Merge pull request 'feat(sidecar): forward the asset manifest, the pixels and the body pass (Phase 3)' (#43) from feat/asset-bridge-p3 into edge

Reviewed-on: #43

.gitea/workflows/pr-checks.yml

@@ -17,10 +17,12 @@
 #   so let this workflow run on one PR first. The `PR Checks / *` glob matches
 #   without needing the dropdown.
 #
-# Scope note: this gates PRs into `main` only. Feature work that lands on an
-# integration branch first (e.g. `edge`) is still caught on the branch's PR into
-# `main`. To gate that earlier hop too, add the branch to the `branches:` list
-# below — nothing else needs to change.
+# Scope note: `edge` is gated as well as `main`. Multi-phase work lands there
+# first, so gating only the `main` hop would run these checks for the first time
+# at the cutover — the one moment a red build is most expensive to discover. This
+# is the same call `RunicGateway/installer` made for the same reason, and it was
+# taken here after a nine-PR Android workstream landed on an ungated `edge` with
+# no CI at all. Adding a branch to the `branches:` list is the whole change.
 #
 # Runner: the same self-hosted `ubuntu-latest` runner release.yml uses. Rust is
 # not assumed to be preinstalled, so the toolchain step bootstraps it the same
@@ -31,7 +33,7 @@ name: PR Checks
 
 on:
   pull_request:
-    branches: [main]
+    branches: [main, edge]
 
 # A newer push to the same PR cancels the in-flight run.
 concurrency:

.gitea/workflows/release.yml

@@ -149,6 +149,39 @@ jobs:
             fi
           fi
 
+          # ── Orphan sweep ────────────────────────────────────────────────
+          #
+          # The check above is VERSION-SCOPED: it only ever asks about the one
+          # version this run computed. That is enough to recover an orphan on
+          # the very next run, and useless afterwards — once any releasable
+          # commit lands, the next run computes a NEW version, never looks at
+          # the old tag again, and the orphan becomes permanent and silent.
+          #
+          # servuo-plugins v0.1.0 is the proof, and the proof is pointed: the
+          # commit that ADDED the recovery above was itself typed
+          # `fix(release): ... recover the orphaned v0.1.0 tag`, so it bumped to
+          # v0.1.1 — and the run that introduced the recovery stepped straight
+          # past the tag it was written to rescue. That tag is still orphaned.
+          #
+          # So every v* tag is checked, and anything missing a release is
+          # WARNED about. Deliberately not recovered: publishing an old version
+          # would mean building today's tree and shipping it under a tag whose
+          # tree it is not, which is worse than the inconsistency it fixes.
+          # A human decides whether to recover or drop it.
+          #
+          # Never fails the run. A sweep that can break a good release is a
+          # sweep someone will delete.
+          ORPHANS=""
+          for T in $(git tag -l 'v*' --sort=-v:refname); do
+            T_HTTP="$(curl -s -o /dev/null -w '%{http_code}' \
+              -H "Authorization: token $(printf '%s' "${REGISTRY_TOKEN:-}" | tr -d '\r\n')" \
+              "https://${GITEA_HOST}/api/v1/repos/${REPO}/releases/tags/${T}" || echo 000)"
+            [ "$T_HTTP" = "404" ] && ORPHANS="${ORPHANS} ${T}"
+          done
+          if [ -n "${ORPHANS}" ]; then
+            echo "::warning::Tags with no release:${ORPHANS} — a run failed after tagging. Publish or delete them; this job will not do either."
+          fi
+
           # Changelog range. A recovery run has nothing after the tag, so
           # summarize what the tag itself contains rather than emitting an empty
           # list: the range that produced it, i.e. previous-tag..this-tag.
@@ -279,33 +312,43 @@ jobs:
           ls -l dist && echo "----" && cat dist/SHA256SUMS
 
       # ── RELEASE ENGINE: commit the bump, tag, push ───────────────────────
-      - name: Commit version bump and push tag
+      # Tag only — `main` is never pushed to.
+      #
+      # This step used to commit the version bump back to main first. Two things
+      # were wrong with that. It has never once executed: an EMPTY template
+      # expression written literally in a comment (the `$`+`{{ }}` token, which
+      # is why it is spelled out here) made the runner fail to build the script
+      # and skip the whole step silently, which is why sidecar/Cargo.toml still
+      # says 0.1.0 after six releases (the tags exist because the release API
+      # creates one when it publishes). And had it executed, it would have been
+      # declined — main is protected, and a release must not depend on a write
+      # to a protected branch.
+      #
+      # So the tag is the version, as it already is in servuo-plugins. The
+      # workflow still writes the real version into Cargo.toml before building,
+      # so a released binary self-reports correctly; what it no longer does is
+      # commit that edit back. The next version is computed from the newest tag,
+      # never from Cargo.toml, so nothing downstream depends on the file.
+      - name: Push the release tag
         if: ${{ steps.plan.outputs.release == 'true' }}
         env:
           REGISTRY_USER: ${{ secrets.REGISTRY_USER }}
           REGISTRY_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
         run: |
           set -euo pipefail
-          VERSION="${{ steps.plan.outputs.version }}"
           TAG="${{ steps.plan.outputs.tag }}"
           # Secrets can arrive with a trailing newline (depending on how they were
           # pasted); a stray CR/LF corrupts the remote URL ("credential url cannot
-          # be parsed"). Strip line breaks before building the URL. Passing them via
-          # env (not inline ${{ }}) also keeps a newline from breaking this script.
+          # be parsed"). Strip line breaks before building the URL. They are passed
+          # via env rather than interpolated into this script, so a newline cannot
+          # break it — do NOT write a template token literally in a comment here,
+          # or the runner will skip this step without failing the job.
           CI_USER="$(printf '%s' "${REGISTRY_USER}" | tr -d '\r\n')"
           CI_TOKEN="$(printf '%s' "${REGISTRY_TOKEN}" | tr -d '\r\n')"
           git config user.name  "uo-link-ci"
           git config user.email "ci@whitlocktech.com"
           git remote set-url origin \
             "https://${CI_USER}:${CI_TOKEN}@${GITEA_HOST}/${REPO}.git"
-
-          git add "${WORKDIR}/Cargo.toml" "${WORKDIR}/Cargo.lock"
-          if ! git diff --cached --quiet; then
-            git commit -m "chore(release): bump version to ${TAG} [skip ci]"
-            git push origin "HEAD:main"
-          else
-            echo "Version unchanged (first release) — no bump commit needed."
-          fi
           # The tag may already exist when finishing a run that died after tagging
           # (see the plan step). `git tag` on an existing name fails under
           # `set -e`; pushing an identical existing tag is a harmless no-op. A
@@ -332,18 +375,75 @@ jobs:
           # corrupt the Authorization header.
           CI_TOKEN="$(printf '%s' "${REGISTRY_TOKEN}" | tr -d '\r\n')"
 
