harness: throttle bubble edits — cap 12/bubble, min 8s interval

Element X chokes loading rooms with long m.replace edit chains (the tool
bubble got an edit per tool call — dozens per turn). Throttle pushes; the
final state still flushes via the reply/closing path.

src/matrix_harness.zig

@@ -201,6 +201,8 @@ const PROGRESS_ARG_CHARS = 120;
 const PROGRESS_RET_CHARS = 400;
 /// Char cap for the bubble body (edits re-send the full body every time).
 const PROGRESS_MAX_CHARS = 3500;
+/// Edit throttling: Element X chokes on long m.replace edit chains.
+const PROGRESS_MAX_EDITS = 12;
 
 const Progress = struct {
     alloc: std.mem.Allocator,
@@ -212,6 +214,8 @@ const Progress = struct {
     reply_event_id: ?[]u8 = null, // the streamed-reply message (edited live)
     reply: std.ArrayListUnmanaged(u8) = .empty, // accumulated reply text
     replied: bool = false, // true once the full reply was edited in
+    last_edit_ms: i64 = 0,
+    edits_done: usize = 0,
 
     fn deinit(self: *Progress) void {
         if (self.event_id) |e| self.alloc.free(e);
@@ -309,6 +313,14 @@ const Progress = struct {
     }
 
     fn push(self: *Progress) void {
+        // Throttle: Element X chokes on long m.replace edit chains, so cap
+        // total edits per bubble and rate-limit to ~1 per 8 seconds. The
+        // final state is always flushed by the caller (flush()).
+        const now_ms: i64 = @intCast(@divTrunc(std.Io.Clock.now(.real, self.io).nanoseconds, std.time.ns_per_ms));
+        if (self.edits_done >= PROGRESS_MAX_EDITS) return;
+        if (now_ms - self.last_edit_ms < 8000 and self.edits_done > 0) return;
+        self.last_edit_ms = now_ms;
+        self.edits_done += 1;
         const id = self.event_id orelse {
             std.debug.print("DBG push: no event_id, blocks={d}\n", .{self.blocks.items.len});
             return;