docs/link/v4.md

# Protocol 4 — Guild membership on the wire

**Status:** Feature-complete on `edge`. All work lands on an `edge` branch in each repo; `edge` → `main` is the v4 cutover.
**Date:** 2026-08-17
**Codebase:** ServUO 57.4, `<servuo>`, net48 / x64, Expansion **EJ**.
**Companion to** [`PLAN.md`](PLAN.md) (1.0 read/event plane), [`PROTOCOL_2.md`](PROTOCOL_2.md) (2.0 provisioning + world-state streams), [`v3.md`](v3.md) (3.0 shard content + the visibility framework), [`INTEGRATION.md`](INTEGRATION.md) (website API).

---

## 1. Why 4

Protocol 2 gave the website a guild board: name, abbreviation, leader, alliance, and a member
**count**. It could say a guild had 155 members. It could not say who they were, and there was no
event for anyone leaving one.

That gap was known when 2.0 shipped. [`PROTOCOL_2.md` §10.1](PROTOCOL_2.md) sketched exactly this
design — hold a member-serial set, diff it each sweep, emit `guild.join`/`guild.leave` — and then
2.0 shipped only the half that needed no new state, folding membership into the board signature as a
serial *sum*. Protocol 4 builds the rest of what §10.1 described.

The immediate consumer is a richer public Guilds page, which is worth the bump on its own merit. The
reason it is being built **now** is that [`../website/TEAMS.md`](../website/TEAMS.md) needs it:
platform Teams are sourced from game guilds, and nothing in Team core can be built against a count.
That is a dependency, not a coupling — nothing in this protocol version knows what a Team is.

**Two kinds, both additive:**

| Kind | Shape | Purpose |
|---|---|---|
| `guild.roster` | full member list, possibly split across frames | the membership itself; supersedes whatever was held |
| `guild.leave` | one departing member | the real-time counterpart to the existing `guild.join` |

Nothing existing changed shape. `GET /guilds` grows a `roster` key; a v3 consumer that ignores it
keeps working.

---

## 2. The shard side

`BridgeSocial.cs` holds each guild's member serial **set** instead of folding it into the signature
as a sum. Two consequences, both the point of the change:

- **A set comparison cannot collide.** The old sum could: one member joining and another leaving
  between two sweeps offset each other, and the guild read as unchanged.
- **A set can be differenced.** Departures are the prior set minus the current one — which is what
  makes a per-member `guild.leave` possible without a core tap, since ServUO raises no event for
  leaving, disbanding, or a leader change.

A changed set re-emits `guild.roster`, the whole member list. That is what lets `guild.leave` stay
**advisory**: a consumer building a "so-and-so left" feed wants the individual events, but a consumer
holding a membership table only needs the roster, so nothing downstream has to replay deltas to
remain correct. A dropped `guild.leave` costs latency, never accuracy.

On a guild's **first** sweep there is no prior set, so nothing is reported as leaving. An unknown
roster becoming known is not 155 people leaving at once.

### 2.1 Frames are capped, and split when they must be

A roster is the only fat frame this bridge emits. Measured against a real 155-member guild: **10,812
bytes, about 69 bytes per member.** The sidecar's `read_line` has no length bound, so an uncapped
roster is an unbounded line.

Members per frame are therefore capped (`Bridge.GuildRosterMembersPerLine`, default 500 ≈ 35 KB), and
a guild over the cap is split into frames carrying `seq`, `more` and `total`:

```jsonc
{"t":1786988708513,"kind":"guild.roster","id":1,"name":"The Silver Hand","abbr":"TSH",
 "total":155,"seq":0,"more":true,"members":[ /* 50 actor objects */ ]}
{"t":1786988708514,"kind":"guild.roster","id":1,"name":"The Silver Hand","abbr":"TSH",
 "total":155,"seq":1,"more":true,"members":[ /* 50 */ ]}
{"t":1786988708514,"kind":"guild.roster","id":1,"name":"The Silver Hand","abbr":"TSH",
 "total":155,"seq":2,"more":true,"members":[ /* 50 */ ]}
{"t":1786988708514,"kind":"guild.roster","id":1,"name":"The Silver Hand","abbr":"TSH",
 "total":155,"seq":3,"more":false,"members":[ /* 5 */ ]}
```

Reading the flags: **`seq` 0 begins a roster and supersedes whatever was held for that guild**;
`more: false` ends it. Every realistic guild is inside the cap and emits exactly one frame with
`seq: 0, more: false` — the same shape as if chunking did not exist. A guild with no members still
emits one frame with an empty array, or a consumer could never learn that a roster it holds has
emptied.