-          REL_ID="$(curl -sSf -X POST "${API}/releases" \
-            -H "Authorization: token ${CI_TOKEN}" \
-            -H "Content-Type: application/json" \
-            -d "$(jq -n --arg tag "$TAG" --arg body "$BODY" \
-                   '{tag_name:$tag, name:$tag, body:$body, draft:false, prerelease:false}')" \
-            | jq -r '.id')"
+          PAYLOAD="$(jq -n --arg tag "$TAG" --arg body "$BODY" \
+                       '{tag_name:$tag, name:$tag, body:$body, draft:false, prerelease:false}')"
+
+          # This POST is the step that orphaned tag v0.1.1 (run 75): it landed one
+          # second after the tag push and Gitea answered 500, having not finished
+          # processing the pushed tag. Re-running the workflow published the same
+          # four assets untouched, so the failure was a race, not a bad request.
+          #
+          # Two things went wrong there, and both are fixed here.
+          #
+          # 1. `curl -sSf` prints NO response body on an error status, so all the
+          #    log carried was "curl: (22) ... error: 500" and the cause had to be
+          #    inferred from timestamps. Capture the body and print it.
+          # 2. Nothing retried, so a transient 5xx became a permanent orphan tag.
+          #    The plan step CAN recover one, but only on a run that reaches it --
+          #    and a later push with no releasable commits stands down before it
+          #    gets there, so in practice the tag sits until a human notices.
+          #
+          # 4xx is deliberately NOT retried: a bad token or a malformed body does
+          # not improve by being sent again, and retrying only turns a clear
+          # failure into a slow one.
+          REL_ID=""
+          for attempt in 1 2 3 4 5; do
+            HTTP="$(curl -s -o /tmp/rel.json -w '%{http_code}' -X POST "${API}/releases" \
+              -H "Authorization: token ${CI_TOKEN}" \
+              -H "Content-Type: application/json" \
+              -d "${PAYLOAD}" || echo 000)"
+
+            if [ "$HTTP" = "201" ] || [ "$HTTP" = "200" ]; then
+              REL_ID="$(jq -r '.id' /tmp/rel.json)"
+              break
+            fi
+
+            echo "::warning::POST /releases attempt ${attempt} returned HTTP ${HTTP}"
+            echo "--- response body ---"
+            cat /tmp/rel.json || true
+            echo
+            echo "---------------------"
+
+            case "$HTTP" in
+              4*) echo "::error::HTTP ${HTTP} is a client error - not retrying."; exit 1 ;;
+            esac
+
+            if [ "$attempt" = 5 ]; then
+              echo "::error::POST /releases still failing after 5 attempts. Tag ${TAG} is pushed but has no release."
+              echo "::error::Re-run this workflow - the plan step detects the orphan tag and republishes it."
+              exit 1
+            fi
+            sleep $(( attempt * 5 ))
+          done
+
+          if [ -z "$REL_ID" ] || [ "$REL_ID" = "null" ]; then
+            echo "::error::Release created but no id came back; refusing to upload assets blind."
+            exit 1
+          fi
           echo "Created release ${TAG} (id=${REL_ID})"
 
           for f in "${BIN}-linux-x86_64" "${BIN}-linux-aarch64" "${BIN}-windows-x86_64.exe" SHA256SUMS; do
-            curl -sSf -X POST "${API}/releases/${REL_ID}/assets?name=${f}" \
+            # Same treatment. An upload that fails quietly leaves a release whose
+            # SHA256SUMS does not cover every binary it advertises, which is worse
+            # than no release at all -- that file IS the trust anchor.
+            HTTP="$(curl -s -o /tmp/asset.json -w '%{http_code}' -X POST "${API}/releases/${REL_ID}/assets?name=${f}" \
               -H "Authorization: token ${CI_TOKEN}" \
-              -F "attachment=@dist/${f}" >/dev/null
+              -F "attachment=@dist/${f}" || echo 000)"
+            if [ "$HTTP" != "201" ] && [ "$HTTP" != "200" ]; then
+              echo "::error::uploading ${f} returned HTTP ${HTTP}"
+              cat /tmp/asset.json || true
+              exit 1
+            fi
             echo "  uploaded ${f}"
           done
 

README.md

@@ -12,11 +12,34 @@ ServUO plugin (C#, net48)  ──loopback TCP, newline-JSON──►  Rust sidec
 The shard never speaks WebSocket and exposes no port of its own — the sidecar is the only
 network-facing component, which is what keeps the game unreachable from the internet.
 
+## Running a shard? Don't build this
+
+The [**Runic Gateway installer**](https://gitea.whitlocktech.com/RunicGateway/installer) installs
+this sidecar for you — the released binary, its config, a hardened service account and the service
+registration — alongside the shard plugin, in one run, on Linux or Windows:
+
+```bash
+sudo ./runicgateway-installer-linux-x86_64 install
+```
+
+It ends by printing the base URL, WebSocket URL, protocol version and auth token to paste into
+**Admin → Shard** on your site. Guide:
+[installer/INSTALL.md](https://gitea.whitlocktech.com/RunicGateway/docs/src/branch/main/installer/INSTALL.md).
+
+Installing it yourself is supported too — the release binaries on this repo's
+[releases page](https://gitea.whitlocktech.com/RunicGateway/link/releases) are the same ones the
+installer fetches, and
+[INSTALL.md Appendix A3–A4](https://gitea.whitlocktech.com/RunicGateway/docs/src/branch/main/installer/INSTALL.md#a3-install-the-sidecar)
+covers placing the binary and registering the service by hand.
+
+Everything below this line is for **developing on the sidecar**.
+
 ## Related repos
 
 | Repo | What |
 |------|------|
 | **this** — `RunicGateway/link` | The Rust sidecar (`sidecar/`). |
+| [RunicGateway/installer](https://gitea.whitlocktech.com/RunicGateway/installer) | The **installer** — deploys this sidecar and the plugin onto a shard host. The supported way to set one up. |
 | [RunicGateway/servuo-plugins](https://gitea.whitlocktech.com/RunicGateway/servuo-plugins) | The **C# ServUO plugin** — the shard side of the bridge (`overlay/`, `patches/`, `deploy.ps1`, test scaffolding). |
 | [RunicGateway/docs](https://gitea.whitlocktech.com/RunicGateway/docs) | All project documentation — design docs, protocol spec, integration guide, research. |
 
@@ -28,9 +51,10 @@ network-facing component, which is what keeps the game unreachable from the inte
 | `.gitea/workflows/pr-checks.yml` | Gates every PR into `main` on `cargo fmt --check`, `cargo clippy -D warnings`, and `cargo test`. |
 | `.gitea/workflows/release.yml` | Builds + releases the sidecar binary (Linux + Windows) on every merge to `main`. |
 
-## Build & run
+## Build & run (development)
 
-The sidecar is a standard cargo crate:
+Building from source is for working *on* the sidecar; a deployment gets its binary from a release,
+via the installer or by hand. The sidecar is a standard cargo crate:
 
 ```bash
 cd sidecar
@@ -39,7 +63,7 @@ cp sidecar.toml.example sidecar.toml   # then edit
 cargo run --release
 ```
 
-Deploying it rather than developing on it: `--config <PATH>` names the config file (as does
+Deploying it by hand rather than developing on it: `--config <PATH>` names the config file (as does
 `$UOLINK_CONFIG`), and `--print-config` prints the resolved settings — **including the auth token
 the website needs** — as JSON, provisioning the config file on first run. That is the supported way
 to read the token back; it is not meant to be scraped from the log.

sidecar/src/app.rs

@@ -14,11 +14,13 @@
 //! - `shutdown` is whatever "stop" means on this host: Ctrl-C and `SIGTERM` on Unix, the SCM's
 //!   `Stop` control on Windows.
 
