When modifying integration-test execution or environment setup, prioritize determinism and behavioral equivalence. 1) Don’t assume test-runner flags are interchangeable - If a flag is known to be “experimental” but was added for harness-specific behavior, treat it as part of the stability contract. - If you replace it (e.g., to avoid toolchain deprecations), require proof that the new flag reproduces the same behavior; otherwise keep the original behavior behind a switch. 2) Make environment cleanliness an explicit, configurable default - Default to the cleanest setup that yields repeatable results (including removing cached state like kernels/data that can affect outcomes). - If you want faster runs by preserving state, require an explicit opt-in environment variable (e.g., `PRESERVE_KERNELS=true`) and document it in the Makefile. Example pattern (Makefile-style): ```make # Deterministic defaults PRESERVE_KERNELS ?= false PARALLEL_WIDTH ?= 2 define CLEAN_APP_DATA @echo "Clearing application data under $(APP_ROOT) ..." ; \ mkdir -p $(APP_ROOT) ; \ if [ "$(PRESERVE_KERNELS)" = "true" ]; then \ find "$(APP_ROOT)" -mindepth 1 -maxdepth 1 ! -name kernels -exec rm -rf {} + ; \ else \ find "$(APP_ROOT)" -mindepth 1 -maxdepth 1 -exec rm -rf {} + ; \ fi endef # Prefer keeping the known-stable harness behavior unless proven equivalent # (gate by toolchain/version or expose a switch) SWIFT_PARALLEL_ARGS ?= --experimental-maximum-parallelization-width $(PARALLEL_WIDTH) define RUN_CONCURRENT_PASS $(SWIFT) test $(INTEGRATION_SWIFT_EXTRA) -c $(BUILD_CONFIGURATION) $(SWIFT_CONFIGURATION) \ $(SWIFT_PARALLEL_ARGS) \ --filter "$(CONCURRENT_FILTER)" endef ``` Apply this standard to any changes in Testing infrastructure code (Makefile scripts, test harnesses, parallelism config, setup/teardown logic) to prevent flaky integration results and toolchain-dependent breakage.