docs/modules/rust/PLAN.md

# `module-rust` — the plan

**Status:** phases 0 and 1 done, 2026-09-15. **Twenty-two decisions of record, no open questions.** Audited against the whole contract, not just the game-facing chapters
(§7); the event and engagement catalogues are §9 and §10; §11 is a second pass over `MODULE_API.md`
itself. **R19–R22 (2026-09-15) added a second modding framework, a Pterodactyl egg, moved the rigs off
the workstation, and put the sidecar's configuration in the egg** — see [`CARBON.md`](CARBON.md) for
the framework reference and §14 for the rig as built.

The [dry run](../rust-dryrun.md) designed this module on paper and deliberately did not build it.
This is the document that builds it. Where the two disagree, this one is later and wins — but the dry
run's four *findings* still stand, and one of them (identity) has moved from a footnote to the
critical path. §2 says why.

Nothing here is normative. [`MODULE_API.md`](../../website/MODULE_API.md) is the contract,
[`MODULE_SYSTEM.md`](../../website/MODULE_SYSTEM.md) the system, [`EVENTS.md`](../../website/EVENTS.md)
the event design of record, and the [Integration Kit][kit] is the teaching text this plan follows
chapter by chapter. This document is a *schedule and a set of decisions*, not a specification.

---

## 1. The shape, and what is already settled

Three new repositories, mirroring the three the platform already has for Ultima Online, plus the
optional fourth part that lives inside the module:

| Part | Repo | Mirrors | What it is |
|---|---|---|---|
| Website module | [`RunicGateway/Module-Rust`](https://gitea.whitlocktech.com/RunicGateway/Module-Rust) | `Module-uo` | Routes, schema fragment, prebuilt client chunk, nav |
| Sidecar | [`RunicGateway/Rust-Link`](https://gitea.whitlocktech.com/RunicGateway/Rust-Link) | `link` | Owns the game connection and the durable copy |
| Oxide bridge plugin | [`RunicGateway/Rust-Plugins`](https://gitea.whitlocktech.com/RunicGateway/Rust-Plugins) | `servuo-plugins` | C# inside the game, dials out, never blocks |
| *(event capability)* | — | — | Declarations inside `Module-Rust` ([kit][kit] ch. 5) |

All three were created empty on **2026-09-15**. The module is built from
[`integration-kit/template/`][kit], which CI holds against a pinned core — currently
`MODULE_API_VERSION` **1.10.0**.

**The repository name is not the module id.** `Module-uo` ships a module whose `id` is `uo`; this one
ships `rust`. §2.1 of the contract requires `id` to equal the directory core loads it from, which is
`modules/rust/`, and it is the prefix of every table and every mount — so the capitalisation in the
repository name reaches nothing inside the bundle.

Settled before this document and unchanged by it (dry run, org lead, 2026-08-19):

- **One server, one sidecar**, on that server's own host. A community with six servers runs six
  pairs; the module holds six clients and core never learns there is more than one.
- **The plugin dials out.** Rust's server is a binary, so the way in is Oxide's published hook API
  rather than source — and the shard-dials-out invariant survives that change of footing unchanged.
- **No RCON.** It was the original design and it was overruled.
- **`wipe_id` on every table that holds gameplay data.** It is the whole shape of the game in one
  column, and it is the first thing a UO-shaped mental model gets wrong.

## 2. Decisions of record

### R1 — identity is an in-game link code for v1

**Decided 2026-09-15 (org lead).** A player proves account ownership by typing a command in-game; the
plugin issues a one-time code; the website confirms it through the sidecar. Exactly the shape
`module-uo` uses.

This is the dry run's finding 1 answered for now rather than closed. There is still **no
`registerAuthProvider`** in the contract at 1.10.0 — verified against `MODULE_API.md` §2.4 on
2026-09-15 — so "Sign in with Steam", which every Rust community expects, is not reachable from a
module today. The link code is one screen worse and needs no core change, so it is what v1 ships.

**What changed since the dry run is the stakes, not the options.** The dry run rated this survivable
because the module only *read*: a site that renders a leaderboard does not need to know which account
owns a Steam ID. R2 makes the site the author of who may do what in the game, and R3 makes it the
thing that hands out loot. Both are grants against a Steam ID. **A weak identity link is now a
privilege-escalation path, not a missing convenience** — so the code must be single-use,
short-lived, rate-limited, and issued in-game to the player who will own it.

Adding `registerAuthProvider` properly stays the first candidate for a future `MODULE_API` bump. It
participates in session creation, which is the one part of core a module must never be able to
weaken, and it must inherit core's existing policy: **SSO is link-only, identities are never
auto-provisioned.** Specified deliberately, not bolted on. It is out of scope here.

### R2 — site-authored permissions are mirrored into Oxide's own permission store

**Decided 2026-09-15 (org lead).** The website is the author of record for groups and grants. The
bridge plugin applies them through Oxide's own API (`permission.GrantUserPermission` /
`RevokeUserPermission`), so **Oxide is an enforcement cache and the site is the thing that
remembers.**

The alternative — the plugin keeping a private table only our own features consult — was rejected
because it cannot reach any third-party plugin, and reaching them is the point: a grant authored on
the site has to gate Kits.

Three properties fall out, and they are the reason this shape is worth its cost:

- **Every third-party plugin honours site-authored grants with no adapter**, because they all already
  call `permission.UserHasPermission`.
- **A wipe stops being a data-loss event for permissions.** The game forgets; the site does not, and
  re-pushes the whole set on the next connect.
- **Hand edits are reported, not overwritten.** Somebody typing `oxide.grant` at the console is
  drift, and drift is surfaced to an operator — the same posture a lease's `restore()` takes when it
  finds a value a human has moved ([kit][kit] ch. 5).

**Phase 0 exercised all three against the real store and they hold — but it found four rules the
push path has to obey (§12.2).** The one that would have cost the most:
**`permission.GrantUserPermission` silently no-ops when the permission is not registered.** It
returns `void`, throws nothing and logs nothing; the grant simply does not happen. A permission is
registered by the plugin that declares it, so **every grant naming an unloaded, renamed or
uninstalled plugin's permission disappears without a trace** — and since R2's whole recovery story is
"the site re-pushes the full set on connect", a re-push into a server missing one plugin is a silent
partial. The push must check `PermissionExists` (or register the name itself) and report the
difference as drift rather than assuming a write landed.

**This is a direction the Integration Kit has no chapter for, and that is a finding.** Chapters 3 and
4 are the read path — data leaving the game. Chapter 5 is one-shot commands with a ledger and a
teardown. This is neither: it is *continuously reconciled state where the website is authoritative*,
and its nearest relative in the contract is the Team provider **inverted** — instead of core asking
the module what the game knows, the module tells the game what the site knows. The mechanism it
borrows is chapter 4's: **every board's current state has exactly one producer, and it runs on
connect**, pointed the other way. Whether this deserves a sixth chapter is a question for phase 19,
after it has been built once.

### R3 — the Kits reward action registers always and refuses with a reason

**Decided 2026-09-15 (org lead).** Kits is required for event rewards, but "required" means the
action always exists and fails honestly where it cannot work — `{ ok: false, retry: false, error:
'Kits plugin not installed' }` — rather than vanishing from the authoring form or refusing to boot
the module.

Two details from [kit][kit] ch. 5 that this depends on and are easy to get wrong:

- **The reason must be in `error`.** Core reads exactly `ok`, `retry` and `error` off a failure
  envelope; a message under any other name is dropped and the operator sees a bare
  `"<action id> refused"`.
- **`retry: false` has to be reachable.** Core's dispatcher enforces `budgetMs` and classifies a
  budget timeout as retry *unconditionally* — so if the sidecar client's timeout is longer than
  `budgetMs`, our own `retry: false` is unreachable code. Derive one constant from the other and
  assert the inequality in a test. The first module this project shipped had exactly that pairing.

### R4 — Rust reaches an operator through the existing installer, behind `--game`

**Decided 2026-09-15 (org lead).** Not a second binary and not a shared-core refactor: the shipped
[`installer`](../../installer/PLAN.md) grows a game dimension, `--game servuo|rust`, and keeps one
release stream, one `doctor`, one `update`, one `uninstall`.

The shape of the work is set by where the coupling already is. **The reusable half is already
game-agnostic** — `src/service.rs` and `src/main.rs` mention ServUO zero times, and `net.rs`,
`diff.rs`, `paths.rs` and `ui.rs` barely more. **The UO-specific half is concentrated in four files**
— `install.rs`, `doctor.rs`, `overlay.rs`, `tier.rs` — plus the bundle manifest, where
`OverlayComponent` and `ServUoCompat` name the game in the schema itself.

So the change is: make the bundle's game payload a **variant** rather than an overlay, and
`ServUoCompat` a per-game compat block. That is a schema change on the published **`bundles`**
branch, and it is the part to design before touching code.

**The Rust payload is much simpler than the UO one**, which is what makes this affordable: no source
tree to overlay, no `patches/`, no opt-in patch tier — a Rust install is a `.cs` file dropped into
`oxide/plugins/`, plus the sidecar and its service, which the installer already knows how to do. What
it gains instead is a **prerequisite check**: is Oxide installed, and is its build current enough.
That is `doctor`'s shape, not a new concept.

The protocol pairing check generalises unchanged. `servuo-plugins/overlay.toml` declares the protocol
the overlay speaks and the installer refuses to pair a disagreeing sidecar; `rust-plugins` needs the
same declaration under whatever name the variant gives it, and the refusal is the same refusal.

### R5 — Teams come from Rust's first-party clans; the uMod Clans plugin is the richer tier

**Decided 2026-09-15 (org lead).** Basic functionality is built on **Rust's own clan system**; the
third-party **uMod Clans plugin** is an optional layer for a much richer experience, and lands in
phase 17 rather than phase 9.

The distinction matters more than the names suggest, so it is worth stating precisely:

- **Rust's first-party clans** are seven hooks in [`HOOKS.md`](HOOKS.md) — `OnClanCreated`,
  `OnClanDisbanded`, `OnClanMemberAdded`, `OnClanMemberKicked`, `OnClanMemberLeft`, plus colour and
  logo — each handing you a `LocalClan`. **All seven are "no return behavior"**, which is exactly
  what a read-only bridge wants: there is nothing to abstain from, so [kit][kit] ch. 4's
  return-`null`-from-every-veto rule has no work to do here. They need no plugin, and the rig is
  already running them (`D:\rust\server\server1\clans.287.db`).
- **Rust's first-party *Teams* are a different system** — twelve hooks, mostly vetoable — and are the
  transient in-game squad, not the persistent organisation. They are **not** what core's Team
  provider should be fed. Clans is the right choice and this is why.
- **The uMod Clans plugin's API is not in our mirror.** [`HOOKS.md`](HOOKS.md) is the *game's* 477
  hooks; a plugin's own API is its own documentation. It is reached through `[PluginReference]` and
  is null when absent, making it an R3-shaped dependency that must refuse with a reason.

**Reading that plugin's source settled R5 far more firmly than the reasoning above did, and in a
direction worth stating plainly: for Teams, the plugin is *worse* than first-party, not richer.**
`Clans` v0.2.10 (k1lly0u, MIT, 2,692 lines) publishes **fifteen `[HookMethod]`s and every one of
them is a mutation** — `CreateClan`, `JoinClan`, `LeaveClan`, `KickPlayer`, `PromotePlayer`,
`DemotePlayer`, `DisbandClan`, and the eight alliance verbs. **There is no read API whatsoever**: no
`GetClan`, no `GetClanOf`, no `GetClanMembers`, no `GetAllClans`.

> **Corrected in phase 0 (§12.3).** This paragraph continued *"it raises exactly **three** hooks —
> `OnClanCreate`, `OnClanChat`, `OnAllianceChat` — none of which is a membership transition"*, and
> concluded that the plugin cannot answer core's provider questions at all. **That was a grep
> artefact and it is wrong.** `Clans` 0.2.10 raises **nine** hooks. The missing six are invisible to
> a search for `CallHook("OnClan…` because the name is a `const` at the call site:
>
> ```csharp
> const string HOOK_NAME = "OnClanMemberJoined";
> Interface.CallHook(HOOK_NAME, tag, ulong.Parse(joining), RustMemberList);
> ```
>
> They are `OnClanMemberJoined(tag, joining, members)`, `OnClanMemberGone(tag, leaving, members)`,
> `OnClanDisbanded(tag, members)`, `OnClanAllianceCreated`, `OnClanAllianceDissolved`, and
> `OnClanUpdate(tag)`. The first three **carry the full member list**, so the plugin *can* answer
> "who is in this clan" without any read API, and `OnClanUpdate` fires on promote and demote — the
> exact transitions first-party lacks.

**The decision does not change, but its reason does.** First-party remains the Team provider's source
because it is the system the *game* maintains and every server has it; the plugin is an optional
install that not every shard will run, and feeding a provider from something optional makes Teams
conditional on a mod. It is no longer true that the plugin *cannot* answer the questions — only that
it should not be the one asked. Feeding the Team provider from first-party clans is still
**permanent, not a first step**.

What the plugin genuinely adds is **alliances and clan/alliance chat** — richer in *features*, not in
roster data. That is what phase 17's adapter surfaces, and it sits beside the Team provider rather
than under it.

One route is deliberately not taken: the plugin persists to its own files under `oxide/data/`, and a
determined integration could read those directly. **That is reaching into another plugin's private
storage, not using an API** — it breaks without warning on any upstream refactor and is not a
contract anybody owes us. If phase 17 wants roster data from the plugin, the honest path is an
upstream pull request adding a read method, not a file reader.

### R6 — the required base set: Kits, Clans, PopupNotifications and ZoneManager

**Named 2026-09-15 (org lead), extended the same day by R17.** All four are MIT; three are k1lly0u's
and BetterChat's author differs only in the optional tier. All fetched at plan time:

| Plugin | Version | Released | Source |
|---|---|---|---|
| [Kits](https://umod.org/plugins/rust-kits) | 4.4.9 | 2026-06-04 | `https://umod.org/plugins/Kits.cs` |
| [Clans](https://umod.org/plugins/clans) | 0.2.10 | 2026-05-06 | `https://umod.org/plugins/Clans.cs` |
| [PopupNotifications](https://umod.org/plugins/popup-notifications) | 0.2.1 | 2026-08-09 | `https://umod.org/plugins/PopupNotifications.cs` |
| [ZoneManager](https://umod.org/plugins/zone-manager) (R17) | 3.1.14 | 2026-09-02 | `https://umod.org/plugins/ZoneManager.cs` |

Every one has a direct `.cs` download, so phase 0's install step is four `curl`s into
`oxide/plugins/` and no manual retrieval. **Pull them fresh rather than using the copies staged in
`Downloads`**, which are an older vintage.

**The four base plugins use three different conventions for their callable API**, which is worth
knowing before someone goes hunting for the wrong one: Kits uses `[HookMethod]`, BetterChat uses
`API_`-prefixed methods, and ZoneManager uses plain private methods resolved by name. All three are
reached the same way from our side — `[PluginReference]` plus `Call()` — but only the first is
greppable as a declared API.

`Clans` is listed by uMod as a **Universal** plugin rather than a Rust one — it is written against
Covalence, and its API takes `IPlayer` rather than `BasePlayer`. That is the portable half of the
uMod surface, and it is why the same plugin serves several games.

**Kits is the opposite of Clans and is a genuinely good dependency.** Twenty-three `[HookMethod]`s,
including the three things the event work actually needs:

- `GiveKit(BasePlayer player, string name)` — the reward action's call;
- `GetKitNames(List<string>)` / `GetAllKits()` — **the option source** for the authoring form, so an
  operator picks a kit from the live server instead of typing an identifier from memory;
- `GetKitInfo` / `KitDescription` / `KitImage` / `KitMax` / `KitCooldown` /
  `GetPlayerKitUses` / `GetPlayerKitCooldown` — form metadata and per-player eligibility.

It also **raises `OnKitRedeemed(BasePlayer player, string kitName)`**, which the bridge can listen on
to report a redemption as an ordinary event, whoever triggered it.

**Two traps in `GiveKit`, both found by reading it rather than by reasoning about it.**

> **R16 moved these off the critical path.** The reward action no longer calls `GiveKit` — it grants
> the kit's `RequiredPermission` and the player redeems it themselves. Both traps are kept here
> because the second one is *why* R16 is the better design, and because anything that ever does call
> `GiveKit` directly — an admin "give this player a kit now" button, say — walks straight into them.

**`GiveKit` returns `null` on a failure path, and `null` is Oxide's idiom for "no opinion".** The
first line is `if (!player) return null;`. Everywhere else in this ecosystem a null return means
*abstain*, so the reflex — treat null as fine — reports a reward as delivered when there was no
player to deliver it to. The success value is the literal `true`; a refusal is a **message string**,
which drops straight into chapter 5's `error` field. So the mapping is
`result is bool ok && ok` for success, a string for a refusal reason, and **`null` is a failure, not
a success.**

**`GiveKit` takes a `BasePlayer`, so the player must be connected.** There is no offline grant in
this API. An event that rewards participants at two in the morning rewards only whoever is online at
that moment, silently. The choice this forced was: accept online-only and say so in the action's
description, or keep a persisted pending-grant queue in the bridge and redeem it on next connect —
a second at-most-once store with its own idempotency, which is not free.

**R16 took a third option and it is the right one: stop granting items.** Grant the entitlement
instead. An entitlement waits without a queue, because waiting is what an entitlement does. The
problem was not hard to solve — it was the wrong problem, produced by an action declared around the
wrong noun.

`PopupNotifications` is small and does exactly one thing: `CreatePopupNotification(string message,
BasePlayer player = null, float duration = 0f)`, where a null player makes it global. That is the
server-side notification surface, and it needs no more than that.

**The one gap to design around:** the seven first-party hooks carry created, disbanded, added,
kicked and left — but **no promote or leader-changed event**. Core's provider requires
`getTeamLeaders`, so leadership is read off `LocalClan` at snapshot time rather than tracked from
transitions. That makes phase 9 *partly* snapshot-driven where the dry run predicted it would be
fully event-driven — a small correction to that document, recorded here rather than silently.

Phase 0 verified the seven against `agent/hooks.tsv` and the gap is real. It also found that the
**Clans plugin closes it** — `OnClanUpdate(tag)` fires on both promote and demote (§12.3). That does
not move the provider off first-party, but it does mean phase 17's adapter can offer *event-driven
leadership* on servers that run the plugin, over a snapshot baseline on servers that do not. Design
phase 9's snapshot so phase 17 can sharpen it rather than replace it.

### R7 — the notifications and engagement set ships in v1

**Decided 2026-09-15 (org lead).** `registerNotificationStreams`, `registerEventTriggers`,
`registerAudiences` and `registerEngagementSeeds` are all in the first release, plus
`registerAnnounceLeg` and `registerPostHook`.

They are a **matched set**, which is why they are one decision and one phase: a trigger declares the
payload contract and the widest audience a rule on it may ever be given, an audience resolves *people*
over module data, and seeds ship the bodies and the rules that use them. Three rules from the kit
that this phase lives or dies on:

- **`ceiling` is required, has no default, and is not a ladder.** A `staff` ceiling does **not** permit
  `owner`, because "one person" for a cheat-detection event is *the player it was detected on*. Fewer
  people is not less exposure.
- **`subjectKey` must name one of your declared variables** — it is what the cooldown keys on. Core
  refuses the module at boot if it names nothing, which is the good failure; the bad one it prevents
  is every subject sharing `undefined`.
- **A rule group is offered once, per group key.** A rule appended to an existing group reaches
  **fresh installs only** — so a rule that must reach existing deployments takes a new group key.

And the emit discipline: **emit on the transition, not on the poll.** Core's cooldown would hide a
module that emitted "the server is still up" as news.

### R8 — multi-server from the start

**Decided 2026-09-15 (org lead).** The server list is the landing page and everything else hangs
under `/rust/servers/:id`. Every gameplay row carries a server id as well as a `wipe_id`, and the
module holds one sidecar client per configured server.

This is the dry run's finding 2 taken at face value: *"one module, one game" is not the same as "one
module, one server"*. Retrofitting an `:id` segment through every route, table and page is the
expensive version, and a Rust community runs several servers by default.

### R9 — the live map, with every layer toggleable

**Decided 2026-09-15 (org lead).** The site offers a live map, and **each layer is an operator switch
set to public / players-only / admin-only** through the existing visibility framework
([`SHARD_VISIBILITY.md`](../../website/SHARD_VISIBILITY.md)).

