v0.5.0

The first tagged release. Earlier versions here describe the tree as it was
developed, not as anything anyone could name and fetch.

Fixed

• The build passed ar rvD, whose D flag is GNU binutils only, so building
  any archive failed on macOS, and with the archiver's stderr discarded it
  failed with no message. Fixed upstream in modular-make 1.8.5, which probes
  for the flag, adds c so the archiver is silent unless something is wrong,
  and no longer discards its stderr. The vendored GNUmakefile moves from
  1.8.0 to 1.8.5.
• nc_crypto and keystore called getrandom(2) unconditionally, so
  neither compiled anywhere but Linux. Both now select a backend at compile
  time: arc4random_buf on macOS and the BSDs, getrandom(2) on Linux with
  a retry on EINTR, and /dev/urandom otherwise, which also covers a Linux
  kernel too old for the syscall. macOS is now built and tested in CI, which
  is how this was found.
• nc_random_id() fell back to an unseeded rand() whenever it could not
  read /dev/urandom, so every process produced the identical sequence of
  connection ids. That was the normal path on Windows, where the file does not
  exist, and reachable anywhere /dev is missing. It now asks the platform
  for real entropy, and the last-resort path mixes the clock with a stack
  address instead of calling rand() at all.
• Three tests read from a loopback socket without ever waiting and treated an
  empty queue as no traffic. Linux queues a loopback datagram inside
  sendto(), so the read always found it; macOS delivers through a separate
  context, so the reads came back empty and the tests gave up early. That
  failed the macOS CI job in nc_udp_test and test_gnet, and left
  test_microchan passing on a few microseconds of luck. The socket tests now
  wait for the datagram, and the protocol tests no longer use a socket at all.

Added

• A version, and a way to get one. The project is SemVer from 0.5.0 on, the
  version is in src/netchan.h as NETCHAN_VERSION_MAJOR, _MINOR, _PATCH,
  _STRING, and a comparable NETCHAN_VERSION integer, and releases are git
  tags of the form v0.5.0. A release workflow refuses to publish a tag that
  disagrees with the header.

• tools/vendor.sh, which copies a release into another project's tree. It
  takes the layers you name plus the ones they depend on, from the source
  snapshot GitHub serves for the tag, and records what it took. It runs
  standalone or piped:

  curl -fsSL https://raw.githubusercontent.com/OrangeTide/netchan/v0.5.0/tools/vendor.sh \
    | sh -s -- --version v0.5.0 --layer auth

• tests/netchan_test: the channel accessors, including a receiver
  identifying channels by the content type carried in the peer's OPEN, and
  the truncation seam where the 64-byte field meets the 63-byte wire cap.

• Windows support for the library. The core takes its monotonic clock from
  GetTickCount64 and its connection ids from rand_s; nc_crypto and
  keystore take entropy from BCryptGenRandom; nc_udp.h includes winsock2
  instead of <sys/socket.h>. crypto/ and auth/ need -lbcrypt, which
  the build now exports to their dependents. The core needs nothing on the
  link line, which is why it uses rand_s rather than the same call: a
  vendored src/ still builds with nothing but a compiler.

  The examples are unchanged and remain POSIX. They own sockets, an event
  loop, and a terminal.

• A CI job that cross-compiles for Windows with mingw-w64 and runs every
  socket-free test binary under wine, so the entropy and clock backends are
  executed rather than merely linked. Without it this would rot unnoticed,
  which is exactly what happened to macOS.

• netchan_chan_content_type(), the accessor for the string given to
  netchan_chan_open() or carried in the peer's OPEN. It tells apart channels
  that share a netchan type, such as two reliable streams named "control" and
  "events". Never NULL; a channel opened without one reads back as "".

• microchan/tests/mc_memlink, an in-memory datagram link for tests:
  delivery is a function call, and loss, duplication, and reordering happen on
  a fixed count. test_microchan gains a check that the Go-Back-N window
  holds its line through reordering and duplication, which loopback UDP could
  not produce on demand.

• microchan/tests/test_mc_udp, covering the host UDP transport on its own:
  the ephemeral bind, the address packing, the return convention of a
  non-blocking recv, and that the address a datagram arrives with is usable as
  a destination.

Changed

• test_microchan and test_gnet run over mc_memlink instead of loopback
  UDP. Neither the core nor the game net layer touches a socket, so the
  socket was testing nothing about them while making both tests
  platform-dependent. They are now deterministic everywhere, and test_gnet
  went from 1.2s to 0.3s.

• microchan/, a second library: the same idea with one allocation per
  connection and none after it, every buffer inside it fixed at compile time,
  Go-Back-N over an 8-message window instead of selective ack, and a core that
  fits in a 16-bit large-model DOS binary.
  Transports are IPX for MS-DOS and UDP for a host. It brings its own tests,
  the four-player thor game the variant exists for, and an Open Watcom
  makefile for the DOS target.

  It is a separate library, not a build option. Its API was renamed on import
  from nc_*/NC_*/struct netchan to mc_*/MC_*/struct microchan,
  because the two trees otherwise both define struct nc_addr with different
  layouts and both ship an nc_addr.h, nc_udp.c/h, and netchan.c/h. Same
  names, different meanings, one repository. Nothing links both.

Vendoring

curl -fsSL https://raw.githubusercontent.com/OrangeTide/netchan/v0.5.0/tools/vendor.sh | sh -s -- --version v0.5.0

The source snapshot below is what that fetches. See
VENDORING.md.