fix(devices): USB truncation mode 2 — close() kills the in-flight usblp URB

The chunked-write fix (81bc2e3) still truncated on hardware: the lab
test slip stopped mid-sentence with no feed and no cut (text hidden
until the feed button). Verified against drivers/usb/class/usblp.c:

- write() returns at URB SUBMISSION, not completion;
- only ONE write URB is in flight (the next write EAGAINs until it
  completes);
- usblp_release() — our close() — KILLS in-flight URBs.

The printer drains bulk data at PRINT speed (tiny internal buffer on
these clones), so closing right after the last accepted write cancels
the still-transferring tail — exactly where the feed + GS V cut bytes
live. Kernel-accepted ≠ printer-received.

Fix: the one-URB rule makes acceptance of write N a completion
certificate for write N−1. writeAllUsb now writes the payload's FINAL
BYTE alone — its acceptance proves everything before it is physically
in the printer — then drains 300 ms for that single packet before the
caller closes. New test pins the final-byte-alone chunking; wiki
printer-usb-transport.md carries the kernel-level account.

Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V

packages/devices/src/drivers/printer-escpos.test.ts

@@ -185,11 +185,12 @@ describe("stamp (Albanian date format)", () => {
   });
 });
 
-describe("writeAllUsb — the char device's partial writes / EAGAIN (field bug 2026-07-06)", () => {
-  // A NONBLOCK usblp fd accepts only what fits the printer's USB buffer per write.
-  // The old single-write path dropped everything past the first buffer — the ICS
-  // XP-K200L printed the ticket's text head but lost the barcode and the cut. A
-  // regular file can't reproduce that, so these drive the loop with a fake handle.
+describe("writeAllUsb — partial writes / EAGAIN / close-cancel (field bugs 2026-07-06/07)", () => {
+  // A NONBLOCK usblp fd accepts only what fits the printer's USB buffer per write,
+  // write() returns at URB submission, and close() KILLS the in-flight URB — so the
+  // loop must deliver every byte AND certify delivery before the caller may close
+  // (final byte written alone; its acceptance proves all prior bytes landed). A
+  // regular file can't reproduce any of that, so these drive a fake handle.
 
   /** Accepts at most `cap` bytes per call; records everything accepted in order. */
   function slowHandle(cap: number) {
@@ -207,10 +208,18 @@ describe("writeAllUsb — the char device's partial writes / EAGAIN (field bug 2
   it("delivers the WHOLE payload across many short writes (barcode + cut included)", async () => {
     const payload = renderTicket({ ticketId: "12345678901", issuedAt: "2026-06-21T10:00:00.000Z" });
     const h = slowHandle(100); // way smaller than the job → many partial writes
-    await writeAllUsb(h, payload, Date.now() + 2000);
+    await writeAllUsb(h, payload, Date.now() + 2000, 5);
     expect(Buffer.concat(h.chunks).equals(payload)).toBe(true);
   });
 
+  it("the FINAL byte is written alone — the delivery certificate before close", async () => {
+    const payload = Buffer.from("x".repeat(5000)); // > one 4K chunk
+    const h = slowHandle(100_000); // accepts anything → chunking is ours, not the cap's
+    await writeAllUsb(h, payload, Date.now() + 2000, 5);
+    expect(Buffer.concat(h.chunks).equals(payload)).toBe(true);
+    expect(h.chunks.at(-1)!.length).toBe(1); // usblp: its acceptance proves the rest landed
+  });
+
   it("retries EAGAIN (buffer full) until the kernel accepts the rest", async () => {
     const payload = Buffer.from("x".repeat(300));
     let calls = 0;
@@ -228,7 +237,7 @@ describe("writeAllUsb — the char device's partial writes / EAGAIN (field bug 2
         return Promise.resolve({ bytesWritten: n });
       },
     };
-    await writeAllUsb(h, payload, Date.now() + 2000);
+    await writeAllUsb(h, payload, Date.now() + 2000, 5);
     expect(Buffer.concat(accepted).equals(payload)).toBe(true);
   });
 

packages/devices/src/drivers/printer-escpos.ts

@@ -604,6 +604,12 @@ function withTimeout<T>(p: Promise<T>, ms: number, msg: string): Promise<T> {
  *  so jobs are pushed in chunks safely under that. */
 const USB_WRITE_CHUNK = 4096;
 
+/** Pause after the FINAL byte's write is accepted, before close. Its acceptance
+ *  proves everything before it is physically in the printer (see writeAllUsb); this
+ *  covers the one-byte URB still in flight — a single bulk packet the printer ACKs
+ *  immediately (it just freed buffer space by ACKing the previous chunk). */
+const USB_DRAIN_MS = 300;
+
 const delay = (ms: number) => new Promise<void>((r) => setTimeout(r, ms));
 