**The map is two problems and they take different paths, which is the thing to get right before
building either:**

- **The map image is static content on the game host** — `proceduralmap.<size>.<seed>.<save>.map`
  beside the world save, regenerated only on a wipe. That is exactly [kit][kit] ch. 3 §2b's case, and
  it takes that shape: **request/reply, never events** (a sidecar that broadcast it would write
  megabytes into its own store and fan them at every client), **one in flight with an explicit busy**,
  **two stages — what exists, then what changed**, and a **derivation version** separate from the
  protocol so improving how we read the file invalidates a cached image whose source hash did not
  move. And **no import on boot**: a wipe is an event the operator knows about and the website does
  not.
- **Everything moving on it is live state** down the ordinary read path — monuments, cargo ship,
  patrol helicopter, airdrops, locked crates, and player positions.

**The layer switches are a security boundary, not a preference.** Public player positions in Rust are
a competitive-advantage leak — anyone, including people who do not play on the server, could locate
players and infer base positions. The default posture is monuments and world events public, player
and base layers admin-only, and an operator opening one up is a deliberate act with the consequence
stated on the switch.

**This is the only asset-bridge work in scope.** Item icons and the 2,590 workshop skin ids stay out
of v1; kill feeds and kit lists render as text.

### R10 — the Android app is in this workstream, capability-driven, trailing by one phase

**Decided 2026-09-15 (org lead).** Full Rust support in the app, not a degradation check. It decides
which screens to show from **`GET /api/v1/public/modules`** — each started module's `capabilities`
array — and each app leg lands **one phase after** the website surface it consumes is merged, so it
is always built against a real endpoint rather than a planned one.

Two endpoints that are *not* this and are easy to confuse with it, both checked on 2026-09-15:
`/api/v1/public/status` returns site mode plus a version block — the first-run probe and
version-mismatch guard, not a feature manifest — and `/api/health` on the internal app is a liveness
probe returning `{status:'ok'}`. Neither can say which pages exist.

§2.9's rule governs: **treat an unknown capability as absent, and never infer a URL from one.**

### R11 — a small read-only set of Discord slash commands

**Decided 2026-09-15 (org lead).** Questions answered from data the module already holds — server
status, wipe schedule, leaderboards, who is online. No write verbs, and the account link stays on the
two surfaces R1 names rather than acquiring a third.

The trap to carry in from the Teams work: **ephemerality is fixed at the deferral**, so a command
that might refuse must defer ephemeral or its refusal goes public in the channel. And the registries
have no removal path, which is an argument for adding a command late rather than early.

### R12 — per-wipe detail plus all-time rollups

**Decided 2026-09-15 (org lead).** Every gameplay row carries `wipe_id`; leaderboards default to the
current wipe; a separate rollup accumulates per player across wipes so a returning player's history
survives the monthly reset.

The truncation is a **runtime operation on a module route, never a schema one** — §2.6's leading-verb
allowlist forbids `DELETE` and `TRUNCATE` in a fragment precisely because the fragment replays at
every boot and would empty the table on each restart.

### R13 — two extension slots: `admin.users.detail` and `site.footer.status`

**Decided 2026-09-15 (org lead).** An operator looking at a user sees their linked Steam identity,
per-server stats and site-authored permission grants in core's own admin user page; the footer
carries a shard-status indicator on every page.

One module per slot, so claiming them also reserves them. And the naming rule matters for anyone
reading this later: **a slot is named for a PLACE, never for a meaning** — `site.footer.status` is
"the status-ish spot in the footer", not core knowing what a game server is.

### R14 — `/rust` on all three tiers

**Decided 2026-09-15 (org lead).** `public`, `admin` and `player` all mount `/rust`, exactly as
`module-uo` mounts `/uo`. Sub-surfaces are path segments: `/rust/servers/:id`, `/rust/map`,
`/rust/clans`.

Chosen deliberately because **prefixes share one namespace with core's own and the loader's collision
probe cannot see all of core's** — several core endpoints are mounted at the tier root rather than
under a prefix. A noun from our own domain that equals the module id cannot collide, where
`/servers` or `/map` very well might.

### R18 — plugin configuration is editable from the site, with a generated form and an auto-reload

**Decided 2026-09-15 (org lead).** An admin edits any loaded plugin's configuration from the website
and it reloads automatically. Two tiers:

- **Base:** the site walks `oxide/config/` **recursively** and **generates a form from the values
  themselves** — a boolean becomes a toggle, a number a numeric field, a string a text box, an array
  a list, a nested object a group. It is derived at read time, so **it works for whatever plugins
  happen to be installed**, including ones added or removed after we shipped.
- **Advanced:** raw JSON editing for anything the generated form cannot express.

This is the same posture as R2 — the site as the authority over the game host — but a different
*shape*. R2 is continuously reconciled state pushed on every connect; this is **request/reply,
on demand** ([kit][kit] ch. 3 §2a). Do not build it on the permission mirror.

**The mechanics Oxide gives us, all verified 2026-09-15:** configs live at
`oxide/config/<PluginName>.json`; `oxide.reload <name>` rereads one; and **`OnPluginLoaded` /
`OnPluginUnloaded` are real hooks** in the Server category, so *whether a reload actually succeeded
is observable* rather than assumed.

#### Discovery is a recursive walk, and it stops at `oxide/config/`

**Amended 2026-09-15 (org lead).** Configuration is **not** one flat `oxide/config/<Plugin>.json` per
plugin. Plugins nest — `oxide/config/<Mod>/whatever.json`, and deeper — and one plugin may own
several files. So discovery is a **recursive walk** of the config tree, and the UI groups by plugin
rather than assuming one file each.

Four things follow, and the first is a boundary rather than a detail.

**`oxide/data/` is not the settings surface, and must not be walked into.** `DataFileSystem` writes
to `oxide/data/`, and that is **live state, not configuration**. The base set makes the point by
itself: Kits keeps `Kits/kits_data.json` and `Kits/player_data.json`, ZoneManager keeps
`ZoneManager/zone_data.json`, Clans keeps `clan_data.json` (and a legacy `clans_data.json` beside it,
which is also a reminder that these names are not stable). Editing those from a web form edits
**players' kit cooldowns and the live zone definitions**, a running plugin overwrites the change on
its next save, and `oxide.reload` does not make most plugins safely re-read them. It is a different
problem with a different answer and it is deliberately out of scope. If a plugin's *settings* genuinely
live under `data/`, that is a per-plugin exception someone opts into knowingly, never something the
walk discovers on its own.

**The reload target cannot be inferred from the path.** `oxide/config/Foo/bar.json` may belong to
plugin `Foo`, or to something else entirely — the folder name is convention, not contract. So the
reload target is an **explicit field with the folder name as its default guess**, confirmable by the
admin. Infer it silently and the failure is the nastiest kind available here: we reload the wrong
plugin, observe `OnPluginLoaded` for *it*, and report success while the plugin that was actually
edited never re-read anything.

**A relative path from a web form is now a path-traversal surface.** Canonicalise the resolved path,
assert it is genuinely under the config root, reject absolute paths, and reject symlinks that resolve
outside. Before this amendment the feature addressed files by plugin name; now it addresses them by
path, and that is exactly the change that introduces the bug class.

**Bound the walk and the file.** A depth limit, a file-count limit and a per-file size cap — a
pathological tree must not be enumerated and a multi-megabyte JSON must not be loaded into a form.
And because one plugin can own several files, **the backup and rollback operate on the whole set a
save touches**, not one file at a time.

#### The trap that would silently corrupt every float

**JavaScript cannot tell `1` from `1.0`, and Oxide configs deserialize into typed C# classes.**
`JSON.parse('{"Rate":1.0}')` yields the number `1`, and `JSON.stringify` writes it back as `1` — so
a naive read-modify-write of a config file **silently rewrites every whole-numbered float as an
integer**, on fields nobody touched. Newtonsoft may coerce it, or may throw, and a throw at load
means the plugin does not come back.

So: **never parse the whole document, mutate, and re-serialize.** Edit the file *textually* — a
surgical replacement of the edited key's value — or use a parser that preserves number literals. The
fields at risk are exactly the ones a Rust server tunes: gather rates, multipliers, scales.

#### Five more limits of inferring a schema from values

Worth stating because the feature's whole promise is that it works without knowing the plugin:

- **Empty arrays and `null` carry no type.** Nothing can be inferred; render them advanced-only.
- **Enum-like strings are indistinguishable from free text.** There is no allowed-value set to read,
  so a string field is a text box and cannot be validated.
- **There are no descriptions, no minimums and no maximums.** The key name is the entire label —
  which is survivable because Oxide convention favours readable keys (BetterChat really does ship
  `"Maximal Titles"` and `"Reverse Title Order"`).
- **Nested objects and arrays of objects need recursion**, a depth limit, and a fall-back to raw JSON
  past it.
- **The file after a reload may not be what we wrote.** Oxide merges missing defaults and saves, so
  re-read after reloading rather than assuming our write is the current state.

#### Safety, and why it needs more than the usual

**A bad config does not fail the write — it fails the next load, and the plugin stays down.** Since
R6 and R17 make four plugins *required*, a broken ZoneManager config takes event participation with
it. So the write path is:

1. Read with a **version** (hash or mtime) and require it back on write — an operator editing on
   disk at the same time gets a conflict rather than a silent overwrite.
2. Validate the edited document parses.
3. **Back up the current file**, write, reload.
4. **Watch for `OnPluginLoaded` within a window.** If it does not arrive, **restore the backup and
   reload again, automatically**, and report the failure with whatever Oxide logged.

That rollback is the feature's real content. Without it this is a web form that can take the shard's
plugins down one typo at a time.

**Two more obligations.**

**Redact secrets.** Plugin configs routinely hold API keys and Discord webhooks. A config reader
hands those to anyone who can open the page. Mask values whose key matches the usual shapes — key,
token, secret, password, webhook — and treat them **write-only**, exactly as the platform already
treats the uo-link token. This is a real leak vector and the generated-form approach walks straight
into it.

**Gate it on its own site permission and audit every write** — who changed which key, from what to
what, and whether the reload succeeded. It is an admin writing to the game host's filesystem, which
is the most powerful thing the site can do to a server.

#### One distinction to keep

**Editing a plugin's config file is not a lease.** A lease (§9) borrows a *convar* for a while and
the game restores it on a deadline. This writes a *file* and is permanent until someone changes it
back. They look similar from a web form and they are not the same mechanism — an event should never
reach for this one.

### R17 — ZoneManager joins the required base set, because events need to know where people are

**Decided 2026-09-15 (org lead).** **[ZoneManager](https://umod.org/plugins/zone-manager)** (k1lly0u,
3.1.14 released 2026-09-02, MIT, Rust, ~218k downloads) is a fourth **required** plugin, not an
optional one. It is what makes an event able to answer *where a player is*.

Reading its source changed two things this plan had been vague about, and promoted one action out of
the optional tier.

**Its API is private methods reached through Oxide's reflection `Call()`** — no `[HookMethod]`, no
`API_` prefix, which is a third convention among the four base plugins and worth knowing before
someone goes looking for one that is not there.

| Surface | What it gives an event |
|---|---|
| `IsPlayerInZone(zoneId, player)` → `bool` | the direct question |
| `GetPlayerZoneIDs(player)` → `string[]` | every zone a player is in |
| **`GetPlayerZoneIDsNoAlloc(player, List<string>)`** | the allocation-free variant — **the one to use on any sweep** |
| `CreateOrUpdateZone(zoneId, args, position)` | make a zone |
| **`CreateOrUpdateTemporaryZone(..., Plugin owner)`** | make a zone *owned by our plugin* |
| **`EraseTemporaryZone(Plugin owner, zoneId)`** | remove one; owner-scoped **only against another plugin's zone**, not against an unowned one (§12.4) |
| `GetZoneIDs` / `GetZoneName` / `GetZoneLocation` / `CheckZoneID` | the catalogue, for an option source |

And nine hooks raised: `OnEnterZone` / `OnExitZone`, `OnEntityEnterZone` / `OnEntityExitZone`,
`OnZoneInitialize` / `OnZoneUpdated` / `OnZoneDestroyed` / `OnZoneErased`, plus `CanSpawnInZone`,
which is a veto our bridge abstains from like every other.

**Three things this settles.**

**Participation stops being the hard part.** [`EVENTS.md`](../../website/EVENTS.md) §H rates
participation *"the hard part"* for UO and *"substantially easier"* for Rust because hooks carry
attacker and victim. ZoneManager makes it exact rather than merely easier: `OnEnterZone` and
`OnExitZone` are **presence transitions delivered as events**, so the participation ledger is fed
from what happened rather than reconstructed from a sweep. An action's reply carries its
`participants` envelope member, and this is where that member gets its content.

**Advance conditions become expressible.** A phase that waits until *"ten players are at the
monument"* is a real gate rather than a wish — it reads zone membership. Without ZoneManager the
same condition is distance arithmetic against a point, recomputed on a timer, which is both more
expensive and less accurate.

**`rust.zone.open` moves from the optional tier into the base catalogue, and it can be
`reversible: 'ledger'` honestly.** This is the nicer half: `CreateOrUpdateTemporaryZone` takes a
**`Plugin owner`** and `EraseTemporaryZone` is **scoped to that owner**, so ZoneManager already has a
first-class notion of a zone belonging to the plugin that made it. And erasing a zone that is gone is
a success, which is what `revert` needs.

> **Narrowed in phase 0 (§12.4), and this one is a safety correction.** The scoping is real but it is
> *one-directional*: it stops us erasing a zone owned by **another plugin**, and does nothing at all
> for a zone owned by **nobody**.
>
> ```csharp
> // Only compare zone owner if the owner param is provided so users can remove temporary zones
> // without needing to unload the plugin that created them
> if (owner && zoneOwner && owner != zoneOwner)
>     return false;
> ```
>
> `zoneOwner` is null for every *permanent* zone — which is every zone an operator made by hand. So
> `EraseTemporaryZone(us, "<operator's zone>")` **deletes it and returns `true`**, indistinguishable
> from erasing our own. Observed live: a zone created with `CreateOrUpdateZone` and no owner was
> erased by an `EraseTemporaryZone` call from an unrelated plugin.
>
> So this is **less** of ch. 4's persisted ownership registry than the paragraph above claims. We
> still keep our own map from core's resource reference to the zone id, and that map is now
> **load-bearing rather than convenient**: phase 12 must refuse to erase any zone id it did not
> record creating. ZoneManager will not refuse on our behalf, and the `true` it returns is not
> evidence the zone was ours.

**One trap to design against, and it is chapter 4's rule meeting a chatty hook.** `OnEnterZone` and
`OnExitZone` fire on the game thread and a large zone with a busy server produces a great many of
them. The emit path already enqueues and returns, so the game cannot stall — but **the bridge should
subscribe selectively rather than forwarding every transition in every zone**. A zone no event cares
about should cost nothing on the wire. Decide the filter with the hooks in front of you at phase 12,
and measure it, because a sweep over `GetPlayerZoneIDsNoAlloc` is cheap and a flood of wire traffic
is not.

### R15 — an optional-integration tier, with BetterChat as the first member

**Decided 2026-09-15 (org lead).** Beyond the required base set (R6) the module carries a tier of
**optional** integrations: each detects its plugin through `[PluginReference]`, degrades cleanly to
absent, and adds something the site already knows how to compute.

The first named member is **[BetterChat](https://umod.org/plugins/better-chat)** (LaserHydra, 5.2.15,
MIT, Universal/Covalence, ~200k downloads) — *"manage chat groups, customize colors, and add
titles"*. The motivating case is **titles earned from leaderboards**: top of the wipe's kill board
gets a tag in chat.

**Its integration point is a pull, not a push, and that is why it is a good first member.**
`API_RegisterThirdPartyTitle(Plugin plugin, Func<IPlayer, string> titleGetter)` registers a callback,
and BetterChat invokes it per player when it renders a chat line. So a leaderboard title is a *pure
function of state we already hold* — nothing is written into BetterChat, nothing can go stale, and
there is no drift to reconcile. Contrast R2, which is a push and needs a whole reconcile story.

**The one trap, and it is chapter 4's rule applied to somebody else's callback:** that getter runs
**synchronously on the chat path**. It must be a cheap in-memory lookup — never a socket call, never
a database query, never anything that can block. A title that costs a round trip is a chat message
that costs a round trip.

Its other two API methods, `API_AddGroup(group)` and `API_SetGroupField(group, field, value)`, pair
naturally with R2: the site already authors permission groups, so a site-authored group can carry a
chat colour and tag. That is a push and would need the same drift posture R2 has; it is a phase-17
decision, not a given.

**The tier is open-ended by design.** Other integrations get added as they prove useful, and the bar
for each is the one this plan applies everywhere: it must fulfil the contract — declare honestly,
degrade to absent, and never make the module's own surfaces depend on something that may not be
installed.

### R16 — the reward action grants the RIGHT to redeem, not the items

**Decided 2026-09-15 (org lead).** An event reward does **not** call `GiveKit`. It grants the
**permission that gates a kit**, and the player redeems it themselves in game, whenever they next
log in.

**Kits already has exactly this model built in**, which is what makes it cheap: every kit carries a
`RequiredPermission`, `GiveKit`'s own path checks it before handing anything over, and the in-game
kit menu renders a kit the player lacks the permission for as locked rather than hiding it.
`GetKitInfo` returns that permission under `["permission"]`, so the module can read which kits are
gated and which are open to everyone.

This is a better design than the one it replaces, and it is worth being explicit about how much it
removes:

- **The offline-grant problem disappears entirely.** `GiveKit` needed a connected `BasePlayer`, so an
  event firing at two in the morning rewarded only whoever happened to be online. An entitlement
  waits. **This closes the open question §3 carried** — no pending-grant queue, no second
  at-most-once store, none of it.
- **It is the same machinery as R2, not a second mechanism.** A reward becomes a permission grant
  authored by the site and mirrored into Oxide — the thing phase 7 already builds. One permission
  authority, one drift story, one audit trail.
- **`reversible: 'ledger'` becomes honest**, where a direct grant could only ever be `'none'`. See
  *"What phase 13 must declare honestly"* below.
- **The player gets agency.** They redeem when they want it, where they want it, with Kits' own
  cooldown and use limits still applying — rather than having items appear in their inventory,
  possibly while they are somewhere it is a liability.

**One design note the option source has to carry.** A kit with an **empty** `RequiredPermission` is
open to everybody, so granting a permission for it rewards nobody with anything. The authoring form's
kit dropdown must surface which kits are permission-gated and refuse — or at minimum warn loudly —
on one that is not. That is a real refusal with a real reason, and exactly what R3's envelope is for.

### R19 — the plugin is framework-agnostic: Oxide **and** Carbon, from now rather than later

**Decided 2026-09-15 (org lead).** Modded Rust runs on two frameworks, not one, and `module-rust`
supports both from the phase it first reads anything — not as a port after phase 18. The bridge
plugin stays **one `.cs` file in the `Oxide.Plugins` namespace deriving from `RustPlugin`**, which is
also Carbon's own documented first example, with `#if CARBON` used only where the APIs genuinely
differ.

**This is affordable because the divergence is concentrated, not spread.** Carbon is not a fork of
Oxide; it is a separate loader shipping an Oxide compatibility layer, and at the level a plugin sees
the two are the same API. [`CARBON.md`](CARBON.md) is the reference — where it came from, what was
read, and the honest note that **none of it has yet run on a live Carbon server.**

Three existing decisions take an amendment, and no decision is reversed:

- **R18 — paths come from the framework, never from a literal.** Carbon's config directory is
  `carbon/configs` (plural) and its data directory `carbon/data`, *and every one of Carbon's
  directories is relocatable from the command line* (`-carbon.configdir`, `-carbon.datadir`,
  `-carbon.rootdir`, and nine more). So the recursive walk is rooted at
  `Interface.Oxide.ConfigDirectory` and the directory it refuses to walk is
  `Interface.Oxide.DataDirectory`. Carbon reimplements both accessors; a hardcoded `oxide/config/`
  is wrong on Carbon *and* on an Oxide server whose operator moved things. The reasoning behind R18
  is untouched — only how the two roots are obtained.
- **R2 — the store was never readable and now it is unreadable by construction.** Oxide persists
  permissions as JSON (`oxide/data/oxide.users.data`); Carbon persists them as **Protobuf or
  SQLite**, switchable at run time. R2 always planned to read the API, so nothing changes — but the
  file-reading shortcut is now permanently closed, which is worth saying once. §12.2's
  `PermissionExists` pre-check also survives intact: Carbon's `GrantUserPermission` returns `bool`
  where Oxide's returns `void`, so **the framework that would have told us whether the write landed
  is the one we cannot portably listen to.**
- **R4 — `doctor` asks *which* framework, not *whether Oxide*.** The payload drops into
  `PluginDirectory` either way. One thing gets weaker: **Carbon's releases are rolling tags**
  (`production_build`, `edge_build`), not an incrementing build number, so "current enough" is a
  claim `doctor` can make about Oxide and can only approximate about Carbon.

