feat(sessions): stream in-flight partial when joining a mid-turn session

Seed the in-flight partial assistant message (text, thinking, and
in-progress tool calls) when opening or reconnecting to a session that is
mid-stream, then continue streaming live deltas on top of it. Replaces the
blocking "Catching up..." placeholder and the end-of-turn transcript reload.

Server stamps every per-session UI event with a monotonic seq at the
SessionEventHub publish choke point and exposes
GET /sessions/:sessionId/stream-snapshot returning { seq, partial }. The
client fetches the snapshot on join, seeds the normalized partial into the
in-memory transcript (never the history cache), and applies buffered/live
events using the seq watermark for exactly-once delivery.

The snapshot is a progressive enhancement: a 404 from an older remote
pi-web or a not-yet-restarted session daemon falls back to an empty seed
(seq 0, drops nothing), so sessions still open and stream normally. The
stream-snapshot route is registered in the federation allowlist for
remote-machine proxying.

src/server/realtime/sessionEventHub.test.ts

@@ -18,7 +18,7 @@ describe("SessionEventHub", () => {
 
     hub.publish("s1", { type: "assistant.delta", text: "hello" });
 
-    expect(sessionSocket.send).toHaveBeenCalledWith(JSON.stringify({ type: "assistant.delta", text: "hello" }));
+    expect(sessionSocket.send).toHaveBeenCalledWith(JSON.stringify({ type: "assistant.delta", text: "hello", seq: 1 }));
     expect(otherSocket.send).not.toHaveBeenCalled();
   });
 
@@ -34,6 +34,7 @@ describe("SessionEventHub", () => {
     expect(socket.send).toHaveBeenCalledWith(JSON.stringify({
       type: "message.end",
       message: { role: "assistant", content: [{ type: "thinking", thinking: "private chain", redacted: true }, { type: "text", text: "visible answer" }] },
+      seq: 1,
     }));
     expect(thinkingBlock.thinkingSignature).toBe("opaque-provider-payload");
   });
@@ -76,4 +77,65 @@ describe("SessionEventHub", () => {
     expect(globalSocket.send).toHaveBeenCalledWith(JSON.stringify({ type: "status.update", status }));
     expect(sessionSocket.send).not.toHaveBeenCalled();
   });
+
+  it("stamps a monotonically increasing per-session seq on published events", () => {
+    const hub = new SessionEventHub();
+    const socket = new FakeSocket();
+    hub.add("s1", socket);
+
+    hub.publish("s1", { type: "assistant.delta", text: "a" });
+    hub.publish("s1", { type: "assistant.delta", text: "b" });
+    hub.publish("s1", { type: "assistant.delta", text: "c" });
+
+    expect(socket.send).toHaveBeenNthCalledWith(1, JSON.stringify({ type: "assistant.delta", text: "a", seq: 1 }));
+    expect(socket.send).toHaveBeenNthCalledWith(2, JSON.stringify({ type: "assistant.delta", text: "b", seq: 2 }));
+    expect(socket.send).toHaveBeenNthCalledWith(3, JSON.stringify({ type: "assistant.delta", text: "c", seq: 3 }));
+  });
+
+  it("advances seq even when no sockets are attached so the watermark stays accurate", () => {
+    const hub = new SessionEventHub();
+
+    hub.publish("s1", { type: "assistant.delta", text: "a" });
+    hub.publish("s1", { type: "assistant.delta", text: "b" });
+
+    expect(hub.currentSeq("s1")).toBe(2);
+
+    const socket = new FakeSocket();
+    hub.add("s1", socket);
+    hub.publish("s1", { type: "assistant.delta", text: "c" });
+
+    expect(socket.send).toHaveBeenCalledWith(JSON.stringify({ type: "assistant.delta", text: "c", seq: 3 }));
+  });
+
+  it("tracks seq independently per session", () => {
+    const hub = new SessionEventHub();
+    const s1 = new FakeSocket();
+    const s2 = new FakeSocket();
+    hub.add("s1", s1);
+    hub.add("s2", s2);
+
+    hub.publish("s1", { type: "assistant.delta", text: "a" });
+    hub.publish("s1", { type: "assistant.delta", text: "b" });
+    hub.publish("s2", { type: "assistant.delta", text: "x" });
+
+    expect(hub.currentSeq("s1")).toBe(2);
+    expect(hub.currentSeq("s2")).toBe(1);
+    expect(s1.send).toHaveBeenLastCalledWith(JSON.stringify({ type: "assistant.delta", text: "b", seq: 2 }));
+    expect(s2.send).toHaveBeenCalledWith(JSON.stringify({ type: "assistant.delta", text: "x", seq: 1 }));
+  });
+
+  it("reports zero seq for a session that has never published", () => {
+    const hub = new SessionEventHub();
+    expect(hub.currentSeq("never")).toBe(0);
+  });
+
+  it("does not stamp seq on global events", () => {
+    const hub = new SessionEventHub();
+    const globalSocket = new FakeSocket();
+    hub.addGlobal(globalSocket);
+
+    hub.publishGlobal({ type: "session.name", sessionId: "s1", name: "Renamed" });
+
+    expect(globalSocket.send).toHaveBeenCalledWith(JSON.stringify({ type: "session.name", sessionId: "s1", name: "Renamed" }));
+  });
 });

src/server/realtime/sessionEventHub.ts

@@ -11,6 +11,7 @@ export interface RealtimeSocket {
 export class SessionEventHub {
   private readonly socketsBySession = new Map<string, Set<RealtimeSocket>>();
   private readonly globalSockets = new Set<RealtimeSocket>();
+  private readonly seqBySession = new Map<string, number>();
 
   add(sessionId: string, socket: RealtimeSocket): void {
     let sockets = this.socketsBySession.get(sessionId);
@@ -30,12 +31,24 @@ export class SessionEventHub {
   }
 
   publish(sessionId: string, event: SessionUiEvent): void {
-    const payload = JSON.stringify(projectBrowserSessionEvent(event));
+    const seq = (this.seqBySession.get(sessionId) ?? 0) + 1;
+    this.seqBySession.set(sessionId, seq);
+    const payload = JSON.stringify({ ...projectBrowserSessionEvent(event), seq });
     for (const socket of this.socketsBySession.get(sessionId) ?? []) {
       if (socket.readyState === socket.OPEN) socket.send(payload);
     }
   }
 
+  /**
+   * Last per-session sequence number stamped by {@link publish} (0 before any
+   * event). Callers building a join-time stream snapshot read this as the
+   * watermark: buffered live events with `seq <= currentSeq` are already
+   * reflected in the snapshot's partial and must be dropped by the client.
+   */
+  currentSeq(sessionId: string): number {
+    return this.seqBySession.get(sessionId) ?? 0;
+  }
+
   publishGlobal(event: GlobalSessionEvent): void {
     this.publishRealtime(event);
   }