Fix Windows CI: pin windows-2022 -- new image's clang-cl 22 rejects bundled SDL's _m_prefetch

[arekbr]

The Windows job dies compiling MTEngineSDL (single error in the whole log):

SDL_endian.h(40,1): error : definition of builtin function '_m_prefetch'

Nothing changed in either repo. The windows-2025-vs2026 runner image was updated between the last green run (June, clang-cl 20.1.8) and now (clang-cl 22.1.3). Since clang 21, _m_prefetch is a compiler builtin, and the SDL 2.28.5 header bundled in MTEngineSDL redefines it under a plain #ifdef __clang__ (an old Clang-11-vs-winnt.h workaround). There were no CI runs between June and the #114/#115 merges, so the failure surfaced on the first push after the image update -- the merged PRs touch neither MTEngineSDL nor the workflow. (Both June and August runs used the same VS-2026 image family and the same MSVC 14.51 -- the only variable that moved is the bundled clang: 20.1.8 green, 22.1.3 red.)

This PR: pin windows-2022 (VS 2022, clang-cl 19.x, servicing-only -- its clang stays below 21). One-line change, buys a working CI immediately.

The proper fix is a one-line guard change in MTEngineSDL's bundled header, mirroring what upstream SDL2 did on their branch (not yet in any SDL2 release -- 2.32.10 still has the plain #ifdef):

-#ifdef __clang__
+#if defined(__clang__) && !_SDL_HAS_BUILTIN(_m_prefetch)

(with the underscore -- _SDL_HAS_BUILTIN is the SDL2 spelling; upstream's first attempt used the SDL3 name SDL_HAS_BUILTIN and did not compile, they fixed it in a follow-up commit. Safe for older toolchains: clang 11-20 report no such builtin, so the workaround stays active for them.) Since AGENTS.md says MTEngineSDL changes need your sign-off: say the word and I'll send that PR there too, then this pin can go back to windows-latest.

🤖 Generated with Claude Code

[arekbr]

Additional measurements backing both the pin and the proper patch, done on real toolchains:

clang-cl 19.1.7 (the version windows-2022 ships): both the current #ifdef __clang__ branch and the patched guard compile fine; __has_builtin(_m_prefetch) = 0 there, so the workaround stays active. Confirms the pin is safe and the patch changes nothing for older toolchains.

clang 22.1.8 (Homebrew LLVM, same major as the failing runner): __has_builtin(_m_prefetch) = 1, and preprocessing shows the patched guard cleanly cuts the redefinition out, letting the compiler builtin take over -- exactly the mechanism that removes the CI error. The patched file compiles clean.

One honest caveat: I could not reproduce the hard error itself outside CI -- upstream LLVM 22.1.8's clang-cl accepts my minimal repro of the header block, while the runner's MS-built clang-cl 22.1.3 clearly rejects it in the log. Microsoft ships its own LLVM build, so the error's severity may differ between distributions. That does not change the math: the pinned clang 19 has no builtin (measured), and the patched guard is inert where the builtin is absent and removes the redefinition where it is present (measured).