**What this decision explicitly refuses.** Carbon publishes 30 hooks Oxide does not, including an
`OnCarbon*` family mirroring its admin module's every moderation action — a tempting staff-audit
feed, and precisely the thing that would quietly make Carbon required. **No Carbon-only hook and no
Carbon-only convar enters a catalogue** unless it has an Oxide answer first, or is advertised
conditionally on the connected server's framework as a deliberate decision. Likewise Carbon's native
`Carbon.Plugins` / `CarbonPlugin` shape is not used: it is the single choice that would make the
source Carbon-only.

**The one thing to hold loosely.** Thirteen hook names our uMod mirror carries are absent from
Carbon's published catalogue ([`CARBON.md`](CARBON.md) §6). At least two look like renames rather
than holes, and none is in a phase today. The protection is the one §6 already requires for a
different reason — *the plugin logs which of its expected hooks have fired at least once* — which
answers this on either framework without trusting either catalogue.

### R20 — a Pterodactyl egg is a Rust-Link deliverable, with the sidecar inside the game container

**Decided 2026-09-15 (org lead).** Most Rust servers are rented, and most rented Rust servers run on
a Pterodactyl panel. So alongside the installer (R4) and the hand install, **a published Pterodactyl
egg is the third supported way the shard side reaches an operator** — and it has to work *in the same
manner* as the other two, not as a degraded variant.

The egg is derived from the community **"Rust Autowipe"** egg, taken as the known-good base, and it
keeps everything that egg already gets right: the steamcmd install script, the wipe-day
`REGEN_SERVER` / `REMOVE_FILES` mechanism, the Rust+ `APP_PORT`, and — the reason it is the right
base — a **`FRAMEWORK` variable already offering `vanilla | carbon | oxide`**. The operator picks the
framework at deploy time, which is R19's justification restated as a deployment fact: we do not get
to choose.

**The sidecar runs inside the game's own container, and that is the load-bearing part.** A
Pterodactyl server gets its own network namespace, so `127.0.0.1` inside it is genuinely private —
which means **D2 survives untouched**: the game link stays loopback and stays unauthenticated,
because loopback *is* the authentication. The startup command becomes a small wrapper that launches
`rust-link-sidecar` and then `RustDedicated`.

The alternative — a second Pterodactyl server running the sidecar — was rejected for exactly that
reason. Two containers have no shared loopback, so it would force a token and a routable bind onto
the game link. That is the case argued at D2 and overruled; it is not reopened here.

Four things the egg must get right, each of which is a way to get it wrong:

- **A second allocation for `[web].bind`.** The sidecar's HTTP/WS side is the half the website
  reaches, so it binds to the container's assigned address on an allocation the panel hands out —
  not to loopback. The token is what guards it, exactly as on a hand install.
- **The sidecar's database must never appear in `REMOVE_FILES`.** That variable is the wipe
  mechanism, and R12 keeps **all-time rollups across wipes**. A sidecar store swept on wipe day is
  the one failure that looks like success: the server comes back, the site repopulates, and every
  player's history is silently gone.
- **Stop means stop the game.** The egg's stop command is `quit`, addressed to RustDedicated. The
  wrapper has to let the sidecar go down with it rather than outliving it or holding the container
  open.
- **The plugin and the sidecar come from a release, never from a copy.** The install script fetches
  the pinned pair the same way the installer resolves a bundle — which makes the egg the third
  consumer of the protocol-pairing check, not an exception to it.

**It lands in phase 18, beside the installer**, because phase 18 is already "how the shard side
reaches an operator", and one story told twice is how two stories drift apart.

### R21 — both Rust rigs move to Pterodactyl, because one install cannot prove two frameworks

**Decided 2026-09-15 (org lead).** Oxide and Carbon **cannot coexist in one install** — Oxide ships
a patched `Assembly-CSharp.dll` and Carbon requires Facepunch's vanilla one. So R19 cannot be proven
on `D:\rust`, or on any single server, at all.

Both rigs move to the existing Pterodactyl panel at **192.168.0.12** (node `Main`): one server with
`FRAMEWORK=oxide`, one with `FRAMEWORK=carbon`, on the same egg. They are started and stopped as
needed rather than both left running.

This replaces `D:\rust` as the rig of record, and it buys more than parity:

- **The egg gets exercised by every phase**, not only by phase 18. R20's deliverable stops being a
  thing written once at the end against a panel nobody has used.
- **It ends the wipe-day maintenance that dominated §4.** `start.bat`'s steamcmd argument ordering,
  re-extracting Oxide after every `app_update`, checking `Assembly-CSharp.dll`'s byte size to tell a
  half-done Oxide install from a working one — all of that becomes the panel's job, through
  reinstall.
- **It is a Linux rig.** Every previous finding came from Windows and Mono; phase 1 spent real time
  on a Mono-specific NUL-padded `SocketException.Message`. Production Rust servers are Linux, so the
  rig moving there makes findings more representative, and makes any remaining Windows-only
  behaviour something we notice rather than depend on.

**A Carbon rig must be a clean install, not a converted one.** Carbon migrates an Oxide install on
first boot — it copies config, data, lang and permission files across. A Carbon rig made by
converting the Oxide rig would start out holding the Oxide rig's state, and would prove less than a
fresh one.

**The access, and the two kinds of key it takes.** `RunicGateway/pterodactyl_claude_api_token` holds
both, one per line: an **application** key (`ptla_…`), which creates and configures servers,
allocations and users and reads eggs but **cannot touch files, power or console**; and a **client**
key (`ptlc_…`), which is where Pterodactyl puts exactly those. An application key is rejected
outright by `/api/client/**` and cannot be widened — they are two credentials, not two scopes of
one. So the deployment loop is three tiers, matched to what each is for:

| What | How | Why that one |
|---|---|---|
| **A release artefact** — the pinned plugin + sidecar pair | The egg's own install script, re-run by a panel **reinstall** | It is the path we ship. Exercising it on the rig is acceptance testing for free |
| **Working-tree iteration** — an uncommitted `.cs` under test | A **client** API key (`ptlc_…`): `files/write`, then `command` to reload | The Pterodactyl analogue of `servuo-plugins/deploy.ps1`, and it carries the same caveat: **if something only works when the push script copies it, it does not ship** |
| **Bulk or binary** — sidecar builds, world files | SFTP on the node, port 2022 | Where the client API's per-file write is the wrong shape |

**Both keys exist and both were exercised on 2026-09-15** — the token file holds them as
`application:` and `user:` lines, and §14 records the rig they built together. The push script itself
lives in **`Rust-Plugins`**, mirroring where `deploy.ps1` lives for ServUO.

### R22 — the sidecar is configured from the egg's variables, not from a file the operator edits

**Decided 2026-09-15 (org lead).** What normally lives in `sidecar.toml` moves into the Rust egg's
variables, so an operator on Pterodactyl configures the sidecar in the panel alongside the game's own
settings rather than opening a file manager to edit TOML. One configuration surface, in the place
they are already looking.

**This is nearly free, because the sidecar already does it.** `rust-link`'s `config.rs` documents its
precedence as *environment overrides file overrides defaults* and already reads all five keys from
the environment: `RUSTLINK_GAME_BIND`, `RUSTLINK_SERVER_ID`, `RUSTLINK_WEB_BIND`,
`RUSTLINK_WEB_TOKEN`, `RUSTLINK_DB_PATH` (plus `RUSTLINK_CONFIG` for the file's own path).
Pterodactyl exposes every egg variable to the container as an environment variable, so the mapping is
one-to-one and **no second configuration mechanism is introduced** — the file stays canonical, the
environment overrides it, the egg sets the environment, and the installer (R4) keeps writing the file
exactly as it does now.

Which gives the two halves of the shard side two different config surfaces, deliberately:

| | Configured from | Mechanism |
|---|---|---|
| The **plugin** | the website, Admin → the R18 config editor | D3: it reads `oxide/config/RunicGateway.json`, so it is inside R18 for free |
| The **sidecar** | the panel, as egg variables | R22: `RUSTLINK_*` in the container environment |

That split is right rather than merely convenient. The plugin is configured by the thing it talks to;
the sidecar is configured by the thing that starts it, and on a panel the operator has no shell.

**Three things the variable set has to get right**, each of which is a way to hand somebody a footgun:

- **`RUSTLINK_GAME_BIND` is not operator-editable.** D2 makes loopback the authentication on the game
  link; a panel field that accepts `0.0.0.0:7799` is a web form that puts an unauthenticated command
  channel on the network. It is set by the egg and marked neither viewable nor editable — the same
  posture R18 takes toward the plugin's own `Host`/`Port`, for the same reason.
- **`RUSTLINK_WEB_BIND` is derived from an allocation, not typed.** It has to match the port the panel
  actually handed out, exactly as the egg already derives `QUERY_PORT` and `RCON_PORT`. A free-text
  bind is a bind that silently does not match the allocation, and the failure is the website never
  connecting with nothing in any log to say why.
- **`RUSTLINK_DB_PATH` must point somewhere `REMOVE_FILES` never sweeps.** Already named in R20 and
  restated here because this is the decision that makes the path an operator-visible field: the wipe
  list and the database path become two settings on the same screen, and they must not be able to
  agree.

**The token is the one place the ergonomics are not automatic.** Today the sidecar generates a token
when it finds none and persists it to its config file, which is what makes it secure out of the box;
`--print-config` is how an operator reads it back. A panel variable cannot be filled in by the
program that generates it, so the choices are: ship an empty default and let the sidecar generate and
persist as it does now, with the operator reading it out of the panel's file manager once; or make
the operator paste one in. The existing precedence already supports both — a set variable wins, an
empty one falls through to generation — so this is a default to choose when the egg is built, not a
mechanism to design. **Whichever is chosen, note that a Pterodactyl variable is visible to anyone
with panel access to that server and appears in the container environment**, which is a different
exposure from a `0600` file and should be stated in the operator guide rather than discovered.

**Lands in phase 18 with the rest of R20's egg.**

## 3. Open questions

**None.** Every question this section carried was closed on 2026-09-15, and so was the one open
*request*: the token file now holds both keys, and **both were exercised end to end on 2026-09-15**
(§14).

One correction belongs here rather than being quietly dropped, because the shape of the mistake is
the reusable part. This section briefly recorded that the client key "authenticates and then lists
zero servers", and built a diagnosis on top of it — including a claim that *includes are broken on
this panel*, because `/api/application/servers?include=user` returned an empty list where the same
route without the include had returned six.

**Both claims were wrong, and they were wrong the same way.** The servers were being deleted while
the probing was happening, so two reads minutes apart were reads of two different worlds. The empty
client list was correct. The empty include was correct. Nothing was broken.

The lesson is not "check twice"; it is that **a differential diagnosis across two API calls silently
assumes the state did not move between them**, and on a live panel somebody else is also holding the
controls. Once a server existed, every one of those calls answered correctly on the first try.

*Clans in the base set while the Team provider reads first-party* was confirmed as the intended
reading: complementary, not in conflict — the plugin is installed for alliances and clan chat, the
provider is fed from the first-party system that actually publishes membership transitions (R5).

*Offline reward grants* was dissolved rather than answered. **R16 changed the noun**: the reward
action grants an entitlement instead of items, and an entitlement does not need the player to be
online. The persisted pending-grant queue that question was weighing is not needed at all.

## 4. The test rig

**Two servers on the Pterodactyl panel at `192.168.0.12`, one per framework (R21).** They replace
`D:\rust`, which was the rig for phases 0 and 1 and whose findings are still recorded in §12 and §13.

| | Oxide rig | Carbon rig |
|---|---|---|
| Panel | node `Main` (192.168.0.12), nest 4 "Rust" | same |
| Egg | ours, derived from **"Rust Autowipe"** (panel egg id 18 is the unmodified base) | same egg |
| `FRAMEWORK` | `oxide` | `carbon` |
| Allocations | game, query, RCON, Rust+, **plus one for the sidecar's `[web].bind`** | same |

Started and stopped as needed rather than both left running; the other servers on the node are
shut down, which is what makes two ~20 GB Rust installs fit a 128 GB disk.

**The Carbon rig is a clean install, never a converted one.** Carbon migrates an Oxide install on
first boot — config, data, lang and permission files all come across — so a Carbon rig made by
switching `FRAMEWORK` on the Oxide rig would start out holding the Oxide rig's state and would prove
strictly less.

**What the panel changes about how work reaches a rig.** The token at
`RunicGateway/pterodactyl_claude_api_token` is an *application* key: it manages servers, allocations
and users, and it **cannot write a file, press a button or run a console command** — Pterodactyl puts
those on the client API. R21's table has the three tiers; the short version is *release artefacts
arrive by reinstall, iteration needs a client key, bulk goes over SFTP on port 2022.*

**Two facts about this panel that are easy to trip over:**

- `/api/client/**` returns **403 `AccessDeniedHttpException`** for the application key — a clear
  error, but only if you are expecting it. It is not a permissions grant that can be widened.
- The application API has **no egg-write endpoint at all** (`/api/application/eggs` is a 404; eggs
  are read through `/api/application/nests/{nest}/eggs`). Importing a new egg version is an admin-UI
  or `php artisan` operation, so the egg's release artefact is a JSON file a human imports — which is
  also exactly how an operator will consume it.

### What moving off the workstation retires

Everything below was true of `D:\rust` and is kept only because it explains findings in §12. **None
of it is maintenance any more** — the panel's reinstall does the same work correctly.

- **`start.bat` never updated anything.** steamcmd requires `+force_install_dir` **before** `+login`
  and the script had it after, so the flag was discarded, the update ran against steamcmd's own
  directory, and the job errored every single time (`Error! App '258550' state is 0x486`). That — not
  the path, which was the first and wrong diagnosis — is why the rig fell a wipe behind.
- **Re-extract Oxide after every `app_update`.** Oxide ships a *patched* `Assembly-CSharp.dll` and a
  Steam update restores Facepunch's, but the update does **not** remove `Oxide.Core.dll` and friends
  — so a half-done install still *looks* Oxided while loading no plugins and raising no hook. The
  tell was file size: on build 25230300, vanilla 9,758,544 bytes against Oxide 2.0.7716's 9,953,280.
- **`C:\oxide_files` is a 2025-04-23 Oxide and must not be copied anywhere** — its
  `Assembly-CSharp.dll` is 6,842,880 bytes, a hard downgrade over a live install.
- **A wipe keeps `server/server1/cfg/`**, which holds `users.cfg` and therefore the `ownerid` line.
  Delete the whole identity directory and you silently remove the operator's own ownership along with
  the map. This one still applies — it is the game's shape, not the host's, and it is why the egg's
  `REMOVE_FILES` list is worth reading carefully rather than trusting.

**Rust force-wipes on the first Thursday of the month, and both frameworks rebuild to match.** "Is
the rig current" stays a recurring question rather than a setup step; what changed is that the answer
is now a reinstall rather than a sequence of manual steps that can half-succeed.

## 5. The phases

**Twenty phases, roughly doubled from the first draft** — the contract audit in §7 is why, and the
honest reading is that the first list was a game-bridge plan with a website module bolted on, where
the kit treats the module as the bulk of the work.

0–4 produce a working read-only multi-server Rust site that an operator can actually install. 6–7b are
the permissions and remote-administration product. 9 is Teams. 10 is the notifications set, whose catalogue is **§10**. 12–13
are events, whose catalogue is **§9**. 14 is the map. 18 is how any of it reaches somebody who is
not us. The Android legs (5, 8, 11, 15) each trail the
website surface they consume by one phase, per R10.

**R19 and R21 do not add a phase — they change what "done" means for several.** Both rigs exist from
phase 3 onward, so from phase 3 a criterion is met when it is met **on both frameworks**, and a
finding that holds on only one is a finding either way. Phase 2's release artefacts and phase 18's
egg are the two places the second framework is visible in the deliverable rather than only in the
proving.

Each phase ends with its findings written down, as every workstream here does.

| # | Phase | Repos | Done when |
|---|---|---|---|
| 0 | **The rig.** ✅ **Done 2026-09-15 — as built and findings in §12.** Updated to the current wipe (the script was fixed *again*, properly), Oxide re-laid, base set installed, the grant path proven end to end and both zone transitions observed live with a player connected. **Both criteria met** | docs | A current server boots with all four loaded, `oxide.grant` demonstrably gates something, and a test zone reports who is standing in it |
| 1 | **Protocol 1, three skeletons, and every bundle seam at once.** ✅ **Done 2026-09-15 — as built and findings in §13.** Plugin, sidecar and module all exist and all three were exercised against the live rig; three org-lead decisions (§13.0), five defects only a running server found (§13.3), and a correction to §11.3 (§13.2). **Both criteria met** | all 3 + docs | One hello line travels game -> sidecar -> module; killing the sidecar does not stall the game; all five guards green on an untouched skeleton |
| 2 | **Packaging and release.** ✅ **Done 2026-09-16 — as built and findings in §15.** `release.yml` *and* the gate that was missing entirely (`pr-checks.yml`, including the frozen-manifest job); the include list with two readers; `v0.1.0` published and installed into a running core from its manifest URL. Three org-lead decisions (§15.0), and the first proof by a core that `/rust` collides with nothing (§15.2). **Criterion met** | Module-Rust + docs | An operator installs the empty module from Admin -> Modules and it reaches `started` |
| 3 | **The read path, on both frameworks.** ✅ **Built and largely proven 2026-09-16 — as built and findings in §16.** Protocol 2: fifteen hooks, an envelope every frame carries, boards, a cursor feed and bounded history; four org-lead decisions (§16.0), two defects only a booted server could find (§16.2), and CI for the two bridge repositories that had none. **The player half of the catalogue is written down as a walk to run rather than measured** — see §16.7. First hook wave from [`HOOKS.md`](HOOKS.md); events and snapshots distinct at the wire; `wipe_id` **and server id** on every row (R8); all-time rollups (R12); every board re-emitted on connect. **First phase to run against the Carbon rig (R19/R21)** — it turns [`CARBON.md`](CARBON.md) from a source-read hypothesis into tested fact, including whether the 13 unlisted hook names are renames or holes | all 3 + docs | A restarted sidecar is fully populated within one connection, a wipe does not erase a player's history, and **the same plugin file does all of that on Oxide and on Carbon** |
| 4 | **The first pages.** ✅ **Done 2026-09-16 — as built and findings in §17.** `/rust` is the list (D12), `/rust/servers/:id` is one server with four tabs (D13), everything selectable in the URL; visibility-gated polling (D14); the `site.footer.status` slot filled with a live count (D15). Four decisions (§17.0) and **four defects a browser walk found, two of them already shipped in phase 3** (§17.2) — an unreachable refresh that erased the server's description, and a "last reported" line reading the wrong timestamp. **Criterion met**, walked against a live rig | Module-Rust | The site renders the last thing each server said while every server is off |
| 5 | **Android leg A** (R10). ✅ **Done 2026-09-17 — as built and findings in §18.** The server list and one server with four tabs, gated on a NEW capability the module had to declare (D16 — its five named surfaces, and a client needs one that names the module); a poll that keeps its rows when it fails, which the app had no shape for (D17); the drawer badge as D15 translated (D19). Four decisions (§18.0) and **three defects an emulator walk found that 644 green tests did not** (§18.4). **Both halves of the criterion walked on one device against two cores** | Android-app + Module-Rust + docs | The app renders a Rust site it has never seen, and a UO site unchanged |
| 6 | **Identity** (R1), and the `admin.users.detail` slot (R13) | 3 + docs | A player links an account in-game; an operator sees the Steam id inside core's own user page |
| 7 | **Site-owned permissions** (R2). Groups and grants authored on the site; full set pushed on connect, deltas after; drift reported. The `PermissionExists` pre-check stays the mechanism on **both** frameworks (R19); Carbon's 14 permission hooks are tested here as a possible live drift signal, and suppressed against our own pushes if they fire | all 3 + docs | A grant made on the website gates a third-party plugin in-game, survives a wipe, and behaves the same against Oxide's JSON store and Carbon's Protobuf/SQLite one |
| 7b | **Mod configuration from the site** (R18). **Recursive** walk of `Interface.Oxide.ConfigDirectory` — never `DataDirectory`, and never either as a literal path (R19) — generated form from the live values, raw-JSON advanced tier, explicit reload target, versioned read/write, auto-reload watched on `OnPluginLoaded`, **automatic rollback** over the whole file set, path-traversal guards, secret redaction, its own permission and an audit trail | all 3 + docs | An admin flips a ZoneManager setting from the website and it takes effect; a deliberately broken config rolls itself back and says why; a nested `<Mod>/x.json` is found and reloads the right plugin |
| 8 | **Android leg B** (R10). Identity and permission surfaces | Android-app | A player links from the app |
| 9 | **Teams from first-party clans** (R5). Membership event-driven, leadership read off `LocalClan` at snapshot; **`declareModuleSlot` × 3** for core's `team.notify` / `team.activity` / `team.forum` | Module-Rust + 2 | The clan page is ours, core's contributions land in places we named, and every slot empty still reads correctly |
| 10 | **Notifications and engagement** (R7). Streams, triggers with `ceiling` and `subjectKey`, audiences, engagement seeds, announce leg, post hook — **the catalogue is §10**, including the in-game-popup question | Module-Rust + docs | The offline raid alert reaches the player whose base it was, and nobody else |
| 11 | **Android leg C** (R10). Inbox and notification preferences for Rust triggers | Android-app | A Rust notification arrives on a phone and can be switched off there |
| 12 | **Events: budgets, option sources and the leases** (§9). [kit][kit] ch. 5's own ordering — leases before actions — and every key verified live before it is advertised | Module-Rust + 2 | A leased value is observed changing in the running game and restored, per key; `rust.group.membership` expires without core asking |
| 13 | **Events: the actions** (§9, R3, R16). `rust.kit.entitle` first, then `rust.prefab.place` and `rust.announce`; `reversible: 'ledger'`; the kit option source flags kits with no permission gate, plus **`reconcile()` and the boot-id watch calling `ctx.events.reconcile()`** (§11.1) | all 3 + docs | A reward granted at 03:00 is waiting in the kit menu when the player next logs in, and a revert withdraws it; a wipe reconciles the ledger instead of stranding it |
| 14 | **The live map** (R9). The map image over the bridge — request/reply, two-stage, one in flight, its own derivation version, no import on boot — plus the live layers and a per-layer public/players/admin switch built on **our own** visibility layer (§11.2 — `shardVisibility` is `module-uo`'s, not core's) | all 3 + docs | The map renders for the current wipe, and a player layer is invisible until an operator deliberately opens it |
| 15 | **Android leg D** (R10). Map and events | Android-app | The map renders on a phone with the same layer gates |
| 16 | **Discord slash commands** (R11). A small read-only set, every refusal deferred ephemeral | Module-Rust + docs | A refusal does not go public in the channel |
| 17 | **Optional mod integrations** (R15). **BetterChat** first — leaderboard titles through `API_RegisterThirdPartyTitle`, a pull with no drift — then the uMod **Clans** adapter (alliances and clan chat, beside the provider rather than under it, R5), then others as they prove useful | Rust-Plugins + Module-Rust + docs | A server missing every optional mod still runs the module, Teams included |
| 18 | **The installer** (R4) **and the Pterodactyl egg** (R20) — the two halves of "how the shard side reaches an operator", built together so one story is not told twice. `--game servuo\|rust`, the bundle payload as a variant, a **framework** prerequisite check in `doctor` (which one, not whether Oxide — R19), the protocol pairing refusal carried over; the egg derived from "Rust Autowipe" with the sidecar inside the game container, a second allocation for `[web].bind`, **the sidecar configured from egg variables** (R22), the sidecar store held out of `REMOVE_FILES`, and its install script fetching the same pinned pair the installer resolves | installer + Rust-Link + docs | An operator sets a Rust server up with the released binary and nothing hand-copied; **and** a second operator imports the egg, deploys, and reaches the same place — on either framework |
| 19 | **Docs, kit feedback, cutover.** `docs/`; **`.profile`** (three repos were added); **`runicgateway.com`** (a second game is a headline change, and Pterodactyl is a hosting claim the site can now make); the Integration-kit question R2 raised; and whether the kit owes a reader anything about **supporting two mod frameworks at once** (R19) — a shape it has no chapter for either | docs + Integration-kit + .profile + runicgateway.com | `docs/` describes what shipped, the front door names the new repos, and R2's missing chapter is answered either way |