+use std::collections::HashMap;
 use std::future::Future;
 use std::sync::atomic::AtomicI64;
 use std::sync::Arc;
 use std::time::Instant;
 
+use serde_json::Value;
 use tokio::sync::{broadcast, mpsc};
 use tracing::info;
 
@@ -88,6 +90,9 @@ where
     let last_event_ts = last_event.clone();
     let replay_handle = handle.clone(); // re-push external news to the shard on (re)connect
     let mut total: u64 = 0;
+    // Partly-received guild rosters, keyed by guild id. Lives in the event-loop task, so it needs
+    // no lock and dies with the loop. See `accumulate_roster`.
+    let mut roster_parts: HashMap<i64, RosterParts> = HashMap::new();
     tokio::spawn(async move {
         while let Some(ev) = event_rx.recv().await {
             // Any line from the shard — including pong heartbeats — is a sign of life.
@@ -161,6 +166,43 @@ where
                             }
                         }
                     }
+                    // Guild roster (Protocol 4): the member list that `guild.update`'s counts cannot
+                    // express. It writes a *different column* of the same row, so it never races
+                    // guild.update. `guild.leave` deliberately has no arm here — the sidecar
+                    // forwards it (persisted and broadcast below, like any event) and the board's
+                    // roster self-corrects on the next `guild.roster`, which the shard re-emits
+                    // whenever the member set changes. Keeping the delta out of the board is what
+                    // keeps the sidecar a forwarder rather than a thing that maintains state.
+                    //
+                    // A roster over the shard's per-line cap arrives as several frames, so it is
+                    // reassembled before it is stored — see `accumulate_roster` for why that happens
+                    // here rather than by appending to the column.
+                    "guild.roster" => {
+                        if let Some(id) = ev.value.get("id").and_then(|v| v.as_i64()) {
+                            let seq = ev.value.get("seq").and_then(|v| v.as_i64()).unwrap_or(0);
+                            let more = ev
+                                .value
+                                .get("more")
+                                .and_then(|v| v.as_bool())
+                                .unwrap_or(false);
+                            let members = ev
+                                .value
+                                .get("members")
+                                .and_then(|m| m.as_array())
+                                .cloned()
+                                .unwrap_or_default();
+
+                            if let Some(complete) =
+                                accumulate_roster(&mut roster_parts, id, seq, more, members)
+                            {
+                                let json = serde_json::Value::Array(complete).to_string();
+                                if let Err(e) = event_store.upsert_guild_roster(id, &json, t).await
+                                {
+                                    tracing::warn!(error = %e, "failed to upsert guild roster");
+                                }
+                            }
+                        }
+                    }
                     "guild.remove" => {
                         if let Some(id) = ev.value.get("id").and_then(|v| v.as_i64()) {
                             if let Err(e) = event_store.delete_guild(id).await {
@@ -279,6 +321,10 @@ where
             // with announce=false so a restart does not re-proclaim every article at once. news.add
             // is idempotent by id, so replaying to a still-populated shard is harmless.
             if ev.kind == "server.hello" {
+                // A (re)connected shard restarts every roster from `seq` 0, so any half-received
+                // one belongs to the previous connection and can never be completed.
+                roster_parts.clear();
+
                 match event_store.news_all().await {
                     Ok(items) => {
                         for mut item in items {
@@ -327,3 +373,199 @@ pub fn now_ms() -> i64 {
         .map(|d| d.as_millis() as i64)
         .unwrap_or(0)
 }
+
+/// A guild roster that has arrived in part: the `seq` expected next, and what has accumulated.
+struct RosterParts {
+    next_seq: i64,
+    members: Vec<Value>,
+}
+
+/// Refuses to accumulate a roster past this many members. The shard caps its own frames, so
+/// exceeding this means a shard that is buggy or not what it claims to be — and the one thing this
+/// buffer must not do is grow without bound on its say-so.
+const MAX_ROSTER_MEMBERS: usize = 50_000;
+
+/// Reassembles a `guild.roster` that the shard split across frames, returning the whole member list
+/// once the final frame arrives and `None` while one is still incomplete.
+///
+/// Reassembly happens **here, in memory, before the store** rather than by appending to the
+/// `members` column, for two reasons. Appending would make the write a read-modify-write — the exact
+/// thing the two-column board design exists to avoid — and it would publish a torn roster: a reader
+/// hitting `GET /guilds` between frames would see a partial member list as though it were the truth.
+/// Buffering keeps the store's write a single atomic upsert of a complete roster.
+///
+/// This is transport-level reassembly, not domain state: it is the same category of work as turning
+/// bytes into a line, and it holds nothing once a roster is complete. That is what keeps it
+/// compatible with the sidecar being a forwarder.
+///
+/// The ordinary case — a guild inside the shard's per-line cap, which is every realistic one —
+/// arrives as `seq` 0 with `more` false and is returned immediately without ever touching the map.
+fn accumulate_roster(
+    parts: &mut HashMap<i64, RosterParts>,
+    id: i64,
+    seq: i64,
+    more: bool,
+    members: Vec<Value>,
+) -> Option<Vec<Value>> {
+    if seq == 0 {
+        // A fresh roster supersedes any partial one: the shard restarts at 0 every time it emits,
+        // so a leftover buffer is from an emission that was interrupted and will never finish.
+        parts.remove(&id);
+
+        if !more {
+            return Some(members);
+        }
+
+        parts.insert(
+            id,
+            RosterParts {
+                next_seq: 1,
+                members,
+            },
+        );
+        return None;
+    }
+
+    let entry = match parts.get_mut(&id) {
+        Some(entry) => entry,
+        // A continuation with nothing to continue: the sidecar started, or the shard reconnected,
+        // midway through an emission. Dropping it is right — the next full roster is complete.
+        None => {
+            tracing::debug!(
+                guild = id,
+                seq,
+                "roster continuation with no start; ignoring"
+            );
+            return None;
+        }
+    };
+
+    if entry.next_seq != seq {
+        tracing::warn!(
+            guild = id,
+            expected = entry.next_seq,
+            got = seq,
+            "roster frames out of order; discarding the partial roster"
+        );
+        parts.remove(&id);
+        return None;
+    }
+
+    entry.members.extend(members);
+
+    if entry.members.len() > MAX_ROSTER_MEMBERS {
+        tracing::warn!(
+            guild = id,
+            len = entry.members.len(),
+            "roster exceeded the reassembly cap; discarding"
+        );
+        parts.remove(&id);
+        return None;
+    }
+
+    if more {
+        entry.next_seq = seq + 1;
+        return None;
+    }
+
+    parts.remove(&id).map(|done| done.members)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn members(names: &[&str]) -> Vec<Value> {
+        names
+            .iter()
+            .map(|n| serde_json::json!({"name": n}))
+            .collect()
+    }
+
+    fn names(vs: &[Value]) -> Vec<String> {
+        vs.iter()
+            .map(|v| v["name"].as_str().unwrap_or_default().to_string())
+            .collect()
+    }
+
+    #[test]
+    fn a_single_frame_roster_is_returned_immediately() {
+        // Every realistic guild takes this path, and it must not depend on the buffer at all.
+        let mut parts = HashMap::new();
+        let out = accumulate_roster(&mut parts, 1, 0, false, members(&["Ada", "Bo"]));
+
+        assert_eq!(names(&out.expect("complete")), ["Ada", "Bo"]);
+        assert!(parts.is_empty(), "nothing should be buffered");
+    }
+
+    #[test]
+    fn a_chunked_roster_reassembles_in_order() {
+        // The case the live rig caught: without this, only the final frame survived and a
+        // 155-member guild appeared on the board with 3 members.
+        let mut parts = HashMap::new();
+
+        assert!(accumulate_roster(&mut parts, 1, 0, true, members(&["Ada"])).is_none());
+        assert!(accumulate_roster(&mut parts, 1, 1, true, members(&["Bo"])).is_none());
+        let out = accumulate_roster(&mut parts, 1, 2, false, members(&["Cy"]));
+
+        assert_eq!(names(&out.expect("complete")), ["Ada", "Bo", "Cy"]);
+        assert!(parts.is_empty(), "buffer is released once complete");
+    }
+
+    #[test]
+    fn a_restarted_roster_supersedes_a_partial_one() {
+        // A shard that reconnects mid-emission starts again at seq 0. The abandoned frames must not
+        // end up spliced onto the front of the new roster.
+        let mut parts = HashMap::new();
+
+        assert!(accumulate_roster(&mut parts, 1, 0, true, members(&["Stale"])).is_none());
+        let out = accumulate_roster(&mut parts, 1, 0, false, members(&["Fresh"]));
+
+        assert_eq!(names(&out.expect("complete")), ["Fresh"]);
+    }
+
+    #[test]
+    fn an_out_of_order_frame_discards_the_partial_roster() {
+        // Better to publish nothing and wait for the next full emission than to store a roster with
+        // a hole in it that nothing downstream could detect.
+        let mut parts = HashMap::new();
+
+        assert!(accumulate_roster(&mut parts, 1, 0, true, members(&["Ada"])).is_none());
+        assert!(accumulate_roster(&mut parts, 1, 2, false, members(&["Skipped"])).is_none());
+        assert!(parts.is_empty());
+    }
+
+    #[test]
+    fn a_continuation_with_no_start_is_ignored() {
+        // The sidecar restarting midway through a shard's emission.
+        let mut parts = HashMap::new();
+
+        assert!(accumulate_roster(&mut parts, 1, 3, false, members(&["Orphan"])).is_none());
+        assert!(parts.is_empty());
+    }
+
+    #[test]
+    fn two_guilds_reassemble_independently() {
+        // Rosters for different guilds interleave freely — the sweep emits one guild after another
+        // and nothing serialises them on the wire.
+        let mut parts = HashMap::new();
+
+        assert!(accumulate_roster(&mut parts, 1, 0, true, members(&["A1"])).is_none());
+        assert!(accumulate_roster(&mut parts, 2, 0, true, members(&["B1"])).is_none());
+        let g2 = accumulate_roster(&mut parts, 2, 1, false, members(&["B2"]));
+        let g1 = accumulate_roster(&mut parts, 1, 1, false, members(&["A2"]));
+
+        assert_eq!(names(&g2.expect("guild 2")), ["B1", "B2"]);
+        assert_eq!(names(&g1.expect("guild 1")), ["A1", "A2"]);
+    }
+
+    #[test]
+    fn an_empty_roster_is_a_complete_roster() {
+        // A guild whose last member left emits one frame with an empty array. Treating that as
+        // "nothing to store" would leave the board showing the roster it had before.
+        let mut parts = HashMap::new();
+        let out = accumulate_roster(&mut parts, 1, 0, false, vec![]);
+
+        assert_eq!(out.expect("complete").len(), 0);
+    }
+}

