StreamTxDemo: fix mingw/Windows-GCC build and binary stdin

[AlaEddine-Zoghlami]

Problem

Building StreamTxDemo (the stream-link TX demo) with mingw-w64 (GCC on Windows) is broken in two independent ways. Both are mingw-only and do not affect the MSVC/vcpkg build.

1. Binary stdin not set on mingw → first frame truncated

txdemo/stream_tx_demo/main.cpp gates both the Windows includes (io.h/fcntl.h) and the _setmode(_fileno(stdin), _O_BINARY) call on #if defined(_MSC_VER). mingw/GCC defines _WIN32 but not _MSC_VER, so it falls into the POSIX #else branch and leaves stdin in text mode. A 0x1A (or CRLF) byte in the binary <u32_le len><PSDU> stream then triggers:

stream_tx_demo: short read on stdin (76/269); record truncated mid-PSDU

and the demo exits before transmitting a single frame.

2. CMake aborts under mingw (no vcpkg)

CMakeLists.txt unconditionally sets CMAKE_TOOLCHAIN_FILE to "$ENV{VCPKG_ROOT}/scripts/buildsystems/vcpkg.cmake" on WIN32. With VCPKG_ROOT unset (the normal mingw / pkg-config case) this expands to the bogus /scripts/buildsystems/vcpkg.cmake and configuration fails with "Could not find toolchain file".

Fix

• Add a __MINGW32__ || __MINGW64__ include branch (io.h + fcntl.h + unistd.h + libusb) and gate _setmode on _WIN32 — identical behavior on MSVC, now also correct on mingw.
• Only set the vcpkg toolchain file when VCPKG_ROOT is non-empty.

Verification

mingw-w64, libusb 1.0.28 via pkg-config, real RTL8812AU. After the fix StreamTxDemo.exe reads the full binary stream and radiates every frame:

<stream-tx>TX #1 ok=1 psdu=269 total=306
...
<stream-tx>done; sent 512 PSDUs

🤖 Generated with Claude Code

[josephnef]

Reviewed the diff against the source and CI. Both fixes are correct and minimal:

• _setmode regate _MSC_VER → _WIN32 + the __MINGW32__/__MINGW64__ include branch — correct. mingw needs io.h+fcntl.h for _setmode/_O_BINARY and unistd.h for sleep() (used at L214); pid_t and <windows.h> aren't referenced in this file, so their absence from the new branch is harmless. MSVC is unchanged — it defines both _MSC_VER and _WIN32, so it still takes the first include branch and still calls _setmode.
• CMake vcpkg guard — correct, and correctly placed before project() (required for CMAKE_TOOLCHAIN_FILE to take effect). CI's Windows job always sets VCPKG_ROOT, so there's no CI behavior change; the only effect is the unset-VCPKG_ROOT (mingw) case, which is the intent.

One gap before merge: txdemo/stream_duplex_demo/main.cpp has the byte-for-byte identical pattern — include block gated on _MSC_VER (L38) and _setmode gated on _MSC_VER (L202–204) — and it reads the same <u32_le len><PSDU> binary stream via read_exact. So it hits the exact first-frame text-mode truncation this PR fixes for StreamTxDemo. Worse: this PR's CMake change is what makes the duplex demo buildable under mingw in the first place, so as written this PR effectively ships a buildable-but-broken StreamDuplexDemo (mingw isn't in CI, so nothing catches it).

Please mirror the same _WIN32 gate + mingw include branch into stream_duplex_demo/main.cpp so the two stdin-driven stream demos stay in lockstep.

For completeness: precoder_demo/main.cpp and txdemo/main.cpp also gate includes on _MSC_VER, but neither reads length-prefixed stdin nor calls _setmode, so they're unaffected — no change needed there.

[AlaEddine-Zoghlami]

Good catch — fixed in 2880c0d.

stream_duplex_demo/main.cpp had the identical _MSC_VER-gated include block + _setmode, and since this PR's CMake guard is exactly what makes the duplex demo buildable under mingw, you're right that as-was it would ship buildable-but-broken. Mirrored the same change: __MINGW32__/__MINGW64__ include branch + _setmode gated on _WIN32.

Verified on mingw-w64 (libusb 1.0.28) with the real RTL8812AU — StreamDuplexDemo now reads the full binary stream end to end:

<stream-duplex>tx #1 ok=1 psdu=269
...
<stream-duplex>tx #500 ok=1 psdu=269
<stream-duplex>tx EOF after 512 PSDUs

(pre-fix it died at short read on stdin (76/269) on frame 1).

Agreed on precoder_demo/main.cpp and txdemo/main.cpp — they gate includes on _MSC_VER too but neither reads length-prefixed stdin nor calls _setmode, so they're unaffected; left them as-is. The two stdin-driven stream demos are now in lockstep.

[josephnef]

Approving — the duplex-demo gap is resolved in 2880c0d (exact mirror of the StreamTxDemo fix; verified on hardware), and all 5 required CI checks are green. The two stdin-driven stream demos are now in lockstep.