### Why the lease comes before the reward action

The stated priority is Kits rewards, and this plan still schedules a lease first (12 before 13). [kit][kit] ch. 5 is
explicit about it and the reasoning survives restating: a lease is *a value that already existed,
changed for a while, and put back*, so reading it first gives you the baseline for nothing. An action
makes something that did not exist. The lease proves the whole command path — correlation, the
idempotency key, `budgetMs` against the client timeout, the deadline the game enforces on its own —
before anything hands out loot. It is a cheaper place to find all four of chapter 5's invisible
failures.

Reversible on request.

### What phase 13 must declare honestly

> **Rewritten 2026-09-15 by R16.** This section previously argued that a kit reward could only be
> `reversible: 'none'`, because there is no honest way to un-grant loot a player has already spent.
> That was correct *about a direct grant* and R16 stopped doing direct grants. The reasoning is kept
> below in its corrected form because the shape of the mistake is the reusable part: **the action was
> declared around the wrong noun.** What the event makes is not loot; it is an entitlement.

**The reward action grants an entitlement, and an entitlement is reversible.** `revert` revokes the
permission, removing one that is not there is a success, and it is idempotent by construction — so
**`reversible: 'ledger'` is the honest declaration**, and core's ledger sweep does real work on every
terminal path.

One consequence to state rather than discover: **a player who redeemed before the revert keeps the
items.** That is correct and not a hole. The ledgered resource is the *grant*, and reverting it
withdraws the entitlement rather than the consumption — the same way cancelling a coupon does not
un-eat the meal. An operator reading the run console should see that distinction in the wording.

**`cost()` counts permission grants, and unlike a kit count it is exactly knowable before dispatch.**
That removes the whole class of problem chapter 5 §4 warns about: there is no "declare the maximum
because you cannot know until the answer comes back". One recipient is one grant. Core prices `cost`
before dispatch and never reconciles it, so being able to count precisely is worth more than it
sounds.

**And the idempotency key largely stops mattering here**, which is the tidiest part of R16. Chapter 5
§2 draws the line itself: *"A key is for a write whose repetition would be a second EFFECT —
creating, granting, announcing. A write that SETS a value to X is idempotent by its own nature."* A
permission grant is a set. Pass the key through anyway — it costs nothing and it is what the
contract expects — but the failure mode it exists to prevent, a socket hiccup producing a second set
of everything, no longer has a way to happen.

The monthly wipe is still the case `revert`'s tolerance rule was written for: if a wipe or a rebuilt
host clears Oxide's permission store, every ledgered grant is invalidated at once and **"gone, and
that is fine" is a success**. Note this is also where R2 pays for itself twice — the site re-pushes
its whole permission set on the next connect, so an entitlement an event granted comes *back* rather
than being quietly lost.

## 6. Risks worth naming now

- **Hooks bind by name and arity, by reflection, with no compile-time check.** A misspelled hook is
  never called, silently, with no warning at load — the single most common way a Rust plugin does
  nothing. The plugin must log which of its expected hooks have fired at least once, so a hook
  Facepunch renamed on a wipe is visible rather than mysterious. See [`README.md`](README.md) §2.
  **R19 gives that mechanism a second job:** it is also the only trustworthy answer to "does this
  hook exist on Carbon", since two published catalogues disagreeing is evidence about the catalogues
  and not about the frameworks.
- **The cheapest way to make Carbon required is to do it by accident.** Carbon's extra 30 hooks, its
  23 extra convars and its `bool`-returning permission API are each individually useful, individually
  small, and collectively a framework lock-in nobody decided on. R19's refusal is written down
  because it will be re-argued, once per convenience.
- **Two rigs is twice the state that can be quietly wrong.** A finding proven on the Oxide rig and
  assumed on the Carbon one is exactly the failure this project keeps finding in source-read claims.
  From phase 3, "done" means done on both, and a phase that could only check one says so.
- **A convar that applies cleanly and does nothing.** Most game config is read once at boot and
  cached; applying it later succeeds, reads back correctly, and changes nothing. Every lease key gets
  verified live — apply, observe in the running game, restore — before it is advertised. The UO
  module surveyed 156 config reads and found roughly eight that were live.
- **The wipe cadence is the schedule.** A monthly force wipe moves the hook list, rebuilds both
  frameworks, and invalidates every ledgered resource. Phases that end near one should expect to
  re-verify rather than assume. **Carbon's self-updating and rolling release tags mean the Carbon rig
  may move under us between two runs on the same day**, where an Oxide build number at least says so.
- **The old rig's RCON password was `letmein` in plaintext with `rcon.web 1`.** Acceptable on a
  loopback dev rig behind a home firewall, and it must never be the shape anything published copies —
  which now matters more, because the panel rigs are reachable on a LAN address and **the egg is a
  published artefact that people will copy defaults out of.**
- **The panel is the rig and the deliverable at once.** Convenient, and a way to prove the wrong
  thing: a rig hand-tuned through the panel UI stops testing the egg. Anything a rig needs belongs in
  the egg or in the push script, never only in a server's saved configuration.

## 7. Contract coverage audit

Added 2026-09-15 after re-reading the whole kit rather than only chapters 3-5. **The first draft of
§5 was built from the game-facing chapters and under-planned the website module by a wide margin** —
the template makes eight of the ten non-event registrations and the plan covered three. Every element
is listed with the phase it now lands in; the column worth reading is *"was it in the first draft"*,
because that is the shape of the mistake.

### The server handshake — ten registrations plus two hooks (ch. 2)

| Call | In the first draft | Phase |
|---|---|---|
| `registerRoutes` | yes | 4 |
| `registerTeamProvider` | yes | 9 |
| `registerEventBudgets` / `OptionSources` / `Leases` / `Actions` | yes | 12-13 |
| `onBoot` / `onShutdown` | implicit only | 1 |
| `registerExtension` | **no** | 4 (`site.footer.status`), 6 (`admin.users.detail`) |
| `registerNotificationStreams` | **no** | 10 |
| `registerEventTriggers` | **no** | 10 |
| `registerAudiences` | **no** | 10 |
| `registerEngagementSeeds` | **no** | 10 |
| `registerAnnounceLeg` | **no** | 10 |
| `registerPostHook` | **no** | 10 |
| `registerSlashCommands` | **no** | 16 |

Triggers, audiences and seeds are one phase because they are a **matched set** in
`template/server/index.js`, not three independent gaps — see R7.

### The bundle's own parts

| Part | In the first draft | Phase |
|---|---|---|
| `module.json` `id` / `coreApi` / `capabilities` | yes | 1, 4 |
| `mounts` and prefix choice | **no** | 1 (R14) |
| `schema.sql` | yes | 1, 3 |
| `purge.sql` | **no** | 1 |
| `swagger-fragment.json` + generator + `check:swagger` | **no** | 1 |
| `vite.config.js` aliases / shims / `checkExternals` | **no** | 1 |
| `checkImports.js` | **no** | 1 |
| `release.yml`, install manifest, `sha256`, host allowlist | **no** | **2** — there was no packaging phase at all |

Phase 1 carries most of these on purpose. [kit][kit] ch. 1's whole argument is to get every seam
working at once with almost nothing in them, so that afterwards you break exactly one at a time.

### The client half

| Part | In the first draft | Phase |
|---|---|---|
| `registry.registerRoutes` / `registerNav` | yes | 4 |
| `declareModuleSlot` | **no** | 9 — the kit: *"you will need it the moment your game has anything like a guild"* |
| `registerFeatureProvider` | **no** | 4 |
| UI kit discipline (`PublicLayout` `shell`, `PageHeader` props) | **no** | 4 |

### Beyond the module

| Item | In the first draft | Phase |
|---|---|---|
| Sidecar rpc correlation | implicit | 1, stated |
| Asset bridge (ch. 3 §2b) | **no** | 14 — **map image only** (R9) |
| `.profile` landing page | **no** | 19 |
| `runicgateway.com` | **no** | 19 |
| Android app | **no** | 5, 8, 11, 15 (R10) |
| `docs/` | yes | every phase, plus 19 |

**What the audit cost the schedule: twelve phases.** That is the honest number, and it is worth
recording because the under-planning had one cause — reading the chapters that describe the game
bridge and treating the module as the thin part, when the kit says in its first paragraph that the
website module is most of the work.

## 8. Questions the audit raised — all answered

Every question §7 produced was put to the org lead on 2026-09-15 and answered the same day. They are
recorded as R7–R14 in §2 rather than repeated here:

| Question | Answer | Decision |
|---|---|---|
| Notifications and engagement in v1? | the full matched set | R7 |
| How many servers does the UI support? | multi-server from the start | R8 |
| The asset bridge? | **only** the live map — not item icons, not skins | R9 |
| Android in this workstream? | full app, capability-driven, trailing by one phase | R10 |
| Discord slash commands? | a small read-only set | R11 |
| What survives a wipe? | per-wipe detail plus all-time rollups | R12 |
| Extension slots? | `admin.users.detail` **and** `site.footer.status` | R13 |
| Which mount prefixes? | `/rust` on all three tiers | R14 |

§3 is empty; both were closed on the same day.

## 9. The event catalogue

Added 2026-09-15. **Phases 12–13 described the event *mechanism* and never the *catalogue*** — one
budget and one lease as a proof of life, which is a skeleton rather than a product. This section is
what the module actually declares.

[`EVENTS.md`](../../website/EVENTS.md) **§H is a Rust/Oxide compatibility section that already
sketched this**, and it should have been read before §5 was written. What follows takes its ids and
its reasoning as the starting point rather than inventing a parallel set.

> **§H's thesis, and it is the one to design around:** *"The lease is the primitive that travels, not
> the spawn. Double gather rate for the weekend is the canonical Rust community event, and it is
> exactly lease-with-expiry. Spawning creatures at a landmark is UO-shaped; holding a value for four
> hours is every game."* It also rates Rust the **easier** case than UO, because Oxide's convars are
> live by default where ServUO's are mostly cached at boot.

### Budgets — what core counts and bounds

| Dimension | Counts |
|---|---|
| `rust.prefabs` | objects placed into the world by a run |
| `rust.zone.minutes` | zone time held — real since R17 |
| `rust.grants` | entitlements granted (R16) |
| `rust.announcements` | in-game broadcasts |

**Caps are per run, and R8 makes that load-bearing.** §H: `run.scope` is part of a run's unique key,
so one definition fanning out to six servers is **six separate budgets, not one shared pool**. An
operator setting a cap of 30 prefabs is setting it per server. Say so on the field.

### Option sources — what fills a dropdown

| Source | Filled from |
|---|---|
| `rust.options.kits` | Kits `GetKitNames` / `GetAllKits`, **flagged by whether `RequiredPermission` is set** (R16) |
| `rust.options.groups` | Oxide permission groups (§H names this one) |
| `rust.options.permissions` | registered permissions |
| `rust.options.prefabs` | a plugin-declared constructible allowlist — the analogue of UO's spawn atlas |
| `rust.options.monuments` | monument names, shared with the map work (R9) |
| `rust.options.zones` | ZoneManager `GetZoneIDs` / `GetZoneName` (R17) |

Every one resolves from live data and returns `[]` on failure rather than defending with a hardcoded
list that will be wrong. A source that refuses degrades its field to free text with a warning and
never blocks the form.

### Leases — values borrowed with a deadline

The heart of it, and the thing to build first.

| Lease | Value |
|---|---|
| `rust.rate.gather` | gather rate multiplier |
| `rust.rate.craft` | craft speed |
| `rust.rate.smelt` | smelting speed |
| `rust.rate.decay` | decay scale |
| `rust.time.night` | night length |
| `rust.population.<kind>` | spawn population multipliers |
| `rust.group.membership` | **a time-limited permission group — weekend VIP** |

**`rust.group.membership` is the one §H names that R16 did not, and the pair is the whole design.**
R16 settled that a *permanent* earned entitlement is an **action** with `reversible: 'ledger'` — grant
the kit's permission, revert revokes it. §H settles that a *time-limited* group is genuinely
**`core.lease`** — held with a deadline the game enforces on its own, restored when it expires
without core having to come back. Same underlying permission mirror (R2), two different shapes,
and choosing the wrong one is the mistake: a weekend VIP implemented as a grant is a VIP who stays
one for ever if the website goes away.

**Every key gets verified live before it is advertised** — apply, observe in the running game,
restore, per key. §H's claim that Rust convars are live by default is an argument for *expecting*
them to work, never a substitute for checking. A value the server reads once at boot applies
cleanly, reads back cleanly, and does nothing at all, and neither core nor review can catch it.

### Actions — verbs a run performs

| Action | `risk` | `reversible` | Notes |
|---|---|---|---|
| `rust.kit.entitle` | `change` | `ledger` | R16 — grants the kit's `RequiredPermission`; revert revokes |
| `rust.prefab.place` | `change` | `ledger` | §H's verb; revert kills the entity, and needs the persisted ownership registry ch. 4 describes |
| `rust.announce` | `notify` | `none` | via PopupNotifications (R6) — global or targeted |
| `rust.zone.open` | `change` | `ledger` | §H's other verb. **Base, not optional, since R17** — `CreateOrUpdateTemporaryZone` takes a `Plugin owner`, so the undo is real. Our own id map decides what may be erased, not ZoneManager's owner check (§12.4) |

**Rewards are not a contract member.** `EVENTS.md` deleted a `registerEventRewards` registry because
it carried four Ultima Online nouns inside a core signature. A reward here is an ordinary action —
which is exactly why R16 could change what it grants without touching anything of core's.

## 10. The engagement catalogue — what Rust can expose

Added 2026-09-15, answering "check the default alerts Rust can expose". R7 settled that the set
ships; this is what goes in it.

**The ceiling lattice is containment, not size** — `self`, `owner`, `subscribers`, `staff`,
`members`, `authenticated`, `everyone`, and the flat reading is the trap. `staff` is **not** a
superset of `owner`: for a cheat-detection event, "one person" is *the player it was detected on*.
Every ceiling below is chosen against that, not against a ladder.

| Trigger | Source | `ceiling` | `subjectKey` |
|---|---|---|---|
| `rust.wipe.started` | `OnNewSave` | `everyone` | server |
| `rust.server.online` / `.offline` | link state transition | `everyone` | server |
| `rust.leaderboard.topped` | our own rollup (R12) | `everyone` | server |
| `rust.base.destroyed` | `OnEntityDeath` on owned building blocks | **`owner`** | player |
| `rust.kit.entitled` | R16's own grant | **`self`** | user |
| `rust.player.linked` | R1's link flow | **`self`** | user |
| `rust.clan.member.added` / `.left` / `.kicked` | first-party clan hooks (R5) | `members` | clan |
| `rust.clan.disbanded` | `OnClanDisbanded` | `members` | clan |
| `rust.player.reported` | `OnPlayerReported` | **`staff`** | player |
| `rust.login.denied` | `CanUserLogin` | **`staff`** | player |
| `rust.player.banned` / `.unbanned` | `OnUserBanned` / `OnUserUnbanned` | **`staff`** | player |

**`rust.base.destroyed` is the one that matters most and the one most likely to be got wrong.** The
offline raid alert is the single most-wanted notification in Rust, and its ceiling is `owner` — the
player whose base it was. Ceilinged `staff` it would be useless to the person who needs it, and
ceilinged `everyone` it would broadcast base locations to the server. This is exactly the case the
lattice exists for.

**Three hooks carry data that must never widen.** `CanUserLogin` and `OnUserApproved` carry **IP
addresses**; `OnPlayerReported` carries player reports. [`README.md`](README.md) §5 already flags
these as admin-channel-only on the live feed, and the same judgement binds their triggers.

### Audiences

| Audience | Resolves to | `ceiling` |
|---|---|---|
| `rust.clan.members` | a clan's linked members | `members` |
| `rust.server.players` | linked accounts seen on a server this wipe | `authenticated` |
| `rust.wipe.participants` | everyone who played the current wipe | `authenticated` |

A resolver returns **user ids and nothing else** — never a template, a channel or an address — and
one that fails resolves to **nobody**, never to everybody and never to its last good answer. Its
params are constant, filled in when an operator saves the rule, so "the clan this event was about"
is not expressible; an event that needs that carries its own recipients.

### Seeds, and one thing to decide when building them

Bodies re-ensure every boot under a seed version; **rule groups are offered once per group key**, so
a rule appended to an existing group reaches fresh installs only. Wipe announcements, raid alerts and
clan transitions each take their own group key for that reason.

**One design note, flagged rather than decided.** `PopupNotifications` gives the module an *in-game*
alert surface, which is not one of core's channels — core resolves ids to email, in-app and push. So
an in-game popup is the module publishing to its own surface off its own trigger, not a fourth
channel core learns about. Worth settling deliberately at phase 10: a raid alert that reaches a
player's phone *and* pops on their screen next login is two mechanisms, and only one of them is
core's.