sidecar/src/main.rs

@@ -46,7 +46,85 @@ use tracing_subscriber::EnvFilter;
 /// `vendor.listing.remove`, with the `GET /ruleset`, `/points` and `/market` reads that serve them
 /// from the store. Same shape as the v2 bump — the kinds are additive, the endpoints are not — and
 /// there is deliberately no feature-negotiation array: v3 implies all three kinds.
-pub const PROTOCOL_VERSION: u32 = 3;
+///
+/// v4 (Protocol 4): adds `guild.roster` and `guild.leave`, giving the guild board a real member list
+/// instead of the member *count* that was all v2 could express. Additive in the same way again — the
+/// kinds are new, `GET /guilds` grows a `roster` key, and nothing existing changed shape. This is the
+/// first bump that also needed a **store migration** (`guilds.members`), because it is the first to
+/// add a column to a table that already exists rather than a whole new table; see `store::migrate`.
+///
+/// v5 (Protocol 5): three enrichments that are additive in the same way again, bumped together
+/// rather than one at a time because a protocol bump is not cheap here — it costs a sidecar
+/// release, a republished bundle and an operator update on every shard, so a field left out costs
+/// a whole second round of that rather than a follow-up commit. They are:
+///
+///   * `house.decay` gains `ownerName` and a decay SCHEDULE — `nextStage`, `decayPeriodSec`,
+///     `dynamicDecay`, and `estimatedCollapse` only where it is exactly knowable (at IDOC under
+///     dynamic decay; at any stage under static decay, which has no randomness to wait out).
+///   * `vendor.listing` gains `ownerAcct` — without which the frame names an owner nobody can
+///     resolve to a person — and a `fees` object carrying the charge, the funds, the pay interval
+///     and the resolved `dismissalAt`.
+///   * `account.login.result` is a NEW kind: the verdict of a login, which the pre-existing
+///     `account.login.attempt` structurally cannot carry (its EventSink fires before the auth
+///     decision is made).
+///
+/// **No store migration this time**, unlike v4. Every frame is persisted whole and the board tables
+/// index only the columns they already had, so the new fields ride inside the stored JSON and the
+/// new kind lands in `events` like any other. That is the dumb-forwarder property doing its job:
+/// the sidecar defines no schema for a frame's contents and so needs no change when they grow.
+///
+/// v6 (Protocol 6): the first bump that is about a GUARANTEE rather than about data, and the first
+/// the sidecar mostly gets for free. Two things:
+///
+///   * **`idempotencyKey` on inbound commands.** A command that carries one is executed by the shard
+///     at most once; a repeat is answered with the original reply rather than re-run. That is what
+///     makes a world-writing verb retryable at all — until now a lost acknowledgement was
+///     indistinguishable from a command that never applied, so the website had to declare every
+///     write un-retryable and accept losing one rather than risk doubling it. The sidecar's part is
+///     to CARRY the key (it rides in the command body, which every write endpoint already passes
+///     through verbatim) and to understand the one new answer the shard can now give: `bridge.busy`,
+///     meaning a command under that key is still in flight. See `web::respond`.
+///   * **`champ.boss.killed` is a new kind**: a champion's defeat, with the damage table only the
+///     shard ever sees. It was previously inferable from `champ.update` going `bossUp` true then
+///     false alongside a nearby `mob.killed`, which is fragile and says nothing about who did the
+///     work. It lands in `events` and on the feed like any other kind, with no code here at all —
+///     the dumb-forwarder property again.
+///
+/// **No store migration.** Nothing gains a column; the new kind is persisted whole like every other.
+///
+/// # Protocol 8 — the Asset Bridge (docs/link/v8.md)
+///
+/// The shard starts sending the operator's own **client assets** over this link: the cliloc string
+/// table, creature and item art, player models. The point is that an operator stops having to run
+/// a GUI converter on a desktop to make their site render a bestiary, and the shard is the only
+/// host that already has the client files — a ServUO server cannot boot without them.
+///
+/// Phase 1 is the transport, and the sidecar's share of it is three things:
+///
+///   * **A new command family, `assets.*`, forwarded verbatim** like every other. The first of them
+///     is `assets.sources` — stage 1 of the import gate: what the client files currently are, and
+///     what version of the shard's extractor would read them. No pixels cross on this call.
+///   * **An inbound line cap** — [`shard::MAX_INBOUND_LINE_BYTES`]. This is the one change that is
+///     not additive. `read_line` had no bound at all, which was survivable while the shard had no
+///     reason to send a large line; protocol 8 gives it one deliberately, and an unbounded read
+///     facing a component that now sends megabytes is a memory-exhaustion shape we would be
+///     inventing ourselves.
+///   * **Nothing else.** Assets ride the request/reply path, so `rpc::try_route` consumes them
+///     before `app.rs` can persist them to the store and fan them out to every WebSocket
+///     subscriber — which is what keeps a 512 KiB reply from being written to SQLite and broadcast
+///     to every connected client. The dumb-forwarder property is doing real work here: the sidecar
+///     does not know what an asset is, and must not learn.
+///
+/// Phase 2 adds the first family that actually carries content: **`cliloc.table`**, served at
+/// `GET /cliloc`. It pages — the shard cuts at a byte budget and the caller echoes a cursor back —
+/// and the sidecar forwards it without keeping any of it, which matters more here than usual: the
+/// payload is five megabytes of EA's strings out of the operator's own client, and the one copy of
+/// it that should exist is the one the website imports. That phase also gave `assets.error` a
+/// `code`, so the status a refusal maps to stops depending on the wording of a human-facing
+/// sentence (see [`web::asset_error_status`]).
+///
+/// **No store migration**, again: nothing on this plane is an event, so nothing is persisted.
+pub const PROTOCOL_VERSION: u32 = 8;
 
