Added a regression suite for the FreeRTOS compatibility layer

[fdesbiens]

Gives the FreeRTOS compatibility layer its first regression coverage in this repository.

Depends on #582. This branch is stacked on the #570 fix, which appears here as the first of two commits, because the queue tests assert the behaviour that PR corrects. Merge #582 first and I will rebase.

Why this suite exists

Every creation function in the layer takes one or two byte pool allocations for its bookkeeping and then creates one or more ThreadX kernel objects. When a kernel object cannot be created, the function returns NULL, or pdFAIL for xTaskCreate(). The caller has no handle, so it cannot call the matching delete function, and whatever the layer failed to release is gone until the system restarts.

That makes these paths invisible from the outside: a leaking version and a correct version return exactly the same thing to the caller. It is how the leak in #570 survived, and inspection is the only thing standing between us and the next one.

So the suite does not test return values alone. It counts the ThreadX primitives the layer reaches for, and checks that each error path gives back precisely what it took.

What it covers

Test	Covers
txfr_queue_create_test	xQueueCreate, xQueueCreateStatic, vQueueDelete, including the two paths fixed in #582
txfr_task_create_test	xTaskCreate semaphore and thread failure paths, xTaskCreateStatic
txfr_sync_create_test	counting and binary semaphores, mutexes, recursive mutexes, event groups, timers

25 checks across three ctest cases. This is deliberately a first course rather than full coverage of a 2,800 line layer: the infrastructure is the hard part, and coverage can now grow a test at a time.

How the injection works

A test asks the harness to fail a chosen kernel creation call, then reads back the counts. Interception uses the linker's --wrap, so tx_freertos.c is compiled exactly as it ships, with no test hooks in product code.

Two traps worth knowing, both documented in test/freertos/readme.md:

• tx_api.h maps the public API onto the error checking entry points, so the symbols that exist at link time are the _txe_ variants. Wrapping tx_semaphore_create does nothing. Worse, --wrap on a name that does not resolve is silently ignored, so a typo produces a test that quietly never injects anything and passes for the wrong reason.
• txfr_malloc() and txfr_free() cannot be wrapped at all, since they are defined in tx_freertos.c and called from within it, so the compiler resolves those calls internally. The byte pool counts stand in for them.

Validation

A test suite that has only ever seen correct code proves nothing, so I built it against the layer as it stands on dev, without the #582 fix. Exactly the two expected checks fail, with the leaked counts:

read_sem failure releases both allocations           FAIL
                                                       got      alloc=2 release=0 create=1 delete=0
                                                       expected alloc=2 release=2 create=1 delete=0
write_sem failure unwinds read_sem and memory        FAIL
                                                       got      alloc=2 release=0 create=2 delete=0
                                                       expected alloc=2 release=2 create=2 delete=1

Against the fixed layer, all three tests pass. Both CI scripts were run end to end locally.

Two findings from building this

Neither is fixed here, since both are in shipped code rather than in the tests:

FreeRTOS.h selects the interrupt primitives by compiler, not by target. A GNU build resolves portDISABLE_INTERRUPTS() to the bare metal __disable_interrupts() intrinsic, so hosting the layer on Linux with GCC fails to link in vPortEnterCritical(). The macros are #ifndef guarded and the header already carries a ThreadX based fallback for the #else branch, so the test config simply pre-empts the choice. A proper fix would gate the __GNUC__ branch on the target rather than the compiler.

The layer passes pointers through ULONG arguments — the task argument and the timer identifier among them — while the Linux port defines ULONG as unsigned int on x86_64. A 64 bit build truncates them, which GCC reports as -Wpointer-to-int-cast and which would crash the timer callback wrapper. The suite builds 32 bit, as the ThreadX and SMP suites do, so this is contained for now, but it is a real constraint on hosting the layer on a 64 bit target.

Notes for review

• Linux only, because --wrap has no MSVC equivalent. The CMake configuration stops with an explicit message rather than failing later at link time.
• Registered with ctest and wired into .github/workflows/regression_test.yml as a freertos job alongside tx and smp, with skip_coverage: true since there is no coverage build configuration yet. The deploy job's needs list is left alone, matching how the RISC-V job was configured.
• The strict warning set is applied to the test code only. Neither ThreadX nor the layer builds clean under -Wextra today, and that is not this suite's to fix.