[andreguzik]

Problem

build.rs unconditionally emits /STACK:8388608 on any Windows build:

#[cfg(windows)]
{
    println!("cargo:rustc-link-arg=/STACK:8388608");
}

/STACK: is MSVC linker syntax. On x86_64-pc-windows-gnu (MinGW), the GNU
linker doesn't understand it and the build fails:

x86_64-w64-mingw32-gcc ... -o rtk.exe ...
ld.exe: cannot find /STACK:8388608: No such file or directory
collect2.exe: error: ld returned 1 exit status

This breaks cargo build/cargo test/cargo install --path . for anyone
whose default Rust toolchain is stable-x86_64-pc-windows-gnu instead of
stable-x86_64-pc-windows-msvc (a common setup for contributors who don't
have Visual Studio Build Tools installed).

Fix

Select the flag syntax based on CARGO_CFG_TARGET_ENV (set by Cargo for
build scripts): keep /STACK:8388608 for msvc, use the GNU-ld equivalent
-Wl,--stack,8388608 otherwise. Also switched the outer guard from
#[cfg(windows)] (host-compile-time) to checking CARGO_CFG_TARGET_OS
(the actual build target), which is the more correct signal for a build
script and matches how CARGO_CFG_TARGET_ENV is already being read.

Testing

• cargo fmt --all --check
• cargo clippy --all-targets (no new warnings)
• cargo test — 2575 passed, 8 ignored, 0 failed (previously: build didn't
  complete at all on this target)
• cargo install --path . succeeds and produces a working rtk.exe on
  stable-x86_64-pc-windows-gnu

No behavior change on x86_64-pc-windows-msvc (still emits /STACK:) or
non-Windows targets (flag is skipped entirely, same as before).

Build-script-only change; not user-facing, so no docs update needed per
CONTRIBUTING.md.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[KuSh]

Closing in favour of #3613, which makes the same windows-gnu stack flag change in build.rs, keys it off the build target rather than the host, and has a signed CLA. Review and comments are welcome on #3613, or reply here if your version covers a case it does not.