Control: replies are buffered, so a 17 MB reply is sent in 0.09 s instead of 34 s on Windows

The control server wrote each reply with writeln!(out, "{r}") on the
bare TcpStream. serde_json formats a Value one piece at a time, and
each piece (a brace, a quotation mark, a key, a colon, a number) was
sent as a write of its own: the 17 MB document.inspect reply of a
document with 50,000 objects on one layer was sent in 7,857,200 writes.
On Windows 11, sending that reply to a client over loopback took 34 s,
and sending the 28 MB document.json reply took 57 s.
vectorcraft-cli mcp reads every reply from a running app over this
connection.

write_line now writes each reply, each refusal and the "too many
control connections" line through a 64 KiB BufWriter, then flushes
it. The 17 MB reply takes 262 writes. Sending it now takes 0.09 s,
and sending the 28 MB reply takes 0.2 s. The bytes are the same.

Docs: docs/target-app-parity.md.

apps/vectorcraft/src/control_server.rs

@@ -7,7 +7,7 @@
 //! bounded ([`MAX_LINE`]) and so is the number of open connections ([`MAX_CONNECTIONS`]); junk
 //! never reaches the UI thread. See `docs/control-protocol.md`.
 
-use std::io::{BufRead, BufReader, Read, Write};
+use std::io::{BufRead, BufReader, BufWriter, Read, Write};
 use std::net::{TcpListener, TcpStream};
 use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
@@ -22,6 +22,8 @@ use vectorcraft_ui_egui::ControlRequest;
 pub const MAX_LINE: usize = 4 * 1024 * 1024;
 /// Connections served at once; further ones get an error line and are closed.
 pub const MAX_CONNECTIONS: usize = 16;
+/// Size in bytes of the buffer that [`write_line`] writes a reply line through.
+const REPLY_CHUNK: usize = 64 * 1024;
 
 pub fn start(port: u16, ctx: egui::Context) -> Receiver<ControlRequest> {
     let (tx, rx) = channel::<ControlRequest>();
@@ -43,7 +45,7 @@ fn accept_loop(listener: TcpListener, tx: Sender<ControlRequest>, ctx: egui::Con
         if open.fetch_add(1, Ordering::SeqCst) >= MAX_CONNECTIONS {
             open.fetch_sub(1, Ordering::SeqCst);
             // Best effort: the connection is dropped either way, and a peer that can't read the reason loses nothing.
-            let _ = writeln!(stream, "{}", json!({"ok": false, "error": "too many control connections"}));
+            let _ = write_line(&mut stream, &json!({"ok": false, "error": "too many control connections"}));
             continue;
         }
         let tx = tx.clone();
@@ -127,6 +129,13 @@ fn read_bounded_line(reader: &mut impl BufRead, buf: &mut Vec<u8>, max: usize) -
     Ok(true)
 }
 
+/// Write `v` and a newline to `out` through a buffer of [`REPLY_CHUNK`] bytes, then flush `out`.
+fn write_line(out: &mut impl Write, v: &Value) -> std::io::Result<()> {
+    let mut out = BufWriter::with_capacity(REPLY_CHUNK, out);
+    writeln!(out, "{v}")?;
+    out.flush()
+}
+
 /// Serve requests from `read`, answering on `out`, until the peer closes, sends something that is
 /// not a request, or `handle` returns `None` (app gone). `handle` gets `(method, params)` and
 /// returns the reply object (the request's `id` is added here).
@@ -145,7 +154,7 @@ fn serve_lines<R: Read, W: Write>(read: R, mut out: W, mut handle: impl FnMut(St
                     if let Some(o) = r.as_object_mut() {
                         o.insert("id".into(), id);
                     }
-                    if writeln!(out, "{r}").and_then(|()| out.flush()).is_err() {
+                    if write_line(&mut out, &r).is_err() {
                         return;
                     }
                     continue;
@@ -153,7 +162,7 @@ fn serve_lines<R: Read, W: Write>(read: R, mut out: W, mut handle: impl FnMut(St
             },
         };
         // Best effort: the connection closes either way, and a peer that can't read the reason loses nothing.
-        let _ = writeln!(out, "{}", json!({"ok": false, "error": reject})).and_then(|()| out.flush());
+        let _ = write_line(&mut out, &json!({"ok": false, "error": reject}));
         return;
     }
 }
@@ -173,6 +182,37 @@ mod tests {
         (replies, calls)
     }
 
