An open's writes in a row to one place go in as one edit: a 10 MB body write is linear

Each body or data write copied and hashed the whole buffer, so a write the mount cut in 8 KB pieces was quadratic: 10 MB took 50 s and a 1 MB insert into 10 MB 11 s, holding the editor's turn. An open's appends to body, or inserts going on at data's address, are now held and put in as one splice (one copy, one undo step, one line-starts pass) before any other request, the close, or the editor's step once the writes pause 20 ms. fs.py-driven: 2 MB 2.19 -> 0.16 s, 10 MB 50.17 -> 0.78 s, 1 MB data into 10 MB 11.49 -> 0.26 s. A body-2m case (2 MB in 256 KB writes on one open) joins the perf gate: 84023 -> 15785 us; baselines re-recorded.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

1 files changed, 3 insertions, 0 deletions

diff --git a/src/ninep/tree.zig b/src/ninep/tree.zig
index 0960f00b..6d09b9a6 100644
--- a/src/ninep/tree.zig
+++ b/src/ninep/tree.zig
@@ -454,6 +454,9 @@ pub fn stagedReply(p: *Pardes, req: Req) Reply {
 // ---- dispatch ----
 
 pub fn handle(p: *Pardes, req: Req) Reply {
+    // Held writes go in before anything but the next write of their open.
+    if (p.fs.batch.bytes.items.len > 0 and !(req.op == .write and req.node == p.fs.batch.node and req.handle == p.fs.batch.handle))
+        pane.flushBatch(p);
     if (req.op == .open and req.create) {
         if (Node.target(req.node)) |t| if (t == .pane and p.paneBySerial(t.pane.serial) == null)
             return Reply.fail(req.tag, E.NOENT);