Story 06 — Native runner tests in CI + LLM-behavior evals

[Lykhoyda]

The riskiest layer has zero automated execution: Swift suites + Kotlin KeyboardGuardTest exist in-tree but CI only compiles them (CodeQL). Phase A: run gradlew testDebugUnitTest (ubuntu) + xcodebuild test -only-testing:RnFastRunnerTests (macOS), path-filtered. Phase B: nightly device smoke — prebuilt artifacts (Story 01), pinned sim+AVD, tiny fixture app, golden command set driven through the real bridge over MCP stdio; 2-consecutive-red alerting. Phase C: mcp-server-tester-style LLM-behavior evals (Maestro: 'test that LLMs can call it correctly… happens less frequently than is expected') — baseline gates Stories 08/12.

Spec: https://github.com/Lykhoyda/rn-dev-agent/blob/limestone-malpais/docs/stories/06-native-runner-ci-and-evals.md (PR #381)
Impact: coverage where the hardest bugs live · Effort: M · Depends: #382 (Story 01, for Phase B)

[Lykhoyda]

Phase A shipped in #464 (merge 277bc81, 2026-07-05).

• .github/workflows/native-tests.yml: Android unit (JVM) (gradlew testDebugUnitTest, 1m32s observed) + iOS unit (simulator) (xcodebuild test on macos-15 via scripts/test-native-ios.sh, 4m26s observed) — path-filtered with green skip-notices, unconditional on main.
• iOS skip-list (CLI-only, by design): RnFastRunnerUITests/RnFastRunnerTests (production server entry) + SnapshotForegroundRegressionTest (device-dependent). Never move these into the xctestplan — build-for-testing would bake SkipTestIdentifiers into the production .xctestrun and the runner's -only-testing:…/testCommand launch would never boot.
• Mutation check (acceptance criterion): both lanes red on a deliberately broken KeyboardGuard — run 28723742068 — then reverted and green.
• Local entry points: npm run test:native:android / npm run test:native:ios.

Remaining in this issue: Phase B (nightly device smoke — unblocked by #382's prebuilt artifacts once a release ships them) and Phase C (LLM-behavior evals — needs a budget decision).