+    /// Counts the `write` calls made on it and keeps the bytes they write.
+    #[derive(Default)]
+    struct Writes {
+        calls: usize,
+        bytes: Vec<u8>,
+    }
+
+    impl Write for Writes {
+        fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
+            self.calls += 1;
+            self.bytes.extend_from_slice(buf);
+            Ok(buf.len())
+        }
+
+        fn flush(&mut self) -> std::io::Result<()> {
+            Ok(())
+        }
+    }
+
+    /// A reply shaped like `document.inspect`'s for a layer of `n` squares in rows of 200.
+    fn inspect_reply(n: u64) -> Value {
+        let paths: Vec<Value> = (0..n)
+            .map(|k| {
+                let (x, y) = ((k % 200) as f64 * 20.0, (k / 200) as f64 * 20.0);
+                json!({"anchors": 4, "bounds": {"height": 10.0, "width": 10.0, "x": x, "y": y}, "closed": true, "fill": "#f4a261",
+                    "id": k + 2, "kind": "Path", "locked": false, "name": "<Path>", "stroke": "None", "strokeWidth": 0.0, "visible": true})
+            })
+            .collect();
+        json!({"ok": true, "result": {"layers": [{"children": paths, "id": 1, "kind": "Layer", "name": "Layer 1"}]}})
+    }
+
     #[test]
     fn requests_are_answered_in_order_with_ids() {
         let (r, calls) = run(b"{\"id\":1,\"method\":\"a\"}\n\n{\"id\":2,\"method\":\"b\",\"params\":{}}\r\n{\"method\":\"c\"}");
@@ -183,6 +223,25 @@ mod tests {
         assert_eq!(r[2]["id"], Value::Null);
     }
 
+    #[test]
+    fn replies_are_written_in_few_large_writes() {
+        for reply in [json!({"ok": true, "result": null}), inspect_reply(2_000)] {
+            let mut out = Writes::default();
+            let mut once = Some(reply.clone());
+            serve_lines(&b"{\"id\":1,\"method\":\"document.inspect\"}\n"[..], &mut out, |_, _| once.take());
+            let mut want = reply;
+            want["id"] = json!(1);
+            let text = String::from_utf8(out.bytes).unwrap();
+            // The writer receives the bytes that `writeln!(out, "{want}")` writes, in at most one write per 4 KiB.
+            assert_eq!(text, format!("{want}\n"));
+            assert!(out.calls <= text.len().div_ceil(4096), "{} writes for {} bytes", out.calls, text.len());
+        }
+        // A refusal is one write as well.
+        let mut out = Writes::default();
+        serve_lines(&b"hello\n"[..], &mut out, |_, _| None);
+        assert_eq!(out.calls, 1);
+    }
+
     #[test]
     fn http_request_closes_before_the_body_runs() {
         // A browser's cross-origin `fetch` POST: the request line is rejected and the
@@ -263,6 +322,37 @@ mod tests {
         assert!(lines[1].contains("closing the connection"));
     }
 
