Two defects the Carbon rig found, neither visible on a development loopback.
**The wipe id was null for every real session.** `Init` runs before the save is
loaded, so `SaveRestore.SaveCreatedTime` is not yet meaningful there and the id
resolved at load time stayed null for the life of the process — every frame
shipped without the field R12 splits history on. It was invisible because a
hot-reloaded plugin reads an already-loaded world and gets the right answer every
time; it took a server that BOOTED with the plugin installed, which is every real
one, to show `wipeId=none` beside a save sitting on disk. Resolved again at
`OnServerInitialized`, and lazily while still unknown.
**Unload blocked the main thread for two seconds against an unreachable sidecar.**
Observed as `hook 'Unload' took longer than 100ms [2002ms]` next to `link thread
did not stop cleanly` — phase 1's stall arriving by a different road. The link
thread sits in a blocking `TcpClient.Connect`, which has no timeout of its own
and cannot be woken; a host that REFUSES answers instantly, which is every
loopback test, and a host that DROPS does not answer at all. The connect is now
bounded, and waits on a stop handle of its own so a reload abandons it at once.
It cannot share `Wake`, which also means 'the queue has something in it' and is
signalled by every hook.
After: the same reload logs no slow-hook warning and no stranded thread.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016wDDVXWMDz82WqE1i969r4
Fifteen hooks: presence, deaths, chat, gathering, raided structures, bans,
reports, login attempts and the wipe. Every frame now carries an envelope —
`type`, `serverId` and `wipeId` — built in one place so nothing can emit a frame
without one.
Three rules the code enforces structurally rather than by intention:
• A read-path hook never vetoes. Four of these are documented as "returning a
non-null value overrides default behavior", so every hook is declared `void`
and cannot answer. `CanUserLogin` is in the wave for what it observes.
• A hook that can fire more than once a second per player is a counter.
`OnDispenserGather` fires on every swing at a tree; it accumulates into a
per-player tally flushed once a minute as one `player.tally` frame, as a
delta rather than a running total.
• `wipeId` is derived here, from the save's creation time, because this is the
only component that can read it. PROTOCOL.md §8.2 reverses protocol 1 on
that point deliberately.
`server.hello` becomes a board rather than a greeting, and `players.online`
joins it; both are re-sent on connect and every 60 seconds, which is what makes
a restarted sidecar repopulate itself without asking.
`rg.hooks` reports which hooks have actually fired. Hooks bind by name and arity
through reflection on both frameworks, so a rename by Facepunch and a name
Carbon's catalogue omits present identically — as silence. This is the standing
answer to both, and it outranks either catalogue because it is a measurement.
The repository had no `.gitea/workflows/` at all. `scripts/checkPlugin.js` asks
the three questions a compiler here cannot: every hook is in `ExpectedHooks`, so
`rg.hooks` can see it; every hook is `void`, unless answering is a decision
written down in `ANSWERS_DELIBERATELY`; and `ProtocolVersion` agrees with
`overlay.toml`, which is what stops a bundle that will not compose. The void
rule is inverted on purpose — a list of *vetoable* hooks would have to be
maintained against a catalogue in another repository, and the first one somebody
forgot to add is the one that would pass. Its own suite breaks it seven ways,
including the failure that would make the other six meaningless: a method parser
that silently matches nothing.
Proven on a live Oxide server: compiled, loaded, the envelope correct, both ban
hooks firing, and the boards repopulating a sidecar whose database had been
deleted 0.3 seconds earlier.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016wDDVXWMDz82WqE1i969r4