 // Not `#[tokio::main]`: on Windows the SCM dispatcher takes over this thread and starts the runtime
 // itself, on its own thread, once the service actually begins. The runtime is built by whichever

sidecar/src/shard.rs

@@ -7,15 +7,130 @@
 //! Framing is newline-delimited JSON, bidirectional: the shard sends events, we send commands. We
 //! accept one shard connection at a time and re-accept when it drops (the shard reconnects on its
 //! own, with a bounded backoff).
+//!
+//! Inbound lines are **capped** (see [`MAX_INBOUND_LINE_BYTES`]). Until protocol 8 they were not:
+//! `read_line` will buffer a line of any length, which was survivable only because the shard had
+//! never had a reason to send a large one. The Asset Bridge gives it one, so the gap had to close
+//! before it became a memory-exhaustion shape we invented ourselves.
 
 use std::sync::Arc;
 
 use serde_json::Value;
-use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader};
+use tokio::io::{AsyncBufRead, AsyncBufReadExt, AsyncWriteExt, BufReader};
 use tokio::net::TcpListener;
 use tokio::sync::{mpsc, Mutex};
 use tracing::{info, warn};
 
+/// The longest line the sidecar will accept from the shard, in bytes.
+///
+/// Set above the largest legal batch rather than at it: the shard cuts a batch when the next item
+/// would take it past `Bridge.AssetBatchBytes` (512 KiB), and always admits the first item of a
+/// page even when that item alone is bigger than the budget — so one page can legitimately
+/// overshoot by one item. Doubling the budget to get this cap is what makes that overshoot safe
+/// instead of a dropped reply.
+///
+/// Over-long lines are **discarded, not buffered**, and the connection stays up. That is the same
+/// disposition `BridgeLink.cs` has always had for its own 1 MiB inbound cap in the other
+/// direction, and it is the right one here: a single malformed frame is not a reason to tear down
+/// a link that live events are flowing over. The dropped reply simply times out and is
+/// re-requested, which is safe because everything on the asset plane is idempotent.
+pub const MAX_INBOUND_LINE_BYTES: usize = 1024 * 1024;
+
+/// What one read off the shard socket produced.
+#[derive(Debug)]
+enum Line {
+    /// A complete line, within the cap.
+    Complete(String),
+    /// A line that ran past the cap. Carries how many bytes were thrown away, for the log.
+    TooLong(usize),
+    /// The shard closed the connection.
+    Eof,
+}
+
+/// A cancel-safe, capped, newline-delimited reader.
+///
+/// Every piece of state that must survive a partial read lives here rather than in a local,
+/// because this is polled inside a `tokio::select!`: the loop below drops the future whenever a
+/// command wins the race, and a `discarding` flag or a half-filled buffer held in a local would be
+/// lost with it. Losing the buffer corrupts the *next* line; losing `discarding` turns the tail of
+/// an over-long line into a line of its own. Both are silent.
+///
+/// The only await point is `fill_buf`, and nothing is consumed until after it returns, so a
+/// cancellation between the two can lose at most the wakeup.
+#[derive(Default)]
+struct LineReader {
+    buf: Vec<u8>,
+    discarding: bool,
+    discarded: usize,
+}
+
+impl LineReader {
+    async fn next<R: AsyncBufRead + Unpin>(&mut self, reader: &mut R) -> std::io::Result<Line> {
+        loop {
+            let consumed;
+            let outcome;
+
+            {
+                let available = reader.fill_buf().await?;
+
+                if available.is_empty() {
+                    return Ok(Line::Eof);
+                }
+
+                match available.iter().position(|&b| b == b'\n') {
+                    Some(at) => {
+                        consumed = at + 1;
+
+                        if self.discarding {
+                            // The tail of a line we already gave up on. Swallow it, terminator
+                            // included, and report the size once.
+                            self.discarded += at;
+                            let total = self.discarded;
+                            self.discarding = false;
+                            self.discarded = 0;
+                            outcome = Some(Line::TooLong(total));
+                        } else if self.buf.len() + at > MAX_INBOUND_LINE_BYTES {
+                            // The cap is reached only now, on the chunk that also holds the
+                            // terminator — so there is nothing left to discard.
+                            let total = self.buf.len() + at;
+                            self.buf.clear();
+                            outcome = Some(Line::TooLong(total));
+                        } else {
+                            self.buf.extend_from_slice(&available[..at]);
+                            let line = String::from_utf8_lossy(&self.buf).into_owned();
+                            self.buf.clear();
+                            outcome = Some(Line::Complete(line));
+                        }
+                    }
+                    None => {
+                        consumed = available.len();
+
+                        if self.discarding {
+                            self.discarded += consumed;
+                        } else if self.buf.len() + consumed > MAX_INBOUND_LINE_BYTES {
+                            // Refuse rather than buffer: this is the whole point of the cap.
+                            // Everything up to the next newline is now dropped on the floor.
+                            self.discarded = self.buf.len() + consumed;
+                            self.buf.clear();
+                            self.discarding = true;
+                        } else {
+                            self.buf.extend_from_slice(available);
+                        }
+
+                        outcome = None;
+                    }
+                }
+            }
+
+            reader.consume(consumed);
+
+            if let Some(line) = outcome {
+                return Ok(line);
+            }
+        }
+    }
+}
+
 /// An event line received from the shard, parsed. `kind` is lifted out for routing.
 #[derive(Debug, Clone)]
 pub struct ShardEvent {
@@ -112,31 +227,41 @@ async fn handle_connection(
     handle.set(Some(cmd_tx)).await;
 