## 11. Second contract pass — `MODULE_API.md` read member by member

Added 2026-09-15, after docs#249 merged. §7 audited the plan against the **Integration Kit** and the
**template**; this pass reads [`MODULE_API.md`](../../website/MODULE_API.md) itself, enumerating every
member rather than grepping for registration names. It found one regression, one mispriced decision,
one missing declaration and a set of `ctx` members the plan had never mentioned.

### 11.1 The regression: `reconcile` was dropped

**`ctx.events.reconcile()` and an action's `reconcile()` appear nowhere in this document.** The
twelve-phase first draft had them — *"phase 6: reconcile, and the boot-id watch"* — and the rewrite to
twenty phases lost them. That is a regression in the plan, not a decision.

It matters **more** for Rust than for the game the contract was written against. [kit][kit] ch. 5
rates `reconcile` the one omission that is *"merely a lower standard rather than a broken promise"* —
but that judgement assumes a world that persists. **Rust wipes monthly, and a wipe invalidates every
ledgered resource for that server at once.** Core cannot tell a wedged sidecar from a game that
rebooted and lost everything an event made: it sees `{ ok: false, retry: true }` either way. It asks
once, at its own boot, and otherwise **waits to be told**.

`ctx.events.reconcile()` is being told, and the thing that triggers it is a **watch on the game's boot
id changing** — which is also the only way to tell a game restart from a sidecar reconnect. They are
not the same event and the second loses nothing. Two rules ride with it: anything that is not an
explicit `{ ok: true, inForce: [...] }` **leaves the ledger alone** — "I do not know" is never read as
"it is gone" — and a resource reported missing becomes `orphaned`, not `reverted`, because nobody
asked for it to go.

**Restored to phase 13**, after the actions exist, with the boot-id watch as its trigger.

### 11.2 R9 was mispriced: the visibility framework is `module-uo`'s, not core's

R9 says the map's per-layer switches work *"through the existing visibility framework
([`SHARD_VISIBILITY.md`](../../website/SHARD_VISIBILITY.md))"*, which reads as reuse. **It is not
reuse.** §6.3 records that `shardVisibility` is **module-owned**, and the tree confirms it — the util,
both models, the admin controller and its tests all live under `module-uo/server/`, and there is
**nothing** by that name left in `website/server`.

§2.7 forbids a module requiring anything outside its own directory, so `module-rust` cannot import a
line of it. **It builds its own**, informed by UO's design and its document but sharing no code.

That is a real cost R9 did not price. It is not a reason to change the decision — per-layer switches
are still right, and `SHARD_VISIBILITY.md` is still the design to learn from — but phase 14 carries
a visibility layer of its own rather than a configuration of somebody else's.

### 11.3 `extensions` is a declared field, not just a call

R13 claims two slots and never says where they are declared. **`module.json` has an `extensions`
array** (§2.1, optional), and the dry run's own manifest carried `"extensions": ["admin.users.detail"]`.
Like `mounts`, it is a statement of surface that the loader holds against reality — so
`admin.users.detail` and `site.footer.status` are declared there as well as registered. Phase 1 adds
it to the list of `module.json` fields that must be got right.

> **Corrected in phase 1 (§13.2), and it is half wrong.** The loader checks only that a named slot EXISTS (`loader.js:681` → `registries.hasSlot`); `checkDeclared` covers `mounts` **alone**, so a declaration with nothing behind it loads cleanly and means nothing. And only ONE of R13's two slots can be declared here at all — `admin.users.detail` is the only server slot core declares, while `site.footer.status` is a CLIENT slot registered from the chunk, and naming it in `extensions` fails the load outright.

### 11.4 The `ctx` members the plan had never named

`ctx` has **29 members** (§2.3). The plan named a handful. The ones that change work:

| Member | Where it lands | Why it matters |
|---|---|---|
| **`ctx.secretBox`** | 1 | Each configured server's sidecar token is a secret at rest. Core encrypts its own (AES-256-GCM, write-only in the API, never returned to any client) and hands a module the same facility — so R8's several tokens get the platform's existing posture rather than a new one |
| **`ctx.middleware.rateLimit`** | 6 | R1 requires the link code be rate-limited. This is the mechanism; `accountChangeLimiter` sits beside it for the account-facing half |
| **`ctx.uploads`** | 14 | Where R9's map image actually lands. The plan described fetching it over the bridge and never said where it goes |
| **`ctx.activity.log`** | 7, 7b | R2's permission changes and R18's config writes both owe an audit trail. Core has an activity log; neither needed inventing one |
| **`ctx.teams.publish`**, **`ctx.teams.activity.push`**, **`ctx.teams.reconcile`** | 9 | Teams is more than the provider. The plan named only `registerTeamProvider`, which answers core's questions — these are how a module *pushes* a change and asks for reconciliation |
| **`ctx.posts`** | 10 | The CMS surface behind `registerAnnounceLeg` and `registerPostHook` |
| `ctx.events.emit`, `ctx.inbox.push`, `ctx.push.publish` | 10 | The three send paths §10's catalogue implies and never named |
| `ctx.users.getById`, `ctx.settings.*`, `ctx.validator`, `ctx.db.query`, `ctx.paths.moduleRoot`, `ctx.log`, `ctx.express`, `ctx.auth.getUserFromRequest`, `ctx.site.baseUrl`, `ctx.moduleId` | throughout | Ordinary plumbing; listed so the narrowing is visible |

`ctx` is **a curated list, not core's internals** — `ctx.auth` is one function rather than core's whole
auth facade, because minting a session is core's job and a module needs to *read* one. Expect to want
something that is not there; that is a minor-version conversation, never a reason to reach around it.

### 11.5 §6.8 — a trigger, a rule and an audience outlive the module that declared them

A constraint on phase 10 and on purge that the plan did not carry.

`engagement_rules.trigger_id` is a plain `VARCHAR` — **no foreign key, no cascade** — deliberately, so
a module can be removed and reinstalled without destroying an operator's rules. The consequence:

- **A rule whose trigger is unregistered shows `dormant`** — never an error, never auto-deleted.
- **The same for an unregistered audience**: it resolves to the empty set and shows dormant, which is
  **not the same answer as "resolved to nobody"** and must not be rendered as if it were.

The failure that prevents is exact: **an id that stops resolving must never silently become a send to
a different set of people.**

### 11.6 The two client lists, in full

Recorded because §7 said "UI kit discipline" without saying what is in it. Both are **closed and
curated** — adding a member is a minor version bump, changing an existing prop is a major one.

**`registry`** — `registerRoutes`, `registerNav`, `registerExtension`, `registerFeatureProvider`,
`declareModuleSlot` (1.6.0), plus the read side, `routesFor` and `featureProviders`.

**`ui`** — `PublicLayout`, `PageHeader`, `Loading`, `ErrorState`, `EmptyState`, `useAsync`, `useAuth`,
`useSite`, `Slot` (1.6.0). Anything else — tables, chips, tabs, editors — **your chunk carries it**.

### 11.7 One confirmation for R10

§2.9: `GET /api/v1/public/modules` returns **only `started` modules**, with four fields and no state,
no failure stage and no failure reason. A `disabled` or `startup_failed` module is simply **absent**.

So R10's capability probe already has the behaviour the app wants: a Rust module that failed to boot
makes the app render a site *without* those screens, rather than one advertising screens that `503`.
The app needs no failure handling for this case because core does not expose the failure.

## 12. Phase 0 as built — the rig, 2026-09-15

The rig is current, the base set runs, and **both acceptance criteria are met**. Six things were
learned that the plan had either wrong or had never asked, and four of them change work in later
phases.

### 12.0 What the rig is now

| | Before | After |
|---|---|---|
| Server build | `24613624` (2026-08-13) | **`25230300`** (2026-09-10) |
| Oxide | `2.0.7585` | **`2.0.7716`** (`OxideMod/Oxide.Rust`, 2026-09-11) |
| World | seed 1234 save v287, previous wipe | regenerated for this wipe; `cfg/` preserved |
| `oxide/plugins/` | empty | Kits 4.4.9 · Clans 0.2.10 · Popup Notifications 0.2.1 · Zone Manager 3.1.14 |

All four compiled and loaded first time on the new build, at exactly the versions R6 and R17 name —
pulled fresh from `https://umod.org/plugins/<Name>.cs`, which still serves those versions and needs
no Cloudflare workaround. Server protocol `2633.288.1`.

The Oxide permission store survived the update untouched (`oxide/` is not a Steam depot directory),
so `76561198038695917` is still in `default` and `admin`.

Two instruments were built and are kept in the phase-0 scratchpad rather than committed: a
dependency-free **WebSocket RCON driver** (Node's global `WebSocket`, no `ws` package), and
**`RGProbe.cs`**, a throwaway Oxide plugin that exposes Oxide's permission API and ZoneManager's
by-name API as console commands. The probe is what made §12.2 and §12.4 observable; phase 1's plugin
skeleton can start from it.

> **One thing the RCON driver had to learn.** Oxide tags its own `Puts()` output and its warnings
> with the **identifier of the command being run**, so a first-match-wins client reads a plugin's log
> line as if it were the reply and discards the real one. It cost two wrong readings before it was
> spotted. Collect every frame in a window; do not correlate one reply per identifier.

### 12.1 The rig's own script was broken in a way the earlier diagnosis missed

Recorded in §4. In short: `start.bat` put `+force_install_dir` **after** `+login`, steamcmd discarded
it, and every update run in the rig's history errored out without updating anything. The 2026-09-15
"fix" changed the path and left the order, so it fixed nothing.

The Oxide re-install in §4 is **not** a finding — pairing a server update with an Oxide re-install is
the routine every Rust host already follows, and saying otherwise would be this plan talking down to
its own audience. One narrow consequence is still worth carrying to **phase 18**: because
`app_update` leaves `Oxide.Core.dll` and the rest in place, a `doctor` check that tests for `oxide/`
or for Oxide's assemblies **passes on a server that is mid-routine**. Compare the
`Assembly-CSharp.dll` against the Oxide build instead, so `doctor` reports the real state rather than
a directory listing.

### 12.2 Four rules the R2 permission push must obey

Verified live against the real store, granting and revoking through both the console command and the
API:

1. **`permission.GrantUserPermission` silently no-ops for an unregistered permission.** `void`, no
   throw, no log. The console `oxide.grant` at least answers `Permission 'x' doesn't exist`; the API
   path R2 uses says nothing at all. This is the finding with teeth — see R2.
2. **A permission exists only because a loaded plugin registered it.** Kits registers `kits.admin`
   and, dynamically, **every kit's `RequiredPermission`** (`Kits.cs:1225`, `:2895`) — which is what
   makes R16's entitlement model real. Unload Kits and those names stop existing.
3. **`RegisterPermission` warns about a foreign prefix but registers anyway.**
   `Missing plugin name prefix 'rgprobe' for permission 'someplugin.vip'` is a warning, not a
   refusal — the permission was created and granted successfully. So the site *can* make a grant
   stick for a plugin that is not currently loaded, at the cost of a console warning. Whether it
   *should* is a phase 7 decision; the mechanism exists.
4. **A player who has never connected is in no group, but can hold direct grants.** A grant to an
   unseen SteamID64 works and reads back immediately. Group membership does not exist for them yet,
   so **anything the site expresses as group membership does not reach a player until their first
   connection**, while a direct grant does. R16's offline entitlement is safe; a group-shaped
   entitlement is not.

Point 4 is the one to carry into phase 7's design: grants and groups have **different reach** for
offline players, and the site's model currently treats them as two spellings of the same thing.

### 12.3 R5's claim about the Clans plugin was a grep artefact

Corrected in R5. The plugin raises nine hooks, not three, and three of them carry full member lists;
the six that were missed are invisible to a literal search because the hook name is a `const` at the
call site. The decision stands on a different reason — first-party is what every server has, the
plugin is optional — and phase 17 gains event-driven leadership as a sharpening rather than a
replacement.

Two smaller things from the same read, both worth having before phase 9 and 17:

- **`Clans` raises the same hook name twice per transition**, once Rust-typed
  (`string, ulong, List<ulong>`) and once Universal-typed (`string, string, List<string>`), plus two
  deprecated arities. Oxide binds by name **and** arity, and both live forms are arity 3 — so a
  loosely typed subscriber catches both and double-counts every join and leave. Type the parameters
  precisely and pick one.
- **`Clans` calls `API_RegisterThirdPartyTitle` itself.** R15's BetterChat integration will be the
  *second* title provider on any server running both, not the first.

Also confirmed, since the plan rests on it: the first-party set is exactly the **seven** hooks in
`agent/hooks.tsv`, all "no return behavior", with no promote and no leader-changed.

### 12.4 ZoneManager's owner scoping is narrower than R17 assumed

Corrected in R17. `EraseTemporaryZone(owner, id)` refuses only when the zone has a *different*
owner; an **unowned** zone — every permanent zone, including every zone an operator made by hand — is
erased by anyone and returns `true`. Phase 12 must gate erasure on its own id map.

Three more things the source and the live rig agreed on:

- **ZoneManager's entire API is plain private methods**, no `[HookMethod]` anywhere in 3.1.14 — so
  `Call()` by name is the only way in, and a typo is silence. Confirmed working live for
  `CreateOrUpdateZone`, `CreateOrUpdateTemporaryZone`, `EraseTemporaryZone`, `GetZoneIDs` and
  `GetPlayersInZone`. The three-conventions finding holds: Kits declares `[HookMethod]` (23 of them),
  ZoneManager declares nothing, BetterChat will use `API_` prefixes.
- **`GetPlayersInZone` cannot distinguish an unknown zone from an empty one** — both return an empty
  list, not null. The participation ledger R17 wants to feed therefore cannot use this call alone to
  answer "is this zone still there", and must check `GetZoneIDs` separately. This is the same
  absence-of-an-answer / answer-of-absence trap earlier phases of other workstreams hit.
- **NPCs never appear in a zone's player list.** `baseEntity is BasePlayer { IsNpc: false }` routes
  them to the zone's *entity* list instead. Useful to know before designing a condition that counts
  "players at the monument" on a server with scientists.

### 12.5 The criterion is closed, and it revealed two ceilings on the rig

> `oxide.grant` demonstrably gates something, and a test zone reports who is standing in it

**Both halves done**, the second with the org lead connected. The zone was created on the player's
own position, and ZoneManager reported both transitions live:

```
[probe] ENTER zone=rgtest player=76561198038695917 (whitlocktech)
[probe] zone=rgtest occupancy=1 [76561198038695917:whitlocktech]
[probe] EXIT  zone=rgtest player=76561198038695917 (whitlocktech)
```

The exit was produced by **moving the zone off the player** rather than walking them out —
`CreateOrUpdateZone` on an existing id relocates the trigger volume and fires `OnExitZone` as it
leaves. Useful for testing presence without choreographing a person.

So **R17's "presence transitions as events" is verified**, which is the claim the participation
ledger and the advance conditions both rest on.

Two ceilings surfaced on the way, and both constrain later phases:

**1. No console session can observe a gate.** The standard idiom is
`return !player || permission.UserHasPermission(...)` — a command from RCON has no `BasePlayer`, so
the console is unconditionally allowed. Anything whose acceptance needs a permission to actually
*refuse* somebody needs a client attached.

**2. An admin account cannot see a refusal either — from most plugins.** The bypass is not uniform,
and the difference decides which phases can be demonstrated on the org lead's own account:

| Plugin | Admin bypass | Demonstrable as owner? |
|---|---|---|
| Popup Notifications | `player.IsAdmin \|\|` — hard | **No** |
| Zone Manager | `authLevel > 0 \|\|` — hard | **No** |
| Kits (`RequiredPermission`) | `Configuration.AdminIgnoreRestrictions && IsAdmin(player)`, and Kits' own `IsAdmin` is the `kits.admin` **permission**, not auth level. The shipped default is **`false`** | **Yes** |

So **phase 13 is demonstrable on this rig as it stands** — R16's entitlement gate applies to a server
owner like anyone else. **Phase 7 is not**, if its acceptance is "a grant made on the website gates a
third-party plugin in-game" against Popup Notifications or Zone Manager: that needs a **second,
non-admin Steam account**. Worth arranging before phase 7 rather than discovering there.

### 12.6 R18's trees, as they actually look

The four base plugins wrote their configs on first boot, so R18's two trees can be compared against
something real rather than predicted:

```
oxide/config/   Clans.json  Kits.json  PopupNotifications.json  ZoneManager.json
oxide/data/     clan_data.json  Kits/kits_data.json  Kits/player_data.json
                ZoneManager/zone_data.json
                oxide.users.data  oxide.groups.data  oxide.covalence.data  oxide.lang.data
```

R18's inventory of `data/` was exactly right. One nuance worth correcting, though: R18 motivates the
**recursive** walk with *"plugins nest (`config/<Mod>/x.json` and deeper)"*, and on a fresh install of
the base set **`config/` is flat — it is `data/` that nests.** The recursive walk is still correct
(other plugins do nest configs), but the nesting the plan cites as its reason is currently visible
only in the tree it must never walk.

And a reason to hold that boundary harder than R18 states: **`oxide/data/` is where Oxide keeps its
own permission store** (`oxide.users.data`, `oxide.groups.data`). A config editor that strayed one
directory over would be editing R2's mirror underneath itself.

## 13. Phase 1 as built — the transport, 2026-09-15

**Both criteria met.** A `server.hello` produced by the live Rust rig travelled game → sidecar →
module → the public website API; killing the sidecar left the game untouched; all five guards are
green on the module skeleton. Three repositories have their first commits:
[`Rust-Link`][rl], [`Rust-Plugins`][rp], [`Module-Rust`][mr], plus
[`rust-link/PROTOCOL.md`](../../rust-link/PROTOCOL.md) and
[`INTEGRATION.md`](../../rust-link/INTEGRATION.md) here.

### 13.0 The three org-lead decisions this phase needed

None of them were settled by §2, and each would have been expensive to reverse later.

- **D1 — the Rust bridge's docs live at `docs/rust-link/`**, a new top-level directory mirroring
  `docs/link/`, rather than under `modules/rust/`. It keeps the `uo`/`link` symmetry and keeps
  *module* docs separate from *bridge* docs, which are different contracts with different audiences.
- **D2 — loopback is the only trust boundary on the game link**, exactly as on the ServUO bridge:
  no token between plugin and sidecar. The alternative was argued on the grounds that Rust servers
  are far more often on GSPs than ServUO shards are, so binding to something other than `127.0.0.1`
  is a realistic operator need. **Overruled**, and the consequence is written into the plugin's
  class docs and `PROTOCOL.md` §1.1: moving that bind puts an unauthenticated command channel on the
  network, and it is documented as the mistake rather than defended against.
- **D3 — the plugin reads its settings from Oxide's own config file**
  (`oxide/config/RunicGateway.json`) rather than a standalone `Bridge.cfg`-shaped file. It is
  idiomatic for Oxide, and it lands inside R18's phase-7b config editor for free. The argument
  against — that editing `Host`/`Port` from the website could cut the link carrying the edit — is
  real and is now phase 7b's problem to guard rather than a reason for a second config mechanism.

### 13.1 What phase 1 deliberately did NOT register

The module registers **routes on three tiers and the two lifecycle hooks, and nothing else**. No
Team provider, no triggers, no audiences, no engagement seeds, no notification streams, no event
budgets/leases/actions/option sources, no extension slots.

That is asserted by a test (`nothing is registered that has nothing behind it yet`) so that removing
it is deliberate. The reasoning is worth keeping: **a declared trigger nothing emits and a declared
slot nothing fills are both surfaces an operator can configure and then wait on**, which is worse
than an absent one, because the absence is visible.

### 13.2 §11.3 was half wrong about `extensions`

§11.3 reads `module.json`'s `extensions` as *"declared, not just registered… like `mounts`, held
against reality by the loader"*. Reading the loader says otherwise, in two ways:

1. **The loader never checks that a declared slot was filled.** `loader.js:681` asks only
   `registries.hasSlot(slot)` — does this slot *exist*. `checkDeclared` covers `mounts` **alone**,
   in both directions. So a declaration with nothing behind it loads cleanly and means nothing.
2. **Only one of R13's two slots can be declared there at all.** `admin.users.detail` is the only
   server slot core declares (`router/v1/admin/users.router.js:202` is the sole `declareSlot` call
   outside the registry). `site.footer.status` is a **client** slot, registered from the chunk —
   naming it in `extensions` fails the load with `unknown extension slot "site.footer.status"`.

So the field is written when phase 6 registers the server half, and never before. Phase 4's
`site.footer.status` work does not touch it.

### 13.3 Five defects the rig found that no test could

Each of these was written from source reading, shipped, and then corrected by a live Rust server.
The hit rate is the phase-0 lesson repeating.

