Fix CMake/CTest configuration issues

[nehalkpatel]

Summary

An audit of the CMake/CTest setup found six issues, each verified empirically against the build rather than inferred. The most serious: Release builds were not actually -O3. Also adds the missing RelWithDebInfo presets, pins the last unpinned dependency, and collapses the three-places-per-test declaration pattern into a single helper.

Every change is verified from a wiped build/ directory. All three workflows pass 29/29.

Release was silently size-optimized

COMMON_COMPILE_OPTIONS appended -Os -g after the per-config flags, and the last -O on the command line wins:

Config	Flags before	Flags after
Debug	-g -Os -g	-g
Release	-O3 -DNDEBUG -Os -g	-O3 -DNDEBUG
RelWithDebInfo	(unreachable)	-O2 -g -DNDEBUG

Proven with a vectorizable loop: -O3 → 243 bytes of .text, -Os → 107, -O3 -Os → 107. So Release was size-optimized and Debug was optimized rather than debuggable.

Both flags were redundant anyway — cmake/toolchain/armgcc.cmake already sets CMAKE_C_FLAGS_DEBUG_INIT "-O0" and CMAKE_C_FLAGS_RELEASE_INIT "-Os -flto" plus -g3, so the ARM path is unaffected.

Every CTest property on the integration test was inert

All three set_tests_properties calls did nothing:

• The first set DEPENDS "blinky;uart_echo;i2c_demo", the second set DEPENDS host_emulator_venv. set_tests_properties replaces rather than appends, so the first was silently discarded — the generated CTestTestfile.cmake showed only the second.
• DEPENDS names tests, not targets. blinky et al. are executables, and host_emulator_venv is created at configure time by execute_process, so it isn't in the test registry at all.
• FIXTURES_SETUP declares the test is a fixture for others; nothing declared a matching FIXTURES_REQUIRED.

The intent was presumably "build the apps first", but ctest does not build — verified by deleting build/host/bin and running the ctest preset alone: 28 tests "Not Run", integration test failed. The workflow presets exist for this and are what the docs, CI and docker-compose already use.

Replaced with TIMEOUT 300 (previously a hang would block until CTest's 1500s default) and LABELS "integration".

Presets

• default was user-visible but unbuildable — cmake --preset=default failed with add_subdirectory called with incorrect number of arguments (EMBEDDED_CPP_BOARD/_MCU unset). Now hidden, matching arm/arm-cm4/arm-cm7.
• RelWithDebInfo had zero presets despite being in CMAKE_CONFIGURATION_TYPES and having build-RelWithDebInfo.ninja generated. It was reachable only via a bare --config. Now has build/test/workflow presets — which matters more since Release no longer carries -g.
• host build and test presets are now hidden inheritance bases. The test preset was identical to host-debug; the build preset carried no configuration. Both relied on a trailing -C/--config to mean anything. The hidden default test preset already held the output/execution settings, but the host presets duplicated them verbatim — chaining default → host → per-config removes that.

Also fixed

-R test_zmq_transport matched no tests. gtest_discover_tests registers ZmqTransportTest.SendReceive — named for the fixture, not the source file — and noTestsAction: error made the documented command fail outright. Now -R ZmqTransportTest.

uv run --extra dev was broken. fe03b14 moved dev deps to a PEP 735 [dependency-groups] table but left CLAUDE.md documenting the old flag, which now errors: Extra 'dev' is not defined in the project's optional-dependencies table. uv syncs groups by default, so plain uv run is correct.

cmake-scripts tracked main. It supplies clang_tidy(), reset_clang_tidy(), add_code_coverage_all_targets() and target_code_coverage(), so upstream drift could change lint enforcement or coverage silently. Pinned to 25.08, verified from a clean build directory.

Test declaration collapsed into one helper

Each unit test was declared in three separate places — executable+options+libs, gtest_discover_tests, and target_code_coverage. Nothing tied them together, so a test added to one list but not the others failed silently. add_host_unit_test(name source [libs...]) makes it one line.

source is a parameter rather than derived as ${name}.cpp because test_host_transport is built from test_zmq_transport.cpp; a convention-based helper would break that target.

Behaviour verified identical by comparison, not just a passing build: 30 instrumented TUs and 136 ccov targets before and after.

Test plan

From a wiped build/:

• host-debug, host-release, host-relwithdebinfo — 29/29 each
• Per-config flags match the config name (table above)
• LABELS "integration" and TIMEOUT "300" present in generated CTestTestfile.cmake
• ctest -L unit → 28, -L integration → 1, total 29
• cmake-scripts fetches 25.08; clang-tidy found; ccov-all generates its HTML report
• All six documented commands run verbatim

Notes

• Breaking for muscle memory: ctest --preset=host and cmake --build --preset=host no longer resolve. All documented commands are updated; docker-compose.yml and CI are unaffected (both use --workflow --preset host-debug).
• ARM presets remain non-functional — src/libs/mcu/arm_cm4/ doesn't exist yet. Expected, and the docs say so.
• ccov-all emits warning: 70 functions have mismatched data. Pre-existing, reproduces from a wiped tree with a single Debug build and no changes. Cause is header-defined inline/template functions (the mcu/board/error INTERFACE libraries) instantiated into all five test binaries, then merged. Benign — the report still generates.
• The branch name predates most of this work; it's now mostly CMake/CTest fixes rather than the presets example.