     let mut reader = BufReader::new(read_half);
-    let mut line = String::new();
+    let mut lines = LineReader::default();
 
     loop {
         tokio::select! {
             // Inbound: a line from the shard.
-            result = reader.read_line(&mut line) => {
-                let n = result?;
-                if n == 0 {
-                    return Ok(()); // clean EOF: shard closed
-                }
-                let trimmed = line.trim_end();
-                if !trimmed.is_empty() {
-                    match serde_json::from_str::<Value>(trimmed) {
-                        Ok(value) => {
-                            let kind = value
-                                .get("kind")
-                                .and_then(|k| k.as_str())
-                                .unwrap_or("")
-                                .to_string();
-                            let _ = event_tx.send(ShardEvent { kind, value });
+            result = lines.next(&mut reader) => {
+                match result? {
+                    Line::Eof => return Ok(()), // clean EOF: shard closed
+                    Line::TooLong(bytes) => {
+                        // Deliberately not a disconnect. See MAX_INBOUND_LINE_BYTES: a reply lost
+                        // this way times out on the caller's side and is re-requested, and tearing
+                        // the link down would take the live event feed with it.
+                        warn!(
+                            bytes,
+                            cap = MAX_INBOUND_LINE_BYTES,
+                            "inbound line over the cap; discarded"
+                        );
+                    }
+                    Line::Complete(line) => {
+                        let trimmed = line.trim_end();
+                        if !trimmed.is_empty() {
+                            match serde_json::from_str::<Value>(trimmed) {
+                                Ok(value) => {
+                                    let kind = value
+                                        .get("kind")
+                                        .and_then(|k| k.as_str())
+                                        .unwrap_or("")
+                                        .to_string();
+                                    let _ = event_tx.send(ShardEvent { kind, value });
+                                }
+                                Err(e) => warn!(error = %e, line = %trimmed, "unparseable event"),
+                            }
                         }
-                        Err(e) => warn!(error = %e, line = %trimmed, "unparseable event"),
                     }
                 }
-                line.clear();
             }
             // Outbound: a command to write to the shard.
             cmd = cmd_rx.recv() => {
@@ -152,3 +277,108 @@ async fn handle_connection(
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Drives `LineReader` over a byte slice, returning every outcome up to EOF.
+    async fn read_all(input: &[u8]) -> Vec<Line> {
+        let mut reader = BufReader::with_capacity(64, input);
+        let mut lines = LineReader::default();
+        let mut out = Vec::new();
+
+        loop {
+            match lines.next(&mut reader).await.unwrap() {
+                Line::Eof => break,
+                other => out.push(other),
+            }
+        }
+
+        out
+    }
+
+    fn complete(lines: &[Line]) -> Vec<&str> {
+        lines
+            .iter()
+            .filter_map(|l| match l {
+                Line::Complete(s) => Some(s.as_str()),
+                _ => None,
+            })
+            .collect()
+    }
+
+    #[tokio::test]
+    async fn splits_on_newlines() {
+        let lines = read_all(b"{\"a\":1}\n{\"b\":2}\n").await;
+        assert_eq!(complete(&lines), vec!["{\"a\":1}", "{\"b\":2}"]);
+    }
+
+    /// The reader's buffer is 64 bytes here, so every one of these lines spans several
+    /// `fill_buf` chunks. Reassembly across chunks is the thing `read_line` did for us.
+    #[tokio::test]
+    async fn reassembles_across_chunks() {
+        let long = "x".repeat(500);
+        let input = format!("{}\n{}\n", long, long);
+        let lines = read_all(input.as_bytes()).await;
+
+        assert_eq!(complete(&lines), vec![long.as_str(), long.as_str()]);
+    }
+
+    /// The cap itself. The over-long line must be reported and thrown away, and — the part that
+    /// actually matters — the line *after* it must still arrive intact. A reader that lost its
+    /// `discarding` flag would emit the tail of the oversized line as a line of its own.
+    #[tokio::test]
+    async fn refuses_an_over_long_line_and_recovers() {
+        let mut input = Vec::new();
+        input.extend_from_slice(&b"a".repeat(MAX_INBOUND_LINE_BYTES + 10));
+        input.push(b'\n');
+        input.extend_from_slice(b"{\"kind\":\"pong\"}\n");
+
+        let lines = read_all(&input).await;
+
+        assert_eq!(lines.len(), 2);
+        assert!(
+            matches!(lines[0], Line::TooLong(n) if n >= MAX_INBOUND_LINE_BYTES),
+            "expected TooLong, got {:?}",
+            lines[0]
+        );
+        assert_eq!(complete(&lines), vec!["{\"kind\":\"pong\"}"]);
+    }
+
+    /// A line of exactly the cap is legal; one byte more is not. Checking both sides is what says
+    /// the comparison is `>` rather than `>=`, which would silently cost a byte of the budget.
+    #[tokio::test]
+    async fn the_cap_is_inclusive() {
+        let at_cap = "b".repeat(MAX_INBOUND_LINE_BYTES);
+        let lines = read_all(format!("{}\n", at_cap).as_bytes()).await;
+        assert_eq!(complete(&lines).len(), 1);
+
+        let over = "b".repeat(MAX_INBOUND_LINE_BYTES + 1);
+        let lines = read_all(format!("{}\n", over).as_bytes()).await;
+        assert!(complete(&lines).is_empty());
+        assert!(matches!(lines[0], Line::TooLong(_)));
+    }
+
+    /// An over-long line whose terminator lands in the very chunk that crosses the cap: the
+    /// reader must not leave itself in `discarding` and eat the next line as well.
+    #[tokio::test]
+    async fn over_long_line_terminating_in_the_crossing_chunk() {
+        let mut input = Vec::new();
+        input.extend_from_slice(&b"c".repeat(MAX_INBOUND_LINE_BYTES + 1));
+        input.extend_from_slice(b"\n{\"kind\":\"pong\"}\n");
+
+        let lines = read_all(&input).await;
+
+        assert!(matches!(lines[0], Line::TooLong(_)));
+        assert_eq!(complete(&lines), vec!["{\"kind\":\"pong\"}"]);
+    }
+
+    /// A partial line at EOF is dropped rather than delivered half-parsed. The shard reconnects
+    /// and re-sends; half a JSON object is not something to hand to the event fan-out.
+    #[tokio::test]
+    async fn trailing_partial_line_at_eof_is_dropped() {
+        let lines = read_all(b"{\"a\":1}\n{\"b\":").await;
+        assert_eq!(complete(&lines), vec!["{\"a\":1}"]);
+    }
+}

sidecar/src/store.rs

@@ -45,6 +45,7 @@ impl Store {
             .await?;
 
         sqlx::query(SCHEMA).execute(&pool).await?;
+        migrate(&pool).await?;
         info!(%path, "store ready");
         Ok(Self { pool })
     }
@@ -236,12 +237,61 @@ impl Store {
         Ok(())
     }
 