1. **A disconnect was silent in the game console.** The link teardown log sat in `LinkLoop`'s
   `catch`, and a connection that ends because the *reader* saw EOF leaves the writer to exit
   cleanly — nothing throws, so nothing is logged. The fix captures `_connected` at the top of the
   `finally`. Worth generalising: **a log in a catch block only covers the failures that throw**, and
   an orderly shutdown of a peer is not one of them.
2. **`Unload` blocked the main thread for 1.9 seconds**, which Oxide reports as
   `Calling 'Unload' on 'RunicGateway v0.1.0' took 1918ms`. The reconnect backoff was
   `Thread.Sleep`, and `Unload` joins the link thread — so every plugin reload froze the server for
   up to the backoff. Waiting on the same `AutoResetEvent` that `Unload` already signals makes it
   immediate. **The ServUO plugin has the same `Thread.Sleep`**, and it is survivable there only
   because ServUO does not hot-reload the way Oxide does.
3. **Mono's `SocketException.Message` is NUL-padded.** A connect refusal came back with ~200 `\0`
   bytes in the middle of the sentence, from a fixed-size OS buffer. `\0` is not whitespace, so
   `Trim()` does not touch it and neither does a whitespace-only collapse — the log line looks like
   it contains a huge run of spaces and no amount of trimming removes it. The flattener has to treat
   `char.IsControl` as a separator too. It took `od -c` on the log to see this at all.
4. **`bootId` regenerated on every PLUGIN load, not every SERVER start.** A fresh `Guid` at `Init`
   meant `oxide.reload RunicGateway` announced a brand-new boot — and §11.1's whole reconcile design
   hangs off that value, so every reload would have asked core to sweep its entire resource ledger
   for a world that never moved. It is now `Process.StartTime`, which is exact, identical on every
   read, and changes when and only when the thing it names changes. **Verified by reloading the
   plugin twice and watching the id hold** (`boot-20260915T194502Z`, matching the process).
5. **A four-connection SQLite pool over `:memory:` hands out four empty databases.** An in-memory
   database is per *connection*, so the schema created on the first pooled connection is invisible
   to the second. It presents as `no such table` from a random subset of queries. The sidecar now
   caps the pool at one connection for an in-memory path — which is the only coherent reading of
   `:memory:` and is what makes it usable at all. The ServUO sidecar never hit this because it only
   ever opens a file.

### 13.4 Two things the kit's own template got wrong for this module

Both are feedback for phase 19, and both are the kit being right about the general case and specific
about the wrong detail.

- **`registration.test.js` reads one page by NAME** (`src/routes/public/Clan.jsx`) to check that
  every declared slot is rendered somewhere. A module that declares no slots — as phase 1 does — dies
  on `ENOENT` before reaching the loop that would have been empty. Generalised here to scan every
  file under `src/routes`.
- **`test/_fakes.js` supplies `validator: {}`.** The template's routers never use express-validator,
  so `{}` is enough for them; an admin router that builds validation chains at file scope cannot be
  *required* with it. The fake now holds the real library, for the same reason it holds a real
  express Router: a fake of either would only ever test the fake.

The kit was also **right in a way worth recording**: `noGameConnection.test.js`'s header predicts, in
so many words, that a module adding a sidecar client will see this check go red and tells the reader
to narrow it rather than delete it — naming `sidecarClient.js` as the file to allow. That is exactly
what happened, on the first run, and the fix was the one line the header names.

### 13.5 The player tier is thin on purpose

R14 puts the module on all three tiers from the start, and the loader holds `mounts` against what is
registered in both directions — so the declaration and the registration land together or not at all.

What the player tier will carry is the signed-in view of a server: the viewer's linked Steam
identity, their presence, their entitlements. None of that exists before phase 6. So the one route
there answers the server list on the authenticated tier, delegating to the **same model** the public
tier uses, so the two cannot drift while they are meant to be the same.

That is a real route rather than a placeholder: it is the address the app and the SPA will call, and
it starts answering correctly now rather than moving later.

### 13.6 Three timeouts in a row, and the ordering is load-bearing

```
sidecar RPC reply timeout (10s)  <  module client timeout (12s)  <  an action's budgetMs
```

Core classifies a `budgetMs` overrun as retryable **unconditionally** — it cannot ask the action,
which is still awaiting a socket. So an action whose budget does not exceed the module's client
timeout can never report `retry: false`, and that branch is unreachable code. Phase 13's actions
must derive their budgets from `sidecarClient.TIMEOUT_MS` rather than writing a number beside it.

The first module this project shipped got this the wrong way round and retried a verb it had
explicitly refused, which is why all three values are now written down in one place
(`PROTOCOL.md` §4.4) instead of three.

### 13.7 The rig, as it stands after phase 1

The phase-0 rig plus the bridge. `RGProbe.cs` is still installed beside `RunicGateway.cs` and is
still useful — its `rg.checkperm` / `rg.zoneonme` commands are phase 7 and phase 12 instruments.

```
D:\rust\oxide\plugins\    Clans.cs  Kits.cs  PopupNotifications.cs  ZoneManager.cs
                          RGProbe.cs  RunicGateway.cs
rust-link-sidecar         game 127.0.0.1:7799 · web 127.0.0.1:8090 · its own scratch config
website (edge)            modules/rust/ installed as a directory; server row "main"
```

A dependency-free WebSocket RCON driver was rebuilt this phase (the phase-0 one lived in a scratchpad
that did not survive). `rcon.web 1`, port 28016 — the password is in `D:\rust\start.bat`. It is the
only way to reach `rg.link` without a console session, and phase 7 will need it again.

**The phase-0 ceiling still stands and still blocks phase 7:** no console session can observe a
permission gate, because every plugin's check short-circuits without a `BasePlayer`, and an admin
account bypasses most of them non-uniformly. A second, non-admin Steam account has to be arranged
before phase 7 — it is the one prerequisite this rig cannot satisfy on its own.

## 14. The Pterodactyl rig as built, 2026-09-15

R21's first server exists, made with the application key and driven with the client key. **Both
credentials work; neither can do the other's job.** What follows is what building it actually taught,
including one finding that changes R20's shape.

### 14.0 The rig

| | |
|---|---|
| Panel | `http://192.168.0.12` (no TLS — `https://` fails outright), node 1 `Main` |
| Servers | `rust-oxide` id **17** / **`e6758c06`**, and `rust-carbon` id **18** / **`87fb1f67`** (§14.5) |
| Egg | 18 `Rust Autowipe`, `ghcr.io/pterodactyl/games:rust` — both rigs, one egg |
| `FRAMEWORK` | `oxide` / `carbon` |
| Limits | 8192 MB memory, 25600 MB disk — deliberately under half the node, so the Carbon rig fits beside it |
| Allocations | oxide 21000-21004, carbon 21005-21009 — game, query, RCON, Rust+, **and one held for the sidecar's `[web].bind`** |
| World | procedural, size 3000, seed 1234 |
| SFTP | `192.168.0.12:2022` |

The RCON passwords are generated 24-byte tokens rather than the old rig's `letmein`, kept out of this
document and out of the repo. §6 named that shape as the thing nothing published should copy; this is
the first rig where it was not copied.

**A Rust server install is about 6 GB, not the ~20 GB this plan assumed** when it worried about node
capacity — measured at 5,894 MB with the game installed and the world generating. Two rigs are
comfortable on a 128 GB node, and the 25600 MB limit is generous rather than tight.

### 14.1 The two keys, and what each one is actually for

Confirmed by use rather than by reading:

| | Application (`ptla_`) | Client (`ptlc_`) |
|---|---|---|
| Create / configure a server, assign allocations | **yes** | no |
| List, read, power, console, **files** | no (`403`) | **yes** |
| Write or import an egg | **no** — `/api/application/eggs` 404s; eggs are an admin-UI or `php artisan` operation | no |

So the full loop needs both, and **a published egg is a JSON file a human imports** — which is also
exactly how an operator will consume ours, so it is a constraint worth designing into rather than
around.

**File operations are refused during install** with `409 ServerStateConflictException` —
*"this server has not yet completed its installation process"*. Anything that pushes files has to
wait for `is_installing: false`, not merely for the server to exist.

### 14.2 The correction: there was never a panel bug

An earlier pass through this section recorded that the client key "authenticates and then lists zero
servers", and reasoned from there to a second claim — that *includes are broken on this panel*,
because `/api/application/servers?include=user` returned an empty list where the same route without
the include had returned six.

**Both were wrong, and wrong the same way.** The servers were being deleted while the probing
happened, so two calls minutes apart read two different worlds. Once a server existed, every one of
those calls answered correctly on the first attempt — the client list, the single-server route, and
`include=user`.

The reusable part is not "check twice". It is that **a differential diagnosis across two API calls
silently assumes the state did not move between them**, and on a live panel somebody else is also
holding the controls.

### 14.2b The upload loop, proven with the real plugin

Not a hello-world: phase 1's actual `RunicGateway.cs` (27,642 bytes, 709 lines) was pushed straight
from the working tree with the client key, and it came back byte-identical on read.

Three things that worked and were not certain to:

- **`files/write` creates missing parents.** `/oxide/plugins/` did not exist — the framework is laid
  down at boot (§14.3), and the server had never been started — and the write created the whole path.
- **A plugin placed before Oxide exists survives Oxide arriving.** The entrypoint's `unzip -o` over
  `oxide/` left the file untouched, so the push does not have to wait for a first boot.
- **It compiled and loaded on Linux**, which no previous phase had ever established. Every prior
  finding came from Windows and Mono:

  ```
  02:19 [Info] RunicGateway was compiled successfully in 0ms
  02:19 [Info] [Runic Gateway] protocol 1, serverId 'main', sidecar 127.0.0.1:7799
  02:19 [Info] Loaded plugin Runic Gateway v0.1.0 by RunicGateway
  02:19 [Info] [Runic Gateway] cannot reach the sidecar: Connection refused - retrying quietly
  ```

  That last line is phase 1's no-stall contract holding on a second platform: no sidecar exists on
  this host yet, the plugin says so once and keeps the game running.

**Read the console without a websocket.** Pterodactyl streams console over a websocket, which is
awkward to drive from a script — but `wrapper.js` also writes `latest.log`, and Oxide writes
`oxide/logs/oxide_<date>.txt`. Both are plain reads through `files/contents`, which is how every log
line quoted in this section was obtained. Worth knowing before anyone writes a websocket client.

### 14.2c The tier-2 loop, end to end

R21's middle tier is the one that has to be pleasant to use, so it was run rather than described.
One pass: patch the working-tree source so the change is visible in the game console, push, reload
through the client API's `command` endpoint, read Oxide's log back, then restore.

```
patched source: True
push -> HTTP 204
oxide.reload -> HTTP 204
02:27 [Info] RunicGateway was compiled successfully in 3392ms
02:27 [Info] Unloaded plugin Runic Gateway v0.1.0 by RunicGateway
02:27 [Info] [Runic Gateway] protocol 1 [PTERODACTYL-PUSH-PROOF], serverId 'main', sidecar 127.0.0.1:7799
02:27 [Info] Loaded plugin Runic Gateway v0.1.0 by RunicGateway
restored source and re-pushed -> 204
```

**Roughly ten seconds from a saved edit to a reloaded plugin**, against a running server with a
generated world, without touching the panel UI. That is the loop `deploy.ps1` gives us for ServUO,
and it is the thing that makes the panel a workable rig rather than only a deployment target.

Four details worth carrying into the push script:

- **Reload is `POST /command`, not a file operation**, and it answers `204` whether or not the plugin
  actually came back. The proof has to be read out of `oxide/logs/` afterwards — the same shape R18's
  auto-rollback needs, and an early rehearsal of it.
- **The unload/load pair straddles the plugin's own `Init` log line.** `Unloaded` is printed, then the
  new instance's startup line, then `Loaded`. A script that waits for `Loaded` before reading has
  already passed the line it wanted.
- **Oxide's compiler idles out and restarts.** The boot compile was `0ms`; the reload compile was
  `3392ms` because `Shutting down compiler because idle shutdown` had happened in between. A timeout
  tuned against a warm compiler will be wrong on the first reload after a quiet period.
- **Restore the working tree and re-push it.** A test that leaves a marker in the source is a test
  that ships a marker. Both were put back and verified byte-identical against the server copy.

### 14.3 The image installs the framework on **every boot**, and neither version is pinnable

`ghcr.io/pterodactyl/games:rust`'s entrypoint is where `FRAMEWORK` is consumed — **not** the egg's
install script, which knows nothing about it. On every single start, before the game runs, it:

- runs `steamcmd +app_update 258550` unless `AUTO_UPDATE=0`;
- for `carbon`, downloads
  `CarbonCommunity/Carbon.Core/releases/download/**production_build**/Carbon.Linux.Release.tar.gz`;
- for `oxide`, downloads `OxideMod/Oxide.Rust/releases/**latest**/Oxide.Rust-linux.zip`.

**Both are moving targets, fetched fresh at every restart.** CARBON.md §8 predicted this for Carbon
from its rolling release tags; the egg makes it true of *Oxide as well*, because `latest` is the same
kind of promise. The consequence is sharper than "the rig may drift":

> **A restart is a framework upgrade.** Two runs of the same test on the same server, minutes apart,
> are not guaranteed to be running the same framework build — and nothing in the panel says so.

That reaches three places. **R4's `doctor`**: the weaker "current enough" claim is not Carbon-specific
after all; under the egg neither framework has a pinned version to check. **§6's wipe-cadence risk**:
the re-verify step is per *restart*, not per wipe. And **R20 itself**: if the egg is our deliverable,
whether it should pin the framework at all is a decision, not an oversight — the upstream egg's
answer is "always newest", which is right for an operator on wipe day and wrong for a test rig
trying to reproduce a finding.

### 14.4 **The trap that changes R20: the startup string is not a safe place to launch the sidecar**

R20 says the startup command becomes "a small wrapper that launches `rust-link-sidecar` and then
`RustDedicated`". The mechanism allows it and the ordering makes it wrong.

`wrapper.js` runs the startup string through `child_process.exec`, which is `/bin/sh -c` — so
`./rust-link-sidecar & ./RustDedicated …` is syntactically fine. **But for Carbon the entrypoint
prepends to the whole string:**

```bash
MODIFIED_STARTUP="LD_PRELOAD=$(pwd)/libdoorstop.so ${MODIFIED_STARTUP}"
```

So a startup beginning with our sidecar becomes:

```bash
LD_PRELOAD=…/libdoorstop.so ./rust-link-sidecar & ./RustDedicated …
```

**The preload lands on the sidecar and not on the game.** Carbon loads through Doorstop rather than
through a patched `Assembly-CSharp.dll`, so the result is a server that starts cleanly, reports no
error, and **is not modded** — no plugins, no hooks, and a bridge that connects to a game it can
never hear from. It is the exact silent-success failure §6 keeps cataloguing, and it would only ever
appear on the Carbon half.

Two further consequences of the same handoff:

- **`quit` SIGTERMs the shell, not the sidecar.** `wrapper.js` kills `gameProcess`, which is the `sh`
  running the startup string; a backgrounded sidecar is not its child in the way that reaches. R20
  already required "stop means stop the game" — this is the mechanism by which it would fail, and it
  leaves an orphan holding port 21004 against the next start.
- **Doorstop also confirms R21's clean-install rule from a second direction.** Switching `FRAMEWORK`
  on an existing install does not undo the other framework: Oxide's patched DLL stays on disk while
  Carbon preloads over it. The migration argument was the soft reason for a fresh Carbon rig; this is
  the hard one.

**So R20 needs a decision it did not know it needed:** the sidecar is launched by something other
than the startup string — our own image or entrypoint layered on the upstream one — or the startup
string is composed so that whatever the entrypoint prepends still lands on `RustDedicated`. The first
is more work and survives upstream changing its entrypoint; the second is free and depends on a line
in somebody else's repository. Raised rather than settled.


### 14.5 The Carbon rig, and R19 proven

`rust-carbon` — id **18**, identifier **`87fb1f67`**, same egg, same world (procedural, 3000, seed
1234), same limits, allocations **21005-21009 with 21009 held for the sidecar**, `FRAMEWORK=carbon`.
**A clean install, never a converted one**, per R21: install 122s, boot 543s, running.

**R19 is proven.** The byte-identical `RunicGateway.cs` that runs on the Oxide rig — no `#if CARBON`
anywhere in it, nothing conditional at all — loaded on Carbon **2.0.259.0** and behaved the same:

```
[INFO] Carbon 2.0.259.0 [2026.09.03.0] 21063e8 on Linux
[INFO] [Runic Gateway] protocol 1, serverId 'main', sidecar 127.0.0.1:7799
[INFO] Loaded plugin Runic Gateway v0.1.0 by RunicGateway [2367ms]
[INFO] [Runic Gateway] cannot reach the sidecar: Connection refused - retrying quietly [RunicGateway Link|26]
```

That last line is the no-stall contract holding on its **third** platform now — Windows/Mono,
Linux/Oxide, Linux/Carbon — from one source file. [`CARBON.md`](CARBON.md) §10 is the full scorecard;
the parts that change decisions are below.

**R18's amendment is confirmed the best way it could have been.** The plugin's own config, written by
the same Oxide-compat API on both rigs, landed at `/oxide/config/RunicGateway.json` on one and
**`/carbon/configs/RunicGateway.json`** on the other. D3 put the plugin's config inside R18's editor;
had that editor used a literal `oxide/config/`, **it would not have found its own plugin's config on
half of all installs.** No test would have caught it; only two rigs would.

**And one claim was refuted — the one with the sharpest consequence.** `CARBON.md` had said Oxide
stores permissions as JSON and Carbon as Protobuf or SQLite, offering the difference as the reason
not to read the file. Both rigs say otherwise:

| | Oxide rig | Carbon rig |
|---|---|---|
| Path | `oxide/data/oxide.users.data` | `carbon/data/`**`oxide.users.data`** |
| First bytes | `0a 16 0a 07 64 65 66 61 75 6c 74` | `0a 17 0a 07 64 65 66 61 75 6c 74` |
| Format | Protobuf | Protobuf |
| Default groups | `default`, `admin` | `default`, `admin`, **`moderator`** |

**Same binary format, same filenames, different directory** — and Carbon writes *its* data into files
named after Oxide. **This makes R2's API-only rule more important, not less.** A file reader would
have worked on both rigs today and broken silently for the one operator who ran `c.migrate_perms_sql`
— no error, no version marker, just a site reporting drift against a store nobody is writing any
more. The rule survives; the reasoning behind it was wrong and is now right.

Two things nobody had thought to claim, found by looking:

- **Carbon auto-creates a third default group, `moderator`**, auto-granted by auth level alongside
  `default` and `admin`. R2 pushes its *full* set on connect, so it has to tolerate a group the
  framework will recreate the moment it is deleted — otherwise the site reports drift for ever.
- **`c.plugins` reports per-plugin `hook fires`, `hook time`, `hook memory`, `hook lag` and
  `hook exceptions`** — most of §6's "log which expected hooks have fired" mechanism, free, and only
  on Carbon. Useful when debugging on Carbon; **not a substitute** for the plugin's own counter, which
  must work on both.

**A warning about how a wrong console command fails here.** Pterodactyl's `command` endpoint returns
`204` whether or not anything happened, and Carbon prints nothing for an unknown command. So
`oxide.plugins` on Carbon — which is simply not a command — is indistinguishable from success at the
API. Anything driving the console has to read a log to know, which is the same conclusion §14.2c
reached about `oxide.reload` and the same shape R18's rollback needs.


## 15. Phase 2 as built — packaging and release, 2026-09-16

The first phase with no game server in it, and the one that turns a directory somebody copied into a
thing an operator can install. [`Module-Rust#2`][mr] is the whole of it.

**The starting position was worse than the phase row implied.** Phase 1 built five guards and ran
every one of them by hand: the repository had no `.gitea/workflows/` at all. So nothing gated the
branch that gets released, and there was no way to release it. This phase adds both halves, and the
org lead widened it to include the gate rather than only the release.

### 15.0 The three decisions this phase needed

- **D4 — the full CI suite, not only `release.yml`.** The phase row names the release; the repo had
  no gate at all. Both ship, and the frozen-manifest job ships with them rather than waiting for a
  later phase — see §15.2 for what it found on its first run.
- **D5 — the bundle carries no `node_modules`, and the emptiness is asserted rather than assumed.**
  The shipped half declares no runtime dependencies: everything it needs arrives on `ctx`. So the
  release runs no `npm ci` and packs no dependency tree. The whole value of that decision is that
  *the day it stops being true is a loud day*, so `checkBundle.js` fails the PR that adds a
  `dependencies` entry without also teaching the release to install and pack it. Module-uo ships
  `ws` and does the opposite; this is a different answer to the same question, not a divergence
  from a rule.
- **D6 — no SonarQube for the three Rust repos yet.** Phase 2 is packaging; scanning is quality
  tooling and phase 19 is already the sweep where `CLAUDE.md`'s project-key table gets updated.
  Three half-populated projects while the repos are skeletons buy nothing.