`guild.leave` is a small frame naming the departed serial:

```jsonc
{"t":1786988770099,"kind":"guild.leave","id":1,"name":"The Silver Hand","who":"0x1F5"}
```

### 2.2 The reconnect baseline is spread

`BridgeLink.OnConnected` clears the diff caches, so after a reconnect **every** guild reads as
changed at once. The outbound queue is not the risk — the cap is 10,000 *lines* and a few hundred
guilds is a few hundred lines — but building hundreds of fat JSON frames in a single Core-thread tick
is exactly the stall this bridge exists to avoid.

So at most `Bridge.GuildRosterGuildsPerTick` guilds (default 25) emit a roster per sweep. A guild over
budget keeps its previously held member set, so it still reads as changed on the next pass; its
`guild.update` has already gone, so the board's counts are current either way. While a baseline is
draining the sweep re-arms itself after 2 s rather than waiting a full `GuildSweepSeconds`, so
catch-up takes seconds instead of one sweep interval per batch.

### 2.3 What a roster member carries

Each entry is the standard actor object — `serial`, `name`, `player`, plus `acct` when the mobile has
an account and `webId` when that account is linked:

```jsonc
{"serial":"0x1F5","name":"Seed000A","acct":"seed_000","webId":"42","player":true}
```

`acct` is **genuinely optional**: a `PlayerMobile` can have no `Account` at all, and the local test
world contains such mobiles. Consumers must not assume it is present.

These identity fields are emitted unconditionally, by design — the sidecar is a forwarder, and
deciding who may see them is the website's job. See §4.

---

## 3. The sidecar side

`guild.roster` writes a **`members` column on the `guilds` board**, not a field inside the existing
`json` column. That column holds the verbatim `guild.update` line, and a roster written into it would
clobber the snapshot — name, abbreviation, leader, online count — that `guild.update` owns. Two
writers across two columns of one row keeps both as plain upserts: neither reads the other's value
first, so there is no read-modify-write and no ordering requirement between the two kinds. Either can
arrive first.

`GET /guilds` folds the roster back in as `roster` at read time. A guild that has had a
`guild.update` but no `guild.roster` yet simply has **no `roster` key** — the honest representation
of "not known", and deliberately distinct from a guild whose roster is genuinely empty.

`guild.leave` gets **no board projection at all**. The sidecar persists and broadcasts it like any
event and leaves the board alone; the roster self-corrects on the next `guild.roster`, which the
shard re-emits whenever the member set changes. Keeping the delta out of the projection is what keeps
the sidecar a forwarder rather than a thing that maintains state.

### 3.1 Split rosters are reassembled in memory

A roster split across frames is reassembled **before** it is stored, and written once on the frame
that closes it. Appending to the column per frame was rejected twice over: it would make the write a
read-modify-write — the exact thing the two-column split exists to avoid — and it would publish a
torn roster, since a reader hitting `GET /guilds` between frames would see a partial member list
presented as the whole truth.

This is transport-level reassembly, the same category of work as turning bytes into a line, and it
holds nothing once a roster is complete. The ordinary single-frame case never enters the buffer at
all. Around it: a fresh `seq: 0` supersedes an abandoned partial, an out-of-order frame discards the
partial rather than storing one with an undetectable hole, a continuation with no start is ignored, a
`server.hello` drops every partial (a reconnected shard restarts each roster at 0), and accumulation
is bounded so a shard that never sends a closing frame cannot grow the buffer without limit.

### 3.2 This is the first bump that needed a store migration

`store.rs`'s `SCHEMA` is `CREATE TABLE IF NOT EXISTS`, which can add a table but **cannot add a
column to a table that already exists**. Every schema change up to and including Protocol 3.0 added
whole tables, so this never mattered and `ALTER TABLE` appeared nowhere in the repo's history.
`guilds.members` is the first column added to an existing table: without a mechanism it would simply
never reach an installed sidecar, and every roster write would fail.