+    /// Upserts one guild's member roster (Protocol 4), keyed by guild id, touching **only** the
+    /// `members` column.
+    ///
+    /// Deliberately not a write to `json`. That column holds the verbatim `guild.update` line, and a
+    /// roster arriving as its own event must not clobber the snapshot — name, abbreviation, leader,
+    /// online count — that `guild.update` owns. Splitting the two writers across two columns of one
+    /// row is what lets both be plain upserts: neither needs to read the other's value first, so
+    /// there is no read-modify-write and no ordering requirement between the two kinds.
+    ///
+    /// The `INSERT` half is not redundant: a roster can arrive before the first `guild.update` for a
+    /// guild, and the row it creates then carries `'{}'` until that update fills it in.
+    pub async fn upsert_guild_roster(
+        &self,
+        id: i64,
+        members_json: &str,
+        t: i64,
+    ) -> anyhow::Result<()> {
+        sqlx::query(
+            "INSERT INTO guilds (id, name, json, updated_t, members) VALUES (?, NULL, '{}', ?, ?)
+             ON CONFLICT(id) DO UPDATE SET members = excluded.members, updated_t = excluded.updated_t",
+        )
+        .bind(id)
+        .bind(t)
+        .bind(members_json)
+        .execute(&self.pool)
+        .await?;
+        Ok(())
+    }
+
     /// The full guild board: every guild's latest snapshot, ordered by name.
+    ///
+    /// The roster is stored in its own column (see [`Self::upsert_guild_roster`]) and folded into
+    /// the projected object as `roster` here, at read time. A guild that has had a `guild.update`
+    /// but no `guild.roster` yet simply has no `roster` key, which is the honest representation of
+    /// "not known" and distinct from a guild whose roster is genuinely empty.
     pub async fn guilds_all(&self) -> anyhow::Result<Vec<Value>> {
-        let rows = sqlx::query("SELECT json FROM guilds ORDER BY name, id")
+        let rows = sqlx::query("SELECT json, members FROM guilds ORDER BY name, id")
             .fetch_all(&self.pool)
             .await?;
-        Ok(parse_json_column(rows))
+
+        Ok(rows
+            .into_iter()
+            .filter_map(|r| {
+                let mut v: Value = serde_json::from_str(&r.get::<String, _>("json")).ok()?;
+                let members: Option<String> = r.get("members");
+
+                if let (Some(obj), Some(raw)) = (v.as_object_mut(), members) {
+                    if let Ok(list) = serde_json::from_str::<Value>(&raw) {
+                        obj.insert("roster".into(), list);
+                    }
+                }
+
+                Some(v)
+            })
+            .collect())
     }
 
     // ---- governor board (Protocol 2.0) ----
@@ -513,6 +563,75 @@ impl Store {
     }
 }
 