### 15.1 What is in a release, and what decides its number

A release is **not source**. It is the directory core's loader expects at `modules/rust/`, already
assembled — the prebuilt client chunk, the schema fragment and the OpenAPI fragment, packed as they
will be unpacked — because an operator never builds anything (`MODULE_SYSTEM.md` §1.14).

| Asset | What it is |
|---|---|
| `module-rust-<version>.tar.gz` | the assembled directory, one top-level entry, no `node_modules` |
| `module-rust-<version>.json` | the install manifest: id, name, version, `coreApi`, url, size, **`sha256`** |
| `SHA256SUMS` | the same hash, in the shape every other repo here publishes |

The version is **derived**, using the engine `link`, `installer` and `Module-uo` already run:
conventional-commit subjects since the newest `v*` tag decide major/minor/patch, nothing releasable
cuts no release, and `module.json`'s version survives as a **floor** rather than as a record. The
number that ships is the **tag**, and CI stamps it into the bundle's own `module.json` at assembly
time — which is why the release also asserts that the assembled `module.json` carries the version
being released, since a bundle that still declared the floor would install under a number that is
not the one it came from.

Module-uo is the argument for deriving rather than declaring: it released only on a hand-edited
version line, and between 2026-08-12 and 2026-08-19 that cost it *every* bundle, because nine phases
of work landed without anyone touching that line. `workflow_dispatch` survives as the backdoor for
the case the rules cannot reach — a widened `coreApi`, a new mount, a capability, with no releasable
code behind it.

**Two failure modes are guarded before anything is built.** The credential check runs first, so a
repository without `REGISTRY_TOKEN` fails with a named error instead of pushing a tag and then 401ing
on the release API — the state `servuo-plugins` got stuck in on its own first release. And a tag that
exists with **no release behind it** is deliberately not treated as "nothing to do": that is the
signature of the same half-failure, and standing down on the tag alone would make it permanent.

### 15.2 The frozen manifest answers the question phase 1 had to take on trust

`§13`'s registration comment says `/rust` "collides with nothing on any of the three tiers, checked
against core's mount tables rather than assumed" — and then names the limit of that check: **core
answers several public routes mounted at the tier root rather than under a prefix**, `/status` and
`/version` among them, which the loader's own collision probe cannot see.

The `frozen-manifest` job is the thing that can. It clones core at the sha pinned in
`ci/core-ref.json`, generates core's route table **without** this module and then **with** it, and
takes the difference. That difference is what the module serves, and it is checked three ways: it
must match the committed `routes.manifest.json`, every route in it must have an operation in
`swagger-fragment.json` *and* every operation must be a route, and — the half that matters most —
**no core route may have been removed or changed**. A module whose mount displaced a core route
cannot show up as an addition, because the URL is unchanged; only the diff sees it.

Run locally against the pinned ref before it was committed, and again in CI against a fresh clone,
both agreeing:

```
route manifest up to date (280 routes)
wrote routes.manifest.json (282 public + 4 internal)
routes.manifest.json is current — 6 routes, all documented
```

Six routes, all documented, nothing of core's moved. **`/rust` is now free by proof rather than by
reading**, and it stays that way on every pull request.

**The pin starts on `main`, unlike Module-uo's.** That repo spent the whole Event System window
pinned to `edge`, because it depended on contract members that had not reached `main` yet. This
module needs `MODULE_API` 1.10.0, which the Event System cutover already put there, so
`ci/core-ref.json` names `efa9db7` on `main` and should stay on `main` until this module comes to
depend on something unreleased.

### 15.3 One declaration, two readers — and it was verified by breaking it

`ci/bundle.json` is an **include** list, never an exclude list: an exclude list ships whatever it
forgot, and the day somebody adds `server/tools/` with a scratch credential in it, an exclude list
packs it and nobody finds out. The cost of that choice is the opposite failure — a new directory
silently drops *out* of every release — which is exactly what happened to Module-uo between v0.3.0
and v1.0.0, where `server/commands/` arrived in a cutover, the list did not learn about it, and the
module installed cleanly and then died at the register stage on the operator's box.

Nothing caught it there because the PR checks copy the **whole repo** into core: they only ever
exercised a tree that had the file. **The subset exists only in the release.** So the list has two
readers — `release.yml`, which packs from it, and `server/scripts/checkBundle.js`, which asks on
every PR whether it still covers everything `server/index.js` can reach.

It reaches requires written **inside `register()`**, which is not a detail: this module's entry point
requires its routers inside the function on purpose, because `core.init(ctx)` has to run before
anything under `router/` is required. A check that only saw file-scope requires would have missed
every router the module has.

Verified the only way a check is worth anything — by breaking it. Dropping `"model"` from the list:

```
ci/bundle.json does not ship everything server/index.js reaches.
  server/model   (2 files reachable)
Add "model" to ci/bundle.json's server[].
```

It names the exact edit, in the units the list is written in.

### 15.4 The release, as published

`v0.1.0`, cut by the first push to `main`, with no tag before it — so the engine took the first-run
branch and shipped what `module.json` declared. Three assets:

| Asset | Size |
|---|---|
| `module-rust-0.1.0.tar.gz` | 41,310 bytes |
| `module-rust-0.1.0.json` | 355 bytes |
| `SHA256SUMS` | 91 bytes |

Downloaded and hashed independently of CI, the artifact is
`7296c76b988c6191840a1dc4ed1a77d96ad35e8c26f9b6b55e88fe4c10878b32` — byte-for-byte the `sha256` the
manifest declares.

**`REGISTRY_TOKEN` was already configured on the repository**, which was the one prerequisite this
work could not verify for itself: the bot identity can read neither repository nor organisation
Actions secrets (`user should be the owner of the repo`). The credential check exists precisely
because the answer was unknowable from here, and the first release ran clean through it.

### 15.5 The acceptance criterion, walked

> *An operator installs the empty module from Admin → Modules and it reaches `started`.*

**Met.** Walked as an operator would, against the local core on `edge`, with the phase-1 state
deliberately torn down first — the hand-copied `modules/rust/` directory moved aside and its
`installed_modules` row deleted, so this was a first install and not an upgrade. Before it,
`GET /api/v1/public/rust/servers` answered **404**.

1. Admin → Modules, with no `rust` row and `gitea.whitlocktech.com` on the allowlist.
2. Pasted the release's `module-rust-0.1.0.json` URL and pressed Install.
   → *"Installed Rust v0.1.0. Restart to load it."*, and a row reading **Restart to start —
   installed, it mounts when the server next starts**, carrying the manifest URL and
   `sha256 7296c76b988c…`.
3. Restarted.
   → `registered module "rust" v0.1.0`, `schema ensured for module "rust" {"statements":2}`,
   `[rust:boot] booted {"refreshMs":30000}`, `module "rust" started`.

The row is now `state = started` with `started_at` set, the source URL and the released hash — where
phase 1's row had `source` and `sha256` **null**, because a directory somebody copied has no
provenance to record. The screen reads **Running — mounted and serving**.

**What the volume received is exactly the include list and nothing else: twenty files.** No tests, no
`server/scripts/`, no `ci/`, no `client/src/`, no `node_modules`, and no dotfiles. `client/dist/entry.js`
is the chunk CI built.

And the three checks that prove it is *serving* rather than merely loaded:

| | |
|---|---|
| `GET /api/v1/public/rust/servers` | **200**, answering with the rig's `main` server |
| `GET /api/v1/player/rust/servers` · `GET /api/v1/admin/rust/servers` | **401** — the tier gates are on, not bypassed |
| `/api/docs.json` | carries all **five** `/rust` paths: the fragment merged into core's own spec |

**Two things this walk is honest about.** The first is that the *restart* was done from the shell
rather than with the screen's own **Restart the server** button: that button runs the same graceful
shutdown a `SIGTERM` does and relies on a supervisor to bring the process back, which the shipped
`docker-compose.yml` provides and a bare `npm start` does not. The button was not exercised, and it
is core's, not this module's. The second is that the module's two tables were left in place when the
row was deleted, so the rig's `main` server row survived the reinstall — the schema fragment replayed
on boot (`statements: 2`) exactly as R12 requires of a fragment that runs every time.

**The one operational finding, and it is about the rig rather than the module.** Docker Desktop's
Linux engine was found dead — its WSL distribution stopped, the `uomm-db` container exited, and
`docker` answered every call with `500 Internal Server Error` rather than anything naming the cause.
Restarting Docker Desktop and the container fixed it. Worth writing down because the failure presents
as the *website* being broken (`ECONNREFUSED` to a database that is simply not there), and because
`CLAUDE.md` points every smoketest at that one container.

## 16. Phase 3 as built — the read path, 2026-09-16

The first phase that had to be true on two mod frameworks, and the first with a
catalogue rather than a message. Four repositories moved: the spec here, the plugin, the sidecar,
and the module.

**Status: the bridge half is done and proven; two proofs are queued on the org lead.** What the
plugin sends and what the sidecar does with it are built, tested and exercised against live Oxide
and Carbon servers. The player-facing half of the catalogue — deaths, chat, gathering, sessions —
cannot fire without somebody holding a mouse, and is written down as a walk to run rather than
guessed at: [`PLAYER_WALK.md`](../../rust-link/PLAYER_WALK.md). §16.7 lists everything still open.

### 16.0 The four decisions this phase needed

- **D7 — the widest first hook wave.** The options ran from "presence, deaths and the wipe" to
  "everything read-only worth having", and the widest was chosen: fifteen hooks, including the
  moderation set that carries IP addresses and player reports. The consequence is real and is
  designed around rather than deferred — those frames arrive **nine phases before** the visibility
  framework phase 14 builds, so the classification and its default-deny allowlist ship now (§16.4).
- **D8 — the plugin derives `wipeId`.** `PROTOCOL.md` §3.2 had reserved that for the website. By
  protocol 2 three components store rows that need it and only one of them can read the value, so
  the reversal is written into §8.2 rather than left as a contradiction.
- **D9 — the live feed is a cursor, and D5 stands.** Core runs Node 20, where a global `WebSocket`
  is still behind a flag, so a socket means taking `ws` — against a release that asserts it ships no
  runtime dependencies. The deciding argument was the other one: **a socket needs a cursor anyway**
  for what it missed while the module was restarting, and the catch-up path is the one that must be
  right. One mechanism exercised every five seconds beats two where the second only runs after an
  outage nobody planned.
- **D10 — rollups permanent, raw bounded.** The website keeps per-player-per-wipe totals for ever
  and a 30-day window of raw events; the sidecar keeps 14 days and prunes hourly. R12's "a wipe does
  not erase a player's history" is met by the totals, which is the row an operator actually reads.
- **D11 — CI for both bridge repositories**, which had none at all. Phase 2 found that hole in
  Module-Rust; it was still open in the two repositories that ship the half running inside somebody
  else's game server.

### 16.1 `type` is the whole of protocol 2 in the sidecar

Protocol 1 routed on `kind`, in a `match` that needed a new arm per addition. Protocol 2 adds
**`type`** — `event`, `snapshot`, `reply`, `control` — and the sidecar files on that and nothing
else. Ten new event kinds are now zero change in Rust-Link, which is the property that matters when
the thing growing fastest is the catalogue.

A frame whose `type` this build does not know is **dropped and counted**, never guessed at.
Defaulting an absent one to `event` would file a *board* as history — the presence board appended a
few thousand times, which nothing reports and nobody notices until they wonder why the database is
large.

**It caught a real mismatch three seconds after it first ran**, which was not planned: a protocol 1
plugin was still live on the retired workstation rig, dialled the new sidecar, and its `server.hello`
went straight into the counter. The game link has no version handshake by design (§2), so
`untyped_frames` on `/health` is the only place that failure is visible — and the symptom without it
is a website showing nothing while the game is plainly up.

### 16.2 Two defects a live server found, and neither could have been found anywhere else

**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 shipping without the field R12 splits history on.

It was invisible for the reason such things usually are: a **hot-reloaded** plugin reads an
already-loaded world and gets the right answer every time. Every development iteration on the
workstation rig was a hot reload. 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. Now resolved again at
`OnServerInitialized`, and lazily while still unknown.

**`Unload` blocked the game's main thread for two seconds.** Carbon reported it exactly:
`hook 'Unload' took longer than 100ms [2002ms]`, next to `link thread did not stop cleanly`. That is
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, and `Unload` joins it from
the main thread.

The reason two phases missed it is worth keeping: **a host that refuses answers instantly, and a
host that drops does not answer at all.** Every loopback test is the first kind. A firewalled
address, a typo, a machine that is off are all the second, and the deployment this phase was being
tested through happened to be one. The connect is now bounded and waits on a stop handle of its own
— it cannot share `Wake`, which also means "the queue has something in it" and is signalled by every
hook that fires. After the fix the same reload logs no slow-hook warning and no stranded thread.

### 16.3 The aggregate, and the rule it generalises

`OnDispenserGather` fires on **every swing at a tree**. A frame per swing would make the bridge the
most expensive thing on a busy server, and nobody wants a killfeed of chickens either, so gathering
and NPC kills are counted in the plugin and flushed once a minute as one `player.tally` frame.

The rule: **if a hook can fire more than once a second per player, it is a counter, not an event.**
R17's warning about chatty zone transitions is the same rule arriving early.

A tally is a **delta, not a running total** — what happened since the last flush — so the consumer
sums rather than diffs, and a dropped frame costs one interval instead of corrupting the series. The
outbound queue is drop-oldest by design, so frames are genuinely allowed to go missing; a running
total over a lossy link is a number that is quietly wrong for ever.

One honest limitation, corrected in the code rather than in the comment that first claimed
otherwise: **a plugin reload loses up to a minute of one player's tally.** `Unload` enqueues the
flush, but the writer stops on the same flag and the queue is cleared after the join. Draining it
first would mean waiting on a socket from the main thread — the stall §16.2 just removed — so the
loss is taken deliberately. A real shutdown flushes at `OnServerShutdown`, and a player leaving
flushes at their disconnect.

### 16.4 The boundary is enforced by the side that serves

The widest hook wave brings IP addresses (`CanUserLogin`, `OnUserApproved`, `OnUserBanned`), one
player's report about another, and the grid reference of somebody's base — nine phases before the
visibility framework §11.2 costed. So the classification ships with the catalogue.

**It is not a field on the wire.** The plugin could have stamped a class on every frame; it
deliberately does not. A boundary declared by the *sender* is one a compromised — or merely
out-of-date — game host can widen. The website's own shard fan-out works the same way: a public
stream with an allowlist of kinds and an admin stream that adds the rest, and what makes it
trustworthy is that the decision lives on the serving side.

So `module-rust/server/catalogue.js` holds it, **default-deny**: a kind this build has never heard
of is not public. That is the shape of the mistake it prevents — the next protocol version adds a
kind, the module stores it happily, and a deny-list filter would publish it the day it first
arrived, before anybody decided whether it should be. A test holds the list against §8.4's table, so
adding a kind to the protocol without classifying it fails a build.

### 16.5 A login denial is not a hook, and §10 says it is

`PLAN.md` §10 sources the `rust.login.denied` trigger from `CanUserLogin`. Reading the hook says that
cannot work: it fires on **every** connection attempt, and the only way to learn of a denial from it
is to *be* the denier — which §8.7 forbids, structurally, by declaring every read-path hook `void` so
it cannot answer. uMod publishes no `OnUserRejected`.

What the game can tell us is two facts: an attempt, and an approval. Protocol 2 emits both, and a
denial is **the absence of an approval** — a deferred read, phase 10's to make. The trigger survives;
its source changes. (The same shape the engagement workstream hit at its own phase 10, which is
either a coincidence or a property of login paths.)

### 16.6 What was proven, and how

| Claim | How | Result |
|---|---|---|
| The read path compiles and loads on **Oxide** | live server, protocol 2, 15 hooks bound | ✅ |
| …and on **Carbon** 2.0.259.0, from the **byte-identical file** | the Pterodactyl rig, config read from `carbon/configs/` | ✅ |
| Every frame carries `type`, `serverId`, `wipeId` | `server.hello` and `players.online` read back off the sidecar | ✅ |
| A wipe id derived from the save, changing only with the save | `rg.link` reports `w-20260915T195817Z` against `saveCreatedAt 2026-09-15T19:58:17Z` | ✅ |
| **A restarted sidecar is fully populated within one connection** | store deleted, process restarted: both boards present **0.3 s** after the listener bound, and **zero** events in history | ✅ |
| Boards are not replayed as history | the same walk: `/events` returned 0 rows while `/boards` returned 2 | ✅ |
| Moderation frames reach the sidecar whole | `banid` / `unban` over RCON | ✅ |
| A protocol mismatch is counted, not mis-filed | a live protocol 1 plugin against the protocol 2 sidecar | ✅ |
| `rg.hooks` answers on both frameworks | Oxide and Carbon, identical output shape | ✅ |
| 44 sidecar tests, 95 module server tests, 20 client tests, every guard | locally, and now in CI on both repos | ✅ |
| The module's route manifest against a **real core** at the pinned ref | 10 routes, all documented, none of core's moved | ✅ |
| **The module ingests a live game** | the working tree installed into a running core, pointed at the workstation sidecar: `cursor started at the feed tail`, then a console `banid`/`unban` arrived as two `rust_events` rows with their wipe id, cursor advanced | ✅ |
| **The allowlist holds against real rows** | with both ban events in the table, `GET /public/rust/servers/:id/events` answered `{"events":[]}` — **and answered the same when asked for `player.banned` by name** | ✅ |
| A wipe row is created by being mentioned | `w-20260915T195817Z` appeared in `rust_wipes` from the first frame carrying it, with no "a wipe started" call anywhere | ✅ |

**One finding about the rigs rather than the code:** the panel rigs cannot reach a sidecar running on
the workstation, because Windows Firewall holds two program-scoped **Block** rules for
`rust-link-sidecar.exe` — created by a dismissed prompt at some point — and a program-level block
beats any port-level allow. Removing them needs elevation. It is a rig problem only: the shipped
design puts the sidecar on the game host's own loopback (D2, R20), where it is the deployment that
never needs a rule at all.

The local workstation rig, which does reach its sidecar on loopback, is what proved everything in the
table above that needs a live socket. **`D:\rust` is therefore not as retired as R21 assumed** — it
survives as the fast loop (a saved file is a reloaded plugin in about ten seconds, against nine
minutes of world generation on the panel), and the panel rigs are what answer "on both frameworks".

### 16.7 What is still open, and who it is waiting on

Three things, all of them measurements rather than decisions, and all of them the org lead's to run:

1. **The player walk** — [`PLAYER_WALK.md`](../../rust-link/PLAYER_WALK.md). Ten minutes on a rig
   with a mouse, and it closes `OnPlayerDeath`, `OnPlayerChat`, `OnDispenserGather`,
   `OnPlayerRespawned`, `OnEntityDeath`-by-a-player, and `sessionSec`. The document says what each
   step should produce, so it can be run without anybody watching the output live.
2. **The wipe walk.** `OnNewSave` fires when a server starts with no save — the panel rigs wipe
   through the egg's own `REMOVE_FILES`, so this is the rig's own mechanism rather than a special
   test. What it proves is the second half of the acceptance criterion: that the old wipe's rows are
   still queryable by `?wipe=` afterwards.
3. **The Carbon socket leg.** Everything up to the socket is proven on Carbon; what is not is frames
   actually arriving over a live link, which is one elevated firewall command away
   (`Remove-NetFirewallRule -DisplayName "rust-link-sidecar.exe"`, then an allow for the rig).

Until 1 and 2 are run, the honest statement of this phase is: **the transport, the envelope, the
boards and the classification are proven on both frameworks; the player half of the catalogue is
built, reviewed against the hook documentation, and unmeasured.** That is written here rather than
in a commit message because it is the kind of thing a later phase will want to know it inherited.

### 16.7b The ingest walk, and the trap it walked into

The module half was exercised against the live rig rather than only against its own tests: the
working tree installed into a running core, pointed at the workstation sidecar with a real Rust
server behind it. It logged `cursor started at the feed tail {at: 1}` — the fresh-install path,
starting at the end rather than replaying — and a console `banid`/`unban` pair then arrived as two
`rust_events` rows carrying the wipe id, with the cursor advancing to 3 and `events_seen` at 2.

Then the boundary, on real rows rather than fixtures: with both ban events sitting in the table, the
public events route answered `{"events":[]}`, **and answered the same when asked for
`player.banned` by name**. That is the difference between a filter and a refusal — the kinds are, as
far as a public caller is concerned, not there.

