fix(sidecar): protocol 13 — a configuration write no longer waits for its reload (F9)

The plugin now answers config.write once the files are on disk and reports
the reload as a config.outcome event, which this process files and feeds by
its type like any other event. The reload window no longer has to fit
inside REPLY_TIMEOUT, so CONFIG_RELOAD_WINDOW and the test that asserted
the pairing are gone, and the 504 body stops pointing at a rollback.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E14m6SuuY6i1vASFeGDBeY

sidecar/src/web.rs

@@ -102,9 +102,8 @@ pub async fn serve(addr: &str, state: AppState) -> anyhow::Result<()> {
         // an operator edited over SSH and the website then overwrote.
         .route("/config/files", get(config_files))
         .route("/config/file", get(config_file))
-        // The only route on this sidecar that causes a WRITE on the game host, and the only one
-        // whose reply can take most of the RPC budget: the plugin holds it open across a reload
-        // and, at worst, across a rollback as well. See `CONFIG_RELOAD_WINDOW`.
+        // The only route on this sidecar that causes a WRITE on the game host. It answers once the
+        // files are on disk; the reload's outcome follows as a `config.outcome` event (protocol 13).
         .route("/config/write", post(config_write))
         // Protocol 8 (the module's PLAN.md §27): an event borrowing a value and giving it back.
         // Three correlated round trips, and the sidecar knows nothing about any of them — not
@@ -694,19 +693,14 @@ async fn config_file(State(st): State<AppState>, Query(q): Query<ConfigPathQuery
     respond(st.rpc.call(&st.game, command, &req_id).await)
 }
 
-/// The window the plugin gives a reload to announce itself, mirrored here from
-/// `ConfigReloadWindowSeconds` in `overlay/oxide/plugins/RunicGateway.cs`.
-///
-/// It is duplicated rather than negotiated because the two numbers are a *pairing*, like the
-/// protocol version: the plugin owns the behaviour and this end owns the budget it has to fit in.
-/// The test below is what keeps them honest — a worst-case write is two windows, and a
-/// [`REPLY_TIMEOUT`](crate::rpc::REPLY_TIMEOUT) that does not cover both would abandon the caller
-/// precisely when a rollback had just saved them, leaving the website to report a timeout over a
-/// server that is perfectly healthy.
-pub const CONFIG_RELOAD_WINDOW: Duration = Duration::from_secs(4);
-
 /// Replace a set of configuration files and reload whatever owns them.
 ///
+/// Since protocol 13 the reply does not wait for the reload. The plugin answers once the files are
+/// on disk (`pending: true`, a `writeId`, and the `ceilingMs` it will wait), and reports the reload
+/// as a `config.outcome` event when it settles. Protocol 12 held this RPC across the reload, which
+/// forced the plugin's window under [`REPLY_TIMEOUT`](crate::rpc::REPLY_TIMEOUT) — and a cold
+/// compile outlasted it and had a valid edit rolled back (the module's PLAN_FIXES.md F9).
+///
 /// Opaque body, exactly as `/permissions/sync`: `cmd` and `reqId` are written over whatever the
 /// caller sent, and nothing else about the object is read here. What the website sends is whole
 /// file *text* rather than a key and a value (D35), so there is nothing on this hop that could
@@ -750,16 +744,15 @@ async fn config_write(State(st): State<AppState>, Json(body): Json<Value>) -> Re
     let result = st.rpc.call(&st.game, command, &req_id).await;
 
     if matches!(result, Err(RpcError::Timeout)) {
-        // A bare `504` on a route that writes reads as "did my change land, and is the plugin
-        // still up?" — and unlike every other timeout on this sidecar, that question has a good
-        // answer. The plugin writes a whole set or restores a whole set, never half of either, so
-        // a re-read settles it; and a write that is still in flight is most likely inside the
-        // rollback this window pays for.
+        // A bare `504` on a route that writes reads as "did my change land?" — and unlike every
+        // other timeout on this sidecar, that question has a good answer. The plugin writes a
+        // whole set or none of it, and restores a whole set or none of it, so a re-read settles
+        // it. Since protocol 13 the reply no longer waits for a reload, so this is a slow host or
+        // a busy main thread, not a rollback in progress.
         return (
             StatusCode::GATEWAY_TIMEOUT,
             Json(json!({
                 "error": "the plugin did not report within the write budget",
-                "reloadWindowSeconds": CONFIG_RELOAD_WINDOW.as_secs(),
                 "hint": "re-read the files: the plugin writes the whole set or restores it",
             })),
         )
@@ -1037,25 +1030,6 @@ mod tests {
         assert_eq!(command["grants"], json!([]));
     }
 
-    /// A command larger than the game link's own line cap is refused here, where the caller learns
-    /// why. Forwarded, it would be discarded by both ends without a word and present as a `504`.
-    /// The pairing in `CONFIG_RELOAD_WINDOW`'s own words, asserted rather than trusted.
-    ///
-    /// The worst path through one configuration write is two windows — wait for the edited
-    /// plugin's reload, give up, restore the files, reload again — plus the write, the reads and
-    /// a line each way. If that does not fit inside the RPC timeout, the caller is abandoned at
-    /// exactly the moment the rollback saved it, and the website reports a timeout over a server
-    /// that is healthy and has the old config back.
-    #[test]
-    fn a_rollback_fits_inside_the_rpc_budget() {
-        let worst = CONFIG_RELOAD_WINDOW * 2;
-        assert!(
-            worst + Duration::from_secs(1) <= crate::rpc::REPLY_TIMEOUT,
-            "two reload windows ({worst:?}) plus a second of slack must fit in {:?}",
-            crate::rpc::REPLY_TIMEOUT
-        );
-    }
-
     /// Protocol 8's routes cannot choose their own command or correlation id either, and they keep
     /// everything else the caller sent, unread.
     #[test]