[renerocksai]

Why

The library module set .link_libc = true, so every consumer linked a libc even though the framework is pure Zig. Consumers such as omajot want one static Linux binary with no interpreter and no shared libraries.

What

• src/sys.zig: the 17 POSIX calls the framework makes outside the Windows transport (socket, setsockopt, bind, listen, getsockname, fcntl, close, connect, send, recv, poll, shutdown, pipe, read, write, clock_gettime, nanosleep) with libc's signatures and -1/errno convention. With libc linked (and on every OS but Linux) they are std.c; on Linux without libc they are direct std.os.linux system calls with a thread-local errno. Call sites in server.zig, transport.zig and transport_linux.zig changed from c. to sys.; transport_macos.zig keeps std.c (macOS always links libSystem).
• failFast keeps _exit semantics: std.c._exit with libc, else std.process.exit, which in Zig 0.16 is exit_group on Linux (RtlExitUserProcess on Windows) without atexit handlers or unwinding. With libc, std.process.exit would call exit(3) and run atexit handlers, hence _exit there.
• The library module, the demo executable and the embedding example no longer set link_libc. The demo's stall uses sys.nanosleep; the example's fatal path uses failFast.
• New Linux test runs without libc for transport.zig, server.zig and sys.zig (part of verify).
• README, USING.md and ARCHITECTURE.md document "No libc required on Linux".

Evidence (exact Zig 0.16.0)

Linux (omarx1, io_uring), under the measurement lock:

• zig build verify: 24/24 steps, 207/211 tests (4 skipped), incl. no-libc runs (transport 9, server 71 + 2 skipped, sys 1).
• Integration suites against the now libc-free demo server: integration 34, continuation 24 groups, batch 29, inline 10, gather 11, arena lifecycle 8.
• Embedding example (an independent consumer), ReleaseSafe, for the native target, x86_64-linux-musl and x86_64-linux-none: statically linked, no PT_INTERP, "There is no dynamic section in this file." The -linux-none binary serves HTTP/1.1 200 OK with a correct Date header (clock_gettime via direct syscall).

macOS (maxross, kqueue, native): zig build verify 18/18 steps, 124/128 tests (4 skipped); integration 30; continuation 24 groups; otool -L shows only libSystem.B.dylib.

Windows: zig build check -Dtarget=x86_64-windows-gnu and a ReleaseSafe cross build pass here; native Windows runs in CI.

🤖 Generated with Claude Code