**The trap, and it is the same one the Pterodactyl work recorded.** The server-state row read as
`online=1` with `hostname=NULL` and `protocol=NULL`, which is a shape no code path writes. The cause
was two cores sharing one database: a second instance left running from earlier in the day still
holds the **phase 2 release** in memory, speaks protocol 1, is refused `409` by a protocol 2 sidecar,
and writes the row back as unreachable every thirty seconds. Nothing was wrong with either of them.

What made it *look* like a defect was the measurement: two `SELECT`s in two round trips, assuming
the state did not move between them. A single atomic read caught the coherent row a moment later —
`online: 1, protocol: 2, hostname: Test Server` — alternating with the other writer's.
**A differential diagnosis across two calls is only valid if nothing else holds the controls**, and
on a shared rig something usually does.

### 16.8 Smaller things worth keeping

- **`rg.hooks` collides with `RGProbe`**, the phase-0 rig plugin, which registered the same console
  command first. Oxide warns and the last loaded wins, which happens to be the bridge. Left alone:
  `RGProbe` is rig scaffolding that never ships, and renaming the shipping command to avoid a
  test tool would be the wrong way round.
- **The IP a console ban reports is the literal string `"0"`**, not an address and not a null, when
  the banned id is offline. Observed, not guessed. The plugin omits the field instead of forwarding
  it — a column full of `"0"` survives every is-it-missing test a reader writes and then fails
  whatever parses it.
- **`--print-config` reports the *effective* configuration and writes the *file* one.** Environment
  variables override the file (R22 depends on that), and the written file never contains them. Not a
  bug, but the two are not the same document and an installer reading one should not assume the
  other.
- **A `cargo clippy` run does not produce a binary.** Two rig readings disagreed with the source
  because the sidecar under test was an older `cargo build`; `clippy` and `test` compile without
  writing one. Rebuild before believing a rig.

## 17. Phase 4 as built — the first pages, 2026-09-16

One repository, and the first phase whose whole deliverable is something a visitor looks at. It
consumes exactly the routes phase 3 built and adds one of its own; nothing here talks to a game
server, which is the point of the criterion it was written against: **the site renders the last
thing each server said while every server is off.**

**Status: criterion met, and met the hard way.** The pages were walked in a browser against a live
rig — a real Rust server behind the real sidecar for `main`, and a second, deliberately unreachable
server for everything the record holds. That walk found four defects, two of them in code phase 3
had already shipped, and all four are fixed here (§17.2).

### 17.0 The four decisions this phase needed

- **D12 — `/rust` is the server list.** Phase 1 registered it at `/rust/servers` and left the
  module's own namespace root answering core's CMS catch-all. R8 calls the list "the landing page",
  so it is registered with an **empty path** — core renders that as `/rust` — and the detail page
  hangs beneath it at `/rust/servers/:id`. One canonical address, and it is the module's name.
- **D13 — one page with in-page tabs**, not four routes. `/rust/servers/:id` carries a header and
  four panels (feed, leaderboard, online, wipes). A tab strip is not in the shared UI kit, so the
  module bundles its own — which is the kit working as designed rather than a gap in it (§3.4 is a
  closed list of nine members, and everything above them is the module's).
- **D14 — poll while the tab is visible.** The feed and the presence list re-fetch every twenty
  seconds, paused by the Page Visibility API and refreshed the instant a viewer comes back. The
  leaderboard and the wipe list load once: a table that re-sorts itself under the reader's cursor is
  worse than one four minutes old. Server-Sent Events were considered and are not this phase — core
  has a fan-out, but the module registers no stream, and a cursor-driven one is phase scope.
- **D15 — the footer slot carries a live count**, not a static link: `2 servers · 42 online`, linking
  to `/rust`. The cost is stated rather than assumed — see §17.4.

### 17.1 What is on the pages

`/rust` is the list: name, map, size, when it was wiped, when it last reported, and the player count
or `Offline`. The whole row is the link.

`/rust/servers/:id` is the server. The header is what a Rust player asks first — map, world size,
seed, wipe date — with live status beside it and a wipe selector that applies to the whole page.
Then four tabs:

| Tab | Reads | Refresh |
|---|---|---|
| Feed | `…/events`, with a filter that maps to the `kind` parameter | 20s, visibility-gated |
| Leaderboard | `…/leaderboard`, sortable, per wipe or all-time | on mount |
| Online | `…/online` — the presence **board**, not counted transitions | 20s, visibility-gated |
| Wipes | `…/wipes`; picking one filters the feed | on mount |

**Everything selectable is in the URL** — tab, filter, wipe, sort. That costs a little ceremony in
the page and buys the thing a community site is for: *"last wipe's leaderboard on Main"* is a link,
the back button undoes a click rather than leaving the page, and a refresh lands where the reader
was. `wipe=current` is a word rather than an id on purpose, so a shared link stays about now.

**The feed renders parts, not sentences.** `lib/feed.js` turns a stored frame into
`{tone, actor, join, verb, subject, detail}`, which keeps the names emphasised without any HTML in a
string, and makes the whole thing testable in a runner with no DOM. A death is four sentences, not
one — `player`, `self`, `npc`, `environment` — because the plugin distinguishes them so a reader does
not have to guess, and a fall reported as a kill by nobody is the failure that avoids. **A kind this
build has never heard of renders as itself** rather than vanishing: the server's allowlist has
already decided the row may be seen, so what is left here is presentation, and the honest
presentation of a kind we have no words for is its own name.

**`player.tally` is public and deliberately not in the feed.** It is an aggregate the plugin flushes
once a minute per active player (PROTOCOL.md §8.6), so a feed carrying it would be mostly wood
counts. It is the leaderboard's input, and that is where it shows up.

### 17.2 Four defects the page walk found, and two of them were already shipped

The walk was the whole point of doing one. None of these is visible in a test that stubs a sidecar
which answers.

1. **An unreachable refresh erased what the server last said.** Phase 3's refresh loop called
   `putState` — the whole-row write — with two fields when a sidecar did not answer, so `hostname`,
   `level`, `seed`, `world_size` and `wipe_id` all went to NULL the first time a game host rebooted.
   The list then read `Offline` with nothing beside it, which is not *"here is what we know about a
   server that is down"*, it is *"we have never heard of it"* — and it defeats this phase's
   criterion exactly. Fixed with `markUnreachable`, which moves three columns and mentions no
   others; `server/test/refresh.test.js` asserts against the SQL, because the defect is about which
   columns a statement names.
2. **"Last reported" was reading the wrong timestamp.** `updated_at` is when this module last WROTE
   the row, which a failed poll does too — so an offline server claimed it had reported just now,
   every thirty seconds, for as long as it stayed down. They are two facts and both are wanted:
   `updated_at` decides staleness, and a new `last_seen_at` records when a `server.hello` last
   arrived. Only a successful refresh moves it.
3. **Every feed row showed a bare time of day.** Correct for today's killfeed and wrong the moment
   the feed is filtered to a past wipe: three events from six weeks ago all rendered as `02:03 PM`.
   Rows from another calendar day now carry the date. The boundary is the calendar rather than a
   duration, because that is what a reader means by "what time was that".
4. **A mistyped address was dressed as a fault.** The detail page rendered core's `ErrorState` under
   its own heading, so `/rust/servers/typo` read "No such server / Something went wrong" and sent a
   reader looking for an outage. A 404 is now its own answer and `ErrorState` is kept for a request
   that failed for a reason nobody can see.

A fifth, smaller: the Online tab listed three players under a header reading `Offline`. An
unreachable sidecar does not clear the presence board — deliberately, the rows are still the best
answer anybody has — but presented bare they read as *who is on right now*, which is the one thing
an offline server cannot be saying. The panel now says which it is.

### 17.3 `useAsync` cannot poll, and that is not a defect in it

Core's fetch hook (§3.4) blanks `data` and sets `loading` on every dependency change. That is right
for a page load and wrong for a poll: bumping a dependency every twenty seconds would clear the
killfeed, render a spinner in its place and re-fill it, four times a minute, for ever.

So the module bundles `hooks/usePolled.js`: a refresh that is **invisible when it succeeds** and
keeps the rows *and* reports the error when it fails — because a site whose premise is "it renders
while the game is off" must not blank itself the first time a request does. `key` (the question)
resets the data; the interval does not. `useAsync` is still the right hook for everything that loads
once, and both are used here.

The live proof: with the tab hidden the log shows no requests at all, and the instant it became
visible there was one refresh followed by one every 20.0 seconds.

### 17.4 The footer slot's real cost, stated

Core renders `SiteFooter` inside `PublicLayout`, and **every public page renders `PublicLayout`
itself** (§3.3) — so a component in that slot mounts once per public page view, not once per
session. D15's live count therefore puts one `/public/rust/servers` request on every public page of
the site, including pages with nothing to do with Rust.

Two things keep that honest rather than merely cheap. It **renders nothing until it has an answer,
and nothing at all if the request fails** — core's `<Slot wrap>` takes its separator with it, so a
failure degrades to exactly the footer an instance with no module installed has. And it **never
polls**: one request per page view is a cost; a timer in the footer of every page is a different
kind of thing. If it ever shows up in an operator's logs, the fix is a short-lived cache in that one
file and nothing else on the site changes.

### 17.5 Smaller things worth keeping

- **The registration fake was *nearly* core, which is worse than obviously not.** `client/test`'s
  fake registry prefixed routes as `` `${id}/${path}` ``; core strips the trailing separator too,
  which is exactly what lets a module register `path: ''` and own its namespace root. The day a
  module did, the fake produced `rust/` where a real core produces `rust`, and the suite failed the
  nav check for a link that works perfectly in a browser. The fake now copies core's line character
  for character.
- **A SQL comment inside a JS template literal may not contain a backtick.** Obvious written down,
  invisible while writing prose about `markUnreachable` inside a query string; the file stops
  parsing several lines later and the error names an argument list.
- **The detail route exists so that a page can 404.** Every other route under `/servers/:id` answers
  an empty list for an id nobody configured — an unknown server genuinely has no events, no
  leaderboard and nobody online, and each of those is a good answer to its own question. Only
  `GET …/servers/:id` can say the server is not there. A disabled server answers the same 404 as a
  missing one: an operator who switched a server off did not switch it into a 403.
- **`capabilities` grew to what the pages serve** — `servers`, `killfeed`, `leaderboard`,
  `presence`, `wipes` — which is the list phase 5's Android leg feature-detects against.
- **The rig had two cores again**, and this time the mechanism was visible rather than inferred: a
  leftover core holding an older module release spoke protocol 1, was refused `409`, and rewrote the
  state row as unreachable every thirty seconds — the phase-3 trap, and also how defect 1 above was
  found. One of the two was stopped with the org lead's say-so before the walk continued.

### 17.6 What was proven, and how

- **The criterion, directly.** A second server was configured pointing at a dead address and seeded
  with a fixture shaped exactly as the plugin emits (two wipes, 26 events, four players, a presence
  board). With that server unreachable and reporting `Offline`, its page still rendered its map,
  size, seed, wipe date, killfeed, per-wipe and all-time leaderboards, last known board and wipe
  history. That is the phase criterion in one screenshot.
- **The allowlist, on real rows.** The fixture includes a kind this build has never heard of. The
  public route answered `{"events":[]}` for it when asked **by name** — a refusal, not a filter.
- **R12's arithmetic, on the page.** All-time equals the two wipes summed (41 + 18 = 59), and a
  player who only appears in the older wipe drops out of the current one rather than reading zero.
- **The footer slot**, live in core's own footer on every public page, linking to `/rust`.
- **The chunk's identity check**, which is the one failure only a browser can show: the console
  carried the module's own registration line and nothing else — no second React, no bare import.

What is **not** proven here and is deliberately left: the pages have not been read on a phone-width
viewport, and the Android leg (phase 5) is where the same surface gets a second client anyway.

## 18. Phase 5 as built — Android leg A, 2026-09-17

The first leg of R10, and the first time this module's surface has had a second client. The app's
own record of it is [`../../android/PLAN.md`](../../android/PLAN.md) **M14**; what follows is what
the phase decided and what the walk found.

**Criterion met, and walked in an emulator rather than asserted:** *the app renders a Rust site it
has never seen, and a UO site unchanged.* Both halves were shown on one device against two running
cores.

### 18.0 Four org-lead decisions

- **D16 — the app gates on a new capability, `rust`.** This module declared five strings and every
  one names a **surface**: `servers`, `killfeed`, `leaderboard`, `presence`, `wipes`. Core flattens
  every started module's capabilities into one list, so a client gating a whole navigation group on
  `servers` would have those screens revealed by any future module that declared the same generic
  word. `module-uo` has exactly one string for this job — `shard` — and this module had none.

  **Gating on the module `id` was considered and rejected.** It is the strongest fact available and
  it is already on the wire, but `id` is a **mount prefix** (§2.1 requires it to equal the directory
  core loads the module from) and `MODULE_API.md` §2.9 forbids a client inferring a route from a
  capability. Letting a client gate on `id` makes the two the same value in practice, and the day one
  builds `/<id>/servers` from it the separation that lets this module move its own pages is gone.

  So `module.json` declares its own name as a sixth capability, and `server/test/entry.test.js`
  asserts it **against `manifest.id`** rather than against the literal `"rust"` — the day the id
  changes, the string a client gates on has to change with it. Module-Rust#5.

- **D17 — the app polls every 20 seconds while its screen is RESUMED.** D14's Page Visibility gate,
  translated. `repeatOnLifecycle(RESUMED)` gives the same three behaviours from one line: nothing at
  all while the app is away, an immediate refresh on return, and a pause behind a dialog. `STARTED`
  was rejected — it keeps polling behind a partially obscured screen, which is precisely the reader
  who is not reading.

- **D18 — the three Rust repositories move to `edge`** for the rest of the workstream, with releases
  cut at the cutover rather than per phase. `edge` branches were created from `main` in Module-Rust,
  Rust-Link and Rust-Plugins; `pr-checks.yml` in all three (and in Android-app) already triggers on
  `[main, edge]`, so this costs no CI. `release.yml` still fires only on a push to `main`, which is
  what makes the cutover the release.

- **D19 — the drawer row carries a live player count, and `NavPaths` learns `/rust`.** The second is
  small and load-bearing: without it an admin's nav override on the module's own `Servers` row, or an
  added link to `/rust`, hands off to a browser rather than opening the native screen.

### 18.1 D15 has no analogue on a phone, so it was translated

The footer slot works on the web because every public page renders the same footer (§17.4). The app
has no footer and no slot. What it has is one drawer row per surface and, since engagement Phase 8, a
precedent for a number beside one — the inbox's unread badge, in `NavigationDrawerItem`'s badge slot,
with a `contentDescription` so a screen reader says *"42 players online"* and not *"42"*.

The count rides there and keeps all three of the website version's rules: **zero renders nothing** (an
empty fleet is not a notification, and a badge reading `0` on a quiet evening is worse than none), a
failed read keeps the last number rather than dropping to zero, and it **never polls**. It is asked
for only where the module is installed, so a UO site makes no request at all.

**The number is players, not servers.** A badge is one integer, and of D15's two halves the live one
is how many people are on — a server count changes when an operator edits configuration, which is not
news, and is on the page the row opens anyway.

### 18.2 What the app had to grow: a refresh that is not a load

The app has had exactly one shape for a read since its first milestone — set `Loading`, ask, replace.
That is right for opening a screen and wrong for a poll, and it is the *same* wall this module hit one
tier along with core's `useAsync` (§17.3). A twenty-second refresh built on it would clear the
killfeed, render a spinner in its place and re-fill it, three times a minute, for ever.

`ui/Polling.kt` is `usePolled`'s other half, and it keeps the same rule: **a refresh is invisible when
it succeeds and keeps the rows when it fails.** Three cases, and the middle one is the whole point:

| Result | What the reader sees |
| --- | --- |
| It answered | New rows. Nothing else. |
| It failed, and there are rows | The same rows, and one quiet line saying the refresh failed. |
| It failed, and there is nothing yet | An ordinary error with a retry — a first load that failed. |

Only the **visible** live panel is polled. The website can afford to mount the one tab it is
rendering; the app's four tabs are one screen, so the refresh asks what the reader is actually looking
at. The leaderboard and the wipe list are never polled at all.

**Changing the question is not a poll.** Filter, sort and wipe blank their panel and load, because
what is on screen is an answer to something the reader has stopped asking — leaving it up would show
last wipe's killfeed under this wipe's heading.

### 18.3 What the walk proved, and how

The rig is phase 4's, unchanged: a core on `:3200` with this module installed, one live server
(`main`, a real sidecar and a real game host) and the seeded `demo` fixture that has never reported.
A second core on `:3100` serves `module-uo` and no Rust.

- **The criterion, first half.** With `demo` unreachable and reading *Offline*, the phone rendered its
  map, size, seed, wipe date, killfeed, per-wipe and all-time leaderboards, its last known presence
  board and its wipe history. Nothing on the screen is a live call to a game host.
- **The criterion, second half.** The same app, switched to the UO core, showed Shard / Rules / Atlas
  / Leaderboards / Market and **no Rust row**.
- **`refreshInto` against a genuinely dead backend.** The core was stopped with the list on screen. A
  poll tick later the rows were unchanged, under one line reading *"Could not refresh just now. This
  is the last thing the site heard."* — no spinner, no error page, nothing blanked.
- **R12's arithmetic, on a phone.** All-time 59 = 41 + 18 across two wipes, and Drift — who appears
  only in the current wipe — **drops out** of the August board rather than reading zero.
- **Every `describe` branch, from real rows**: a player kill with weapon, distance and grid; an NPC
  kill with the prefab read as words; a suicide; an environment death (the fall that must not read as
  a kill by nobody); chat with its colon in the join and a non-Global channel beside it; a disconnect
  with reason and session length; and *while sleeping*.
- **The calendar-day rule.** Filtering to the August wipe produced three rows six weeks old, each
  carrying its date — the §17.2 defect, not re-introduced in Kotlin.
- **The presence panel saying which it is.** The offline server's board rendered under *"The last
  board this server sent. It is offline, so this is who was on then — not who is on now."*
- **The badge**, showing a live count on the drawer row.

### 18.4 The walk found three defects, and a green suite found none of them

1. **The drawer's live count resolved once per process.** It was keyed on the capability answer alone,
   so it was read when the app connected and never again — which is not what *live* means on a row
   somebody opens the drawer to look at. It now refreshes on resume, beside the unread badge and for
   the same reason: coming back to the app is exactly when a stale number would be noticed. *Visible
   only by backgrounding the app and returning to it.*
2. **Every card's text sat flush against its edge.** The app's themed `ShardCard` is a `Card` and
   nothing more — it carries no padding, and each caller pads its own content. Four new call sites did
   not, and on a phone the first glyph of each line read as clipped.
3. **A name touched its own kill count.** Five numeric columns beside an equal-weight name column left
   *Brannock* and *50* reading as one field. The name now takes a wider share and ellipsizes — and the
   **active sort moved to the header**, because the header is the control: tinting a column of numbers
   says *these are special* where tinting the header says *this is what the table is ordered by*.

### 18.5 The rig note worth keeping

The app's debug `network_security_config.xml` permits cleartext to **`127.0.0.1` and `localhost`
only** — not `10.0.2.2`. An emulator walk against a local core therefore needs
`adb reverse tcp:<port> tcp:<port>` and the loopback address. Typed as `10.0.2.2`, every request fails
with `UnknownServiceException: CLEARTEXT communication to 10.0.2.2 not permitted`, and the connect
screen reports *"Couldn't reach that site"* — correct, and indistinguishable from a core that is not
running.

Two smaller things: the first AVD tried had 95% of `/data` used and refused a 44 MB install with
*"Requested internal only, but not enough space"* — `pm trim-caches` freed nothing, and the second AVD
was the answer. And the app's own `pm clear` is the way to reach the first-run connect screen, because
an `install -r` over an earlier install keeps the stored base URL.

### 18.6 What is not proven here

- **A phone-width read of the website's own pages**, which §17.6 deliberately left open. This phase
  gave the surface a second client rather than re-reading the first, and the pages have still not been
  looked at in a narrow browser.
- **The badge's non-zero case on real traffic.** Nobody was playing on the rig, so the count was shown
  by seeding a player count on the unreachable server and reading the badge before the 30-second
  refresh zeroed it. The arithmetic and the rendering are proven; a fleet with people on it is not.

---

[rl]: https://gitea.whitlocktech.com/RunicGateway/Rust-Link
[rp]: https://gitea.whitlocktech.com/RunicGateway/Rust-Plugins
[mr]: https://gitea.whitlocktech.com/RunicGateway/Module-Rust

[kit]: https://gitea.whitlocktech.com/RunicGateway/Integration-kit