+/// The schema version this build expects. Bump it, and add the matching arm to [`migrate`], for
+/// every change that `SCHEMA` alone cannot make to a database that already exists.
+const SCHEMA_VERSION: i64 = 1;
+
+/// Brings an existing database forward to [`SCHEMA_VERSION`].
+///
+/// `SCHEMA` is `CREATE TABLE IF NOT EXISTS` only, which is enough to *add a table* but cannot add a
+/// column to a table that is already there. Every schema change up to and including Protocol 3.0
+/// happened to add whole tables, so this never mattered and `ALTER TABLE` appears nowhere in this
+/// repo's history. `guilds.members` (Protocol 4) is the first column added to an existing table, so
+/// the mechanism has to exist now.
+///
+/// The version counter is SQLite's own `PRAGMA user_version`: an integer in the database header, so
+/// it needs no table of its own and cannot be separated from the file it describes. Each step runs
+/// in a transaction **together with** the bump that records it, so a step either lands completely or
+/// not at all, and an interrupted run resumes at the right place rather than re-applying half of one.
+///
+/// A failure here propagates and aborts startup, deliberately. A half-migrated store answers the
+/// website with confusing partial data, which is worse than being plainly absent — and the shard
+/// dials *out* to the sidecar, so a sidecar that refuses to start never stalls the game.
+async fn migrate(pool: &SqlitePool) -> anyhow::Result<()> {
+    let mut version: i64 = sqlx::query_scalar("PRAGMA user_version")
+        .fetch_one(pool)
+        .await?;
+
+    // A database written by a *newer* sidecar than this binary. This is not an error: every step
+    // here is additive, so a newer schema has only columns and tables an older reader ignores, and
+    // refusing to start would turn "roll the binary back" — a recovery path — into a dead end.
+    if version > SCHEMA_VERSION {
+        tracing::warn!(
+            found = version,
+            expected = SCHEMA_VERSION,
+            "store was written by a newer sidecar; continuing, as migrations are additive"
+        );
+        return Ok(());
+    }
+
+    while version < SCHEMA_VERSION {
+        let next = version + 1;
+        let mut tx = pool.begin().await?;
+
+        match next {
+            // Protocol 4: the guild board carries a member roster. Its own column rather than a
+            // field folded into `json`, because `json` holds the verbatim `guild.update` line and
+            // the two writers must not overwrite each other — see `upsert_guild_roster`.
+            1 => {
+                sqlx::query("ALTER TABLE guilds ADD COLUMN members TEXT")
+                    .execute(&mut *tx)
+                    .await?;
+            }
+            // Unreachable while SCHEMA_VERSION and this match are edited together, which is the
+            // point of failing loudly rather than silently leaving the counter short.
+            n => anyhow::bail!("no migration step defined for schema version {n}"),
+        }
+
+        // `PRAGMA` takes no bind parameters, so this is formatted — safe because `next` is an i64
+        // this loop produced, never anything from outside the process.
+        sqlx::query(&format!("PRAGMA user_version = {next}"))
+            .execute(&mut *tx)
+            .await?;
+
+        tx.commit().await?;
+        info!(version = next, "schema migration applied");
+        version = next;
+    }
+
+    Ok(())
+}
+
 fn parse_json_column(rows: Vec<sqlx::sqlite::SqliteRow>) -> Vec<Value> {
     rows.into_iter()
         .filter_map(|r| serde_json::from_str(&r.get::<String, _>("json")).ok())
@@ -611,3 +730,195 @@ CREATE TABLE IF NOT EXISTS ruleset (
     updated_t INTEGER NOT NULL
 );
 "#;
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// A unique scratch database path. Matches `config`'s idiom — `std::env::temp_dir()` plus the
+    /// test name — so the cases stay independent under the parallel test runner.
+    fn scratch(name: &str) -> String {
+        let dir = std::env::temp_dir().join(format!("uo-link-store-test-{name}"));
+        let _ = std::fs::remove_dir_all(&dir);
+        std::fs::create_dir_all(&dir).expect("create scratch dir");
+        dir.join("uo-link.db").to_string_lossy().into_owned()
+    }
+
+    /// The `guilds` table exactly as a pre-Protocol-4 sidecar left it: no `members` column, and
+    /// `user_version` still 0. This is the shape a real operator's database is in before an update,
+    /// and the only starting point where the migration does anything.
+    async fn legacy_db(path: &str) -> SqlitePool {
+        let pool = SqlitePoolOptions::new()
+            .max_connections(1)
+            .connect_with(
+                SqliteConnectOptions::new()
+                    .filename(path)
+                    .create_if_missing(true),
+            )
+            .await
+            .expect("open legacy db");
+
+        sqlx::query(
+            "CREATE TABLE guilds (
+                 id        INTEGER PRIMARY KEY,
+                 name      TEXT,
+                 json      TEXT NOT NULL,
+                 updated_t INTEGER NOT NULL
+             )",
+        )
+        .execute(&pool)
+        .await
+        .expect("create legacy guilds table");
+
+        pool.close().await;
+        pool
+    }
+
+    async fn user_version(store: &Store) -> i64 {
+        sqlx::query_scalar("PRAGMA user_version")
+            .fetch_one(&store.pool)
+            .await
+            .expect("read user_version")
+    }
+
+    async fn guild_columns(store: &Store) -> Vec<String> {
+        sqlx::query("PRAGMA table_info(guilds)")
+            .fetch_all(&store.pool)
+            .await
+            .expect("table_info")
+            .into_iter()
+            .map(|r| r.get::<String, _>("name"))
+            .collect()
+    }
+
+    #[tokio::test]
+    async fn an_existing_pre_protocol_4_database_gains_the_members_column() {
+        // The case that matters: `SCHEMA`'s CREATE TABLE IF NOT EXISTS is a no-op against a table
+        // that is already there, so without `migrate` this database would never get the column and
+        // every roster write would fail against a live install.
+        let path = scratch("legacy-upgrade");
+        legacy_db(&path).await;
+
+        let store = Store::open(&path).await.expect("open migrates");
+
+        assert!(
+            guild_columns(&store).await.contains(&"members".to_string()),
+            "the migration must add guilds.members to a database that already had the table"
+        );
+        assert_eq!(user_version(&store).await, SCHEMA_VERSION);
+    }
+
+    #[tokio::test]
+    async fn a_fresh_database_lands_at_the_current_version() {
+        let path = scratch("fresh");
+        let store = Store::open(&path).await.expect("open");
+
+        assert!(guild_columns(&store).await.contains(&"members".to_string()));
+        assert_eq!(user_version(&store).await, SCHEMA_VERSION);
+    }
+
+    #[tokio::test]
+    async fn reopening_an_already_migrated_database_is_a_no_op() {
+        // Every sidecar restart re-runs this path, so a second run must not attempt the ALTER again
+        // — which would fail with "duplicate column name" and, since a migration failure aborts
+        // startup, would leave the sidecar unable to start at all after its first upgrade.
+        let path = scratch("idempotent");
+        legacy_db(&path).await;
+
+        Store::open(&path).await.expect("first open");
+        let store = Store::open(&path).await.expect("second open must succeed");
+
+        assert_eq!(user_version(&store).await, SCHEMA_VERSION);
+    }
+
+    #[tokio::test]
+    async fn a_roster_does_not_clobber_the_guild_update_snapshot() {
+        // The invariant the two-column split exists to give. `json` holds the verbatim guild.update
+        // line; if a roster write touched it, name/abbr/online would vanish from the board.
+        let path = scratch("no-clobber");
+        let store = Store::open(&path).await.expect("open");
+
+        store
+            .upsert_guild(
+                7,
+                Some("The Cartographers"),
+                r#"{"kind":"guild.update","id":7,"name":"The Cartographers","abbr":"MAP","members":2,"online":1}"#,
+                100,
+            )
+            .await
+            .expect("upsert guild");
+
+        store
+            .upsert_guild_roster(
+                7,
+                r#"[{"serial":"0x1","name":"Ada"},{"serial":"0x2","name":"Bo"}]"#,
+                200,
+            )
+            .await
+            .expect("upsert roster");
+
+        let guilds = store.guilds_all().await.expect("read board");
+        assert_eq!(guilds.len(), 1);
+        let g = &guilds[0];
+
+        assert_eq!(
+            g["name"], "The Cartographers",
+            "guild.update's name survived"
+        );
+        assert_eq!(g["abbr"], "MAP", "guild.update's abbr survived");
+        assert_eq!(g["online"], 1, "guild.update's online count survived");
+        assert_eq!(g["roster"].as_array().expect("roster is an array").len(), 2);
+        assert_eq!(g["roster"][0]["name"], "Ada");
+    }
+
+    #[tokio::test]
+    async fn the_two_writers_are_order_independent() {
+        // A roster can arrive before the first guild.update for a guild — on a reconnect the shard
+        // re-emits both and nothing orders them. Neither write may depend on the other's row.
+        let path = scratch("either-order");
+        let store = Store::open(&path).await.expect("open");
+
+        store
+            .upsert_guild_roster(9, r#"[{"serial":"0x3","name":"Cy"}]"#, 100)
+            .await
+            .expect("roster first");
+        store
+            .upsert_guild(
+                9,
+                Some("Late Arrivals"),
+                r#"{"kind":"guild.update","id":9,"name":"Late Arrivals","abbr":"LTE"}"#,
+                200,
+            )
+            .await
+            .expect("update second");
+
+        let guilds = store.guilds_all().await.expect("read board");
+        assert_eq!(guilds.len(), 1, "one row, not two");
+        assert_eq!(guilds[0]["name"], "Late Arrivals");
+        assert_eq!(guilds[0]["roster"].as_array().expect("roster").len(), 1);
+    }
+
+    #[tokio::test]
+    async fn a_guild_with_no_roster_yet_has_no_roster_key() {
+        // "Not known" and "known to be empty" are different, and the board must not conflate them:
+        // a website reading `roster: []` would render an empty roster as fact.
+        let path = scratch("absent-roster");
+        let store = Store::open(&path).await.expect("open");
+
+        store
+            .upsert_guild(
+                11,
+                Some("Unswept"),
+                r#"{"kind":"guild.update","id":11,"name":"Unswept"}"#,
+                100,
+            )
+            .await
+            .expect("upsert guild");
+
+        let guilds = store.guilds_all().await.expect("read board");
+        assert!(
+            guilds[0].get("roster").is_none(),
+            "a guild with no roster event must not grow a roster key"
+        );
+    }
+}