The counter is SQLite's own **`PRAGMA user_version`** — an integer in the database header, so it
costs no table and cannot drift from the file it describes. Each step runs in a transaction together
with the bump recording it, so a step lands completely or not at all and an interrupted run resumes
in the right place.

- **A failure aborts startup**, which was already the behaviour. A half-migrated store answers the
  website with confusing partial data, which is worse than being plainly absent — and the shard dials
  *out*, so a sidecar that refuses to start never stalls the game.
- **A database written by a newer sidecar warns and continues.** Every step is additive, so a newer
  schema has only columns an older reader ignores; refusing to start would turn rolling the binary
  back — a recovery path — into a dead end.

`sqlx::migrate!` was considered and rejected: it checksums each migration file, so editing an
already-released migration hard-fails startup, which is a poor trade for a schema of eleven
JSON-blob tables.

---

## 4. Visibility

Both kinds map to the existing **`guilds`** feature in the website's visibility framework
([`v3.md` §3](v3.md)). That mapping is required, not cosmetic: rule 2 fails an *unmapped* kind closed
to admin-only, which would have quietly kept rosters off the public Guilds page forever.

Mapping them is safe because of how the framework strips fields. A roster is the first frame carrying
locked fields inside an **array** of actors rather than a single nested actor — but the projection
walker already recurses into arrays and matches `acct`/`webId` **by suffix, on meaning rather than
spelling**. So a member's account name is stripped below `admin` by exactly the rule that already
strips `guild.leader.acct`.

The website stores `acct`/`web_id` per member, because that is what lets a linked member be matched to
a site user at all. It never projects them below `admin`. The difference this rule makes is a public
page listing character names versus one publishing 150 account names, so it carries a test of its own.

---

## 5. Cross-repo obligations

| Repo | Change |
|---|---|
| `servuo-plugins` | member-set diff, `guild.roster` + `guild.leave`, `BridgeJson.Actors`, two `Bridge.cfg` keys, **`overlay.toml` protocol → 4** |
| `link` | `PROTOCOL_VERSION` → 4, `members` column + `user_version` migration, roster reassembly, `GET /guilds` projection |
| `module-uo` | `shard_guild_members`, `guild.roster`/`guild.leave` ingest, the kind→feature map entries |
| `installer` | `backup.rs`'s stated reason for skipping the sidecar DB (§3.2 falsifies it) — docs only |
| `docs` | this file, `PROTOCOL_2.md` §10.1, `INTEGRATION.md` |

**`overlay.toml` must move in the same PR as the emitters.** CI folds it into the release manifest and
the installer refuses to pair an overlay and a sidecar whose protocol numbers disagree, so a bump
landing separately would silently fail to compose into a bundle.

---

## 6. Verification

`servuo-plugins` has no CI build — the plugin compiles only inside ServUO — so "it compiled" is not
evidence and neither is a clean boot. What was actually run:

1. **A throwaway one-guild spike first** (TEAMS.md Phase 0), against the real Rust sidecar rather than
   a stub, to retire the unknowns before committing to a four-repo bump. It proved the two-column
   board shape, measured the line, and surfaced §3.2 — the missing migration mechanism — which no
   amount of reading would have.
2. **A live end-to-end run**: 155 members seeded from real `PlayerMobile`s on the local ServUO tree,
   the cap forced down to 50 so the split path actually fired, producing 50/50/50/5 across four
   frames; two members then removed on a timer, producing exactly two `guild.leave` frames with the
   correct serials and a re-emitted roster at `total: 153`.
3. **A restart test with the shard stopped first**, so its reconnect could not re-emit and fake
   persistence. The board's row came back byte-identical.

Step 2 is what caught the reassembly bug in §3.1: every unit test passed through it, because they all
exercised a single-frame roster. The case does not arise until a guild exceeds the cap.

Still outstanding for the cutover: the five-rung shard visibility walk against a live shard, confirming
`acct`/`webId` never reach a caller below their rung.