iw4L/docs/DEPLOY.md

Shipping a release

What make deploy does, what reaches the Windows folder with no human involved, and what does not. That folder is WINDOWS.md; your own master rather than ours is MASTER.md.

make setup-windows              # rustup target + cargo-xwin, once
make release prod|dev           # local directory, VPS untouched
make publish prod|dev           # upload a finished RELEASE=
make deploy prod|dev            # release + publish; PROFILE=play
make deploy prod PROFILE=release
make provision                  # users, Caddy, systemd, certificates
make logs prod|dev SINCE=2h

Each recipe is one cargo xtask command — windows, release, publish, provision, logs, certs — and make forwards only PROFILE and RELEASE; cargo xtask alone lists them. prod/dev is the publication channel, not the compilation profile; address and root are IW4L_DEPLOY_HOST / IW4L_DEPLOY_ROOT in .env. prod: udp/4433 and https://<host>:8443/prod/; dev: 4434 and /dev/. Git, protocol and SHA freeze into deployment.json at release, not at publish.

An ordinary publish compiles nothing, calls no zstd, rewrites nothing under /etc and leaves the master running if its SHA matched. The uncompressed iw4l.exe stays local, the archive is iw4l-<sha>.exe.zst, the old one on the VPS survives, and the manifest is swapped in by rename.

Changing master_protocol::PROTOCOL_VERSION also changes ALPN: an old client is refused at the handshake instead of silently decoding a drifted format. The master and the clients of such a release ship together, and publish restarts the master because its SHA changed.

make launcher windows builds the portable ZIP in dist/windows/; it is not a deploy, and a fix to the updater arrives only by copying iw4launcher.exe by hand. What the player's launcher does with IW4L_UPDATE_URL: WINDOWS.md.

Cutting a public release. Tag v0.1.0-demo.N, titled IW4L Technical Demo N, and mark it a pre-release. The notes carry four things and are not worth writing without them: the commit it was built from; the platforms the archive was actually run on; the demo scenario (which game, version and map, and what to do); the known limitations. A target that merely compiled is not a verified platform, and a Linux release waits for no unstarted Windows archive, nor the reverse.

Every archive carries LICENSE, NOTICE and the two font licences (release.rs::LEGAL_FILES) and no game data. Check the archive, not just git ls-files: make publish-check reads the tracked tree and cannot see what was staged into a ZIP. Published tags are never moved, and history is never recreated once anything has been pushed.