[Fix] zig: the Windows target builds again (timespec, socket and nap)

zig/src/cuda/roce.zig

@@ -449,11 +449,24 @@ pub const Gather = struct {
         for (0..g.n_hca) |h| try g.drain(h);
     }
 
+    /// A 20 us nap; where std.c has no timespec (Windows) the nap is simply absent.
+    fn napTimespec() ?std.c.timespec {
+        const T = @TypeOf(std.c.timespec);
+        if (T == void) return null;
+        switch (@typeInfo(T)) {
+            .@"struct" => {},
+            else => return null,
+        }
+        const v: std.c.timespec = .{ .sec = 0, .nsec = 20_000 };
+        return v;
+    }
+
     /// The proxy: spins on the doorbell and posts each sequence in order (a missed one from its slot's size).
     fn proxy(g: *Gather) void {
         const c = g.ctrl();
         var idle: u64 = 0;
-        const nap: std.c.timespec = .{ .sec = 0, .nsec = 20_000 };
+        // std.c.timespec is void on Windows (no CRT nanosleep), so the nap is built only where it exists.
+        const nap: ?std.c.timespec = napTimespec();
         while (g.running.load(.monotonic)) {
             const seq = @atomicLoad(u32, &c[c_seq], .acquire);
             if (seq == g.last_seq) {
@@ -462,7 +475,11 @@ pub const Gather = struct {
                     g.failed.store(true, .release);
                     return;
                 };
-                if (idle >= idle_spins) _ = std.c.nanosleep(&nap, null);
+                if (idle >= idle_spins) {
+                    if (nap) |*t| {
+                        if (@TypeOf(std.c.nanosleep) != void) _ = std.c.nanosleep(t, null);
+                    }
+                }
                 continue;
             }
             idle = 0;

zig/tests/cuda/exl3_tp2_tests.zig

@@ -1,5 +1,6 @@
 //! EXL3 tensor parallel: each rank's shard of real RED-SNOW projections and experts vs the engine's, all-gathered.
 const std = @import("std");
+const builtin = @import("builtin");
 const posix = std.posix;
 const cuda = @import("cuda");
 const check = @import("check.zig");
@@ -336,6 +337,8 @@ fn bf16r(v: f32) u16 {
 
 /// Rank 0 binds host:port and sends its unique id to the one peer (70 s for rank 1's retries).
 fn serve(addr_text: []const u8, bytes: []const u8) !void {
+    // Windows hands socket() back as SOCKET (usize), so the fd-shaped cast below cannot compile there.
+    if (builtin.os.tag == .windows) return error.SocketFailed;
     const a = try parseAddr(addr_text);
     var storage: std.Io.Threaded.PosixAddress = undefined;
     const len = try posixLen(a.host, a.port, &storage);
@@ -375,6 +378,8 @@ fn serve(addr_text: []const u8, bytes: []const u8) !void {
 
 /// Rank 1 connects to host:port, retrying for 60 s (rank 0 may not have bound yet), and reads the id.
 fn receive(gpu: Gpu, addr_text: []const u8, bytes: []u8) !void {
+    // Windows hands socket() back as SOCKET (usize), so the fd-shaped cast below cannot compile there.
+    if (builtin.os.tag == .windows) return error.SocketFailed;
     const a = try parseAddr(addr_text);
     const ip = try std.Io.net.IpAddress.parseIp4(a.host, a.port);
     var fd: posix.socket_t = -1;

zig/tests/cuda/flashnext_rank_tests.zig

@@ -1,5 +1,6 @@
 //! Real two-process/two-device NCCL sums and bounded disconnect/timeout behavior over the framed control link.
 const std = @import("std");
+const builtin = @import("builtin");
 const cuda = @import("cuda");
 const flash = @import("flashnext");
 const check = @import("check.zig");
@@ -45,8 +46,7 @@ pub fn run(gpu: check.Gpu, rank: usize, port: u16, mode: []const u8) !void {
             if (std.mem.eql(u8, mode, "eof")) {
                 _ = std.c.shutdown(link.fd, std.posix.SHUT.RDWR);
             } else {
-                const delay: std.c.timespec = .{ .sec = 0, .nsec = 600_000_000 };
-                _ = std.c.nanosleep(&delay, null);
+                sleepSixTenths();
             }
             check.pass("rank 1: injected {s}", .{mode});
             return;
@@ -119,3 +119,11 @@ fn completeSum(c: *flash.collective.Collective, stream: cuda.Stream, values: u64
     try c.sum(stream, values, count);
     try c.finish(stream);
 }
+
+/// The injected 0.6 s peer delay. Windows' CRT has neither nanosleep nor a real std.c.timespec,
+/// and this test never runs there.
+fn sleepSixTenths() void {
+    if (builtin.os.tag == .windows) return;
+    const delay: std.c.timespec = .{ .sec = 0, .nsec = 600_000_000 };
+    _ = std.c.nanosleep(&delay, null);
+}