19.0
+        with:
+          bazelisk-cache: true
+          repository-cache: true
+          disk-cache: coverage
+      - name: bazel coverage
+        run: |
+          bazelisk coverage //... --config=ci --combined_report=lcov \
+            --instrumentation_filter='//runtime[/:]'
+      - name: summarize and render
+        run: |
+          sudo apt-get update -q || true
+          sudo apt-get install -y -q lcov
+          REPORT="$(bazelisk info output_path)/_coverage/_coverage_report.dat"
+          lcov --summary "$REPORT"
+          lcov --list "$REPORT"
+          genhtml --output-directory coverage-html "$REPORT"
+          cp "$REPORT" coverage-html/coverage.lcov
+      - uses: actions/upload-artifact@v4
+        with:
+          name: coverage-report
+          path: coverage-html
+
   # Informational (PLAN Phase 7): publish numbers, no pass/fail threshold yet.
   benchmarks:
     name: benchmarks (informational)

CHANGELOG.md

@@ -93,5 +93,16 @@ via `git_override` until then.
   content-type/method/route) the way jsonrpc2's generated suite always did —
   including the previously-unasserted simpleRestJson `@pattern`-violation
   wire message.
+- **Union x protocol conformance cells filled** (`protocol-tests/unions/`):
+  the cbor and jsonRpc2 union cells — previously reliant on coin-flip random
+  integration tests that only prove serde self-consistency — now pin the
+  wire subdocument for every union variant deterministically in all four
+  directions (client encode/decode, server decode/echo), plus the reject
+  cells (empty, multi-member, unknown-member, null-member) and the `__type`
+  discriminator tolerance.
+- **Code-coverage tooling**: a `coverage` CI job runs
+  `bazel coverage --combined_report=lcov` over the runtime, prints the
+  per-module summary, and uploads the rendered HTML report as an artifact;
+  `make coverage` runs the same locally. Measurement only — no gate yet.
 
 [Unreleased]: https://github.com/aaylward/smithy-cpp/commits/main

Makefile

@@ -64,6 +64,14 @@ fuzz-smoke:
 consumer:
 	cd examples/bazel-consumer && $(BAZEL) test //... && ./model-evolution-check.sh
 
+# Line coverage for the runtime; the combined lcov report path prints at the
+# end (render with genhtml, or read the CI job's artifact).
+.PHONY: coverage
+coverage:
+	$(BAZEL) coverage //... --combined_report=lcov \
+		--instrumentation_filter='//runtime[/:]'
+	@echo "combined report: $$($(BAZEL) info output_path)/_coverage/_coverage_report.dat"
+
 # Informational, never gates (PLAN Phase 7).
 .PHONY: benchmarks
 benchmarks:

docs/development.md

@@ -24,8 +24,9 @@ make verify-full   # + sanitizers, fuzzer smoke runs, the consumer module, clang
 ```
 
 Each aggregate is also callable piecemeal (`make test codegen goldens lint
-sanitize fuzz-smoke consumer tidy benchmarks format`); the recipes mirror
-`.github/workflows/ci.yml`, one target per job. The underlying commands:
+sanitize fuzz-smoke consumer tidy coverage benchmarks format`); the recipes
+mirror `.github/workflows/ci.yml`, one target per job. The underlying
+commands:
 
 ```sh
 # C++ runtime: build + run all tests

protocol-tests/unions/BUILD.bazel

@@ -0,0 +1,40 @@
+load("@rules_cc//cc:defs.bzl", "cc_test")
+
+# Union x protocol conformance cells (issue #48): union round-tripping was
+# only pinned for simpleRestJson; the cbor and jsonRpc2 cells relied on
+# seeded random integration tests that flip a coin on whether the union
+# appears and can only prove serde self-consistency. These suites pin the
+# wire subdocument for every SinkChoice variant deterministically — client
+# encode/decode and server decode/echo — plus the reject cells and the
+# __type discriminator tolerance. Hand-written (not generated), so they
+# live outside the golden generated/ trees.
+
+cc_test(
+    name = "union_cbor_test",
+    size = "small",
+    srcs = ["union_cbor_test.cc"],
+    deps = [
+        "//examples/roundtrip/rpc/generated:client",
+        "//examples/roundtrip/rpc/generated:server",
+        "//runtime:cbor",
+        "//runtime:client",
+        "//runtime:http",
+        "//runtime:protocol_test_support",
+        "@googletest//:gtest_main",
+    ],
+)
+
+cc_test(
+    name = "union_jsonrpc2_test",
+    size = "small",
+    srcs = ["union_jsonrpc2_test.cc"],
+    deps = [
+        "//examples/roundtrip/jsonrpc/generated:client",
+        "//examples/roundtrip/jsonrpc/generated:server",
+        "//runtime:client",
+        "//runtime:http",
+        "//runtime:json",
+        "//runtime:protocol_test_support",
+        "@googletest//:gtest_main",
+    ],
+)