[chrisuthe]

A Windows consumer that includes <windows.h> before the sendspin headers, as most Win32 code does, fails to compile them. The library's NOMINMAX / WIN32_LEAN_AND_MEAN / NOGDI from #210 are private, so they never covered the public headers.

Summary

• LogLevel::ERROR. wingdi.h defines ERROR as 0, which turned the enumerator into 0 = 1. client.h now sets the macro aside around LogLevel with #pragma push_macro("ERROR") / #undef ERROR / #pragma pop_macro("ERROR"), so the consumer's own ERROR is back after it. The pragmas are supported by GCC, Clang and MSVC, so the header needs no #ifdef _WIN32, and they are no-ops where ERROR is not defined.
• std::numeric_limits<uint16_t>::max() in PlayerRoleConfig::pipeline_lead_time_ms() was expanded by windows.h's function-like max() macro. It is now (std::numeric_limits<uint16_t>::max)(), with a comment so the parentheses are not tidied away.

Notes

• Consumer code that itself names LogLevel::ERROR after <windows.h> still needs NOGDI or its own #undef ERROR; the comment above LogLevel says so. Renaming the enumerator would remove that, but it would break the API for every consumer, so it is left alone.
• There is no unit test: on Linux, simulating windows.h's max() macro breaks the standard library headers before it reaches ours. The natural guard is a Windows CI job that compiles one translation unit including <windows.h> first and then every public header; that is planned as a follow-up along with building the library and basic_client on Windows in CI.

Testing

• Windows 11, clang-cl 22.1.3 (Visual Studio 2026), /W4, every role enabled, one translation unit including every public header:
  • <windows.h> first: failed with 3 errors before this change, compiles cleanly after it.
  • <windows.h> after the sendspin headers, with NOMINMAX + WIN32_LEAN_AND_MEAN, and without <windows.h>: all compile cleanly.
  • static_assert(ERROR == 0) after the headers holds, so pop_macro restores wingdi's macro.
  • With NOGDI, consumer code can name LogLevel::ERROR after <windows.h>; without it, it fails in the consumer's own line, as the header comment says.
• Windows 11, clang-cl: the sendspin target builds with no warnings from the library's own sources.
• Linux (WSL), GCC 13, -Wall -Wextra -Wpedantic: both headers compile, a consumer's own ERROR macro survives the include, and pipeline_lead_time_ms(0xFFFF) still saturates at 0xFFFF. The rest of Linux is this PR's CI.
• clang-format: no changes.

[Copilot]

🟡 Changes recommended

The macro workarounds lack automated regression coverage, so either could be removed without failing the existing host tests.

1 open issue

• Add regression coverage for macro workarounds and lead-time saturation · New

What changed in this PR

This PR extends the Windows build support from #210 by making public headers tolerate a consumer including <windows.h> first, without changing the public API or the consumer’s macro settings.

Changes:

• Temporarily saves and undefines ERROR around LogLevel, then restores it.
• Protects numeric_limits::max() from macro expansion while preserving lead-time saturation.

File	Description
include/​sendspin/​config.h	Parenthesizes max to prevent Windows macro expansion.
include/​sendspin/​client.h	Shields the enum declaration from ERROR and documents the remaining consumer-side limitation.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The Windows macro collision test omits the seven public role headers, allowing collisions in those headers to pass CI.

1 open issue

• Include all public role headers in Windows macro collision test · New

1 resolved since last review

• Add regression coverage for macro workarounds and lead-time saturation

_{Review effort: Balanced}