+    #[test]
+    fn ten_inspect_replies_for_10000_objects_arrive_within_20_s() {
+        let listener = TcpListener::bind(("127.0.0.1", 0)).unwrap();
+        let addr = listener.local_addr().unwrap();
+        let reply = inspect_reply(10_000);
+        let mut want = reply.clone();
+        want["id"] = json!(1);
+        let want = format!("{want}\n");
+        std::thread::spawn(move || {
+            let (s, _) = listener.accept().unwrap();
+            let read = s.try_clone().unwrap();
+            serve_lines(read, s, |_, _| Some(reply.clone()));
+        });
+        // The client sends each request line and reads its reply line through a `BufReader`, as `vectorcraft_mcp::Remote` does.
+        let (tx, rx) = channel();
+        std::thread::spawn(move || {
+            let mut c = TcpStream::connect(addr).unwrap();
+            let mut read = BufReader::new(c.try_clone().unwrap());
+            let mut lines = vec![];
+            for _ in 0..10 {
+                c.write_all(b"{\"id\":1,\"method\":\"document.inspect\"}\n").unwrap();
+                let mut line = String::new();
+                read.read_line(&mut line).unwrap();
+                lines.push(line);
+            }
+            tx.send(lines).ok();
+        });
+        let lines = rx.recv_timeout(Duration::from_secs(20)).unwrap_or_else(|_| panic!("ten replies did not arrive within 20 s"));
+        assert!(lines.iter().all(|l| *l == want), "a reply differs from the bytes `writeln!` writes");
+    }
+
     #[test]
     fn extra_connections_are_refused() {
         let listener = TcpListener::bind(("127.0.0.1", 0)).unwrap();

docs/target-app-parity.md

@@ -382,6 +382,7 @@ Moved from the ROADMAP's "Shipped so far" on 2026-10-10. Update it when a featur
   - Actions panel (record/playback, persisted), generic parameter dialogs for every "…" command.
   - Variables (data merge): text and visibility variables bound to objects (an object may carry several), dataset rows applied in one undo step (`variable.*`, `dataset.*`) from a Variables panel and Window › Variables, with dialogs to define, bind and edit, Make Text/Visibility Dynamic, Capture/Update Data Set, and a row in italics when an edit leaves the art not matching it; bindings let go of deleted art and locked art is left alone; every command, like the rest, is drivable over MCP and the CLI.
   - JSON-lines control channel with real egui pointer and keyboard injection.
+  - The control server writes each reply through a 64 KiB buffer. On Windows 11 it sends the 17 MB `document.inspect` reply of a document with 50,000 objects on one layer to a client over loopback in 0.09 s instead of 34 s.
   - MCP server with 25 tools (drawing, text, effects, Pathfinder, transforms, graphs, text wrap, export, screenshots, any command), attached to the running app or headless. Tool descriptions and replies agree with the engine: format lists come from its format table, `save_file` says it keeps the document's own format, `inspect_document` reports the paint type's characters show, and selection exports report the encoder's warnings. `--automation-read-root` / `--automation-write-root` (on `vectorcraft-cli mcp` and on `vectorcraft --control`) confine the files commands read and write to chosen folders, links followed, with the flags PhotoCraft takes (#832).
   - MCP protocol: five prompts (`prompts/list` / `prompts/get`: poster, icon set, recolour, trace-and-style, export set), `completion/complete` for prompt arguments and template variables read from the live catalogues (formats, effect ids, command ids, trace presets, swatch and object ids), four resource templates (`vectorcraft://object/{id}`, `//command/{id}`, `//effect/{id}`, `//swatch/{name}`) so an agent reads one thing instead of the whole document, and `logging/setLevel` with `notifications/message`. `window.panel` ids are case-insensitive (display labels work too), `vectorcraft://object/{id}` reads through `document.node {summary: true}` (the same reply bytes from one lookup instead of summarizing the whole document: a leaf read on a 2,000-object document drops from ~20 ms to ~0.06 ms), large containers slice with `document.node {depth?, childLimit?}` (truncated levels report `childCount`), `document.inspect` and the `inspect_document` tool slice the same way so a depth-0 read is the layer skeleton (~9.7 KB instead of ~2 MB on a 2,000-object document), `document.find` locates nodes by name/kind/text with ancestor paths, and the CLI sweeps the protocol surface over stdio against the compiled binary. `docs/mcp.md` has a Protocol section covering the capability table and what is deliberately absent (subscriptions, progress, cancellation, sampling, elicitation, Streamable HTTP — all of which need the transport to interleave reads).
   - Headless CLI (`vectorcraft-cli run`, `convert`, `info`, `bench`, `perf`, `mcp`). A `run` step, like a `command_batch` step over MCP, can use an earlier step's result (`"$1.id"`, #785).