 /** The slice of FileHandle the USB write loop needs (injectable for tests — a real
@@ -613,31 +619,45 @@ export interface UsbWriteHandle {
 }
 
 /**
- * Push the WHOLE payload through a non-blocking usblp fd. On O_NONBLOCK the kernel
- * takes only what fits the printer's USB buffer and returns a SHORT write (or EAGAIN
- * when full) — a single fire-and-forget write() silently drops the tail of any job
- * bigger than one buffer. That was a real field bug (2026-07-06, ICS XP-K200L over
- * USB): the text head printed, but the barcode mid-payload and the CUT at the end
- * were in the dropped tail — "prints, but no barcode and no cut", while the same
- * bytes over TCP were fine. So: loop until every byte is accepted, retrying EAGAIN
- * and zero-byte writes with a short pause, bounded by the caller's deadline.
+ * Push the WHOLE payload through a non-blocking usblp fd AND ensure the printer has
+ * physically received it before the caller may close. TWO field-verified truncation
+ * modes on the ICS XP-K200L (same symptom: text head prints, barcode/feed/CUT tail
+ * lost; TCP fine):
+ *
+ * 1. SHORT WRITES (2026-07-06): a single fire-and-forget write() only delivers what
+ *    the kernel accepts. Fix: chunked loop, retry EAGAIN, until all bytes accepted.
+ * 2. CLOSE CANCELS THE LAST TRANSFER (2026-07-07, lab bench): per usblp.c, write()
+ *    returns at URB *submission*, only ONE write URB is in flight at a time, and
+ *    usblp_release() (our close) KILLS in-flight URBs. The printer consumes bulk
+ *    data at PRINT speed (tiny internal buffer), so closing right after the last
+ *    accepted write cancels the still-transferring tail — which is exactly where
+ *    the feed + GS V cut live ("have to press the feed button to see the text").
+ *
+ * The delivery guarantee follows from usblp's one-URB rule: ACCEPTANCE OF WRITE N
+ * PROVES WRITE N−1 FULLY COMPLETED (the driver EAGAINs until the previous URB's
+ * completion). So the payload is pushed as chunks, then its FINAL BYTE alone: when
+ * that 1-byte write is accepted, every byte before it is physically in the printer.
+ * A short drain pause then covers the lone final-byte URB (one bulk packet), and
+ * close is safe. `drainMs` is parameterised only for tests.
  */
 export async function writeAllUsb(
   handle: UsbWriteHandle,
   payload: Buffer,
   deadlineMs: number,
+  drainMs: number = USB_DRAIN_MS,
 ): Promise<void> {
+  if (payload.length === 0) return;
+  const lastByteAt = payload.length - 1;
   let off = 0;
   while (off < payload.length) {
     if (Date.now() > deadlineMs) {
       throw new Error(`usb write timeout (${off}/${payload.length} bytes accepted)`);
     }
     try {
-      const { bytesWritten } = await handle.write(
-        payload,
-        off,
-        Math.min(USB_WRITE_CHUNK, payload.length - off),
-      );
+      // Never let the final byte ride a bigger chunk: it is written ALONE so its
+      // acceptance certifies delivery of everything before it (see doc above).
+      const len = off === lastByteAt ? 1 : Math.min(USB_WRITE_CHUNK, lastByteAt - off);
+      const { bytesWritten } = await handle.write(payload, off, len);
       off += bytesWritten;
       if (bytesWritten === 0) await delay(5); // buffer full, no error — breathe
     } catch (err) {
@@ -648,6 +668,9 @@ export async function writeAllUsb(
       }
     }
   }
+  // All bytes accepted; only the 1-byte final URB can still be in flight. Give it a
+  // moment to land before the caller closes (close would cancel it).
+  await delay(drainMs);
 }
 
 /** Write an ESC/POS payload to a USB-lp character device (e.g. /dev/usb/lp0). usblp

wiki/concepts/printer-usb-transport.md

@@ -104,6 +104,23 @@ deadline with a `(N/M bytes accepted)` diagnostic. Driven by fake-handle tests (
 EAGAIN interleave, wedged-printer timeout, non-EAGAIN passthrough) since a real file can't
 reproduce the char device's behaviour.
 
+**Second truncation mode — close() cancels the in-flight transfer (lab bench, 2026-07-07).** The
+chunked loop alone STILL truncated on hardware (test slip stopped mid-sentence, no feed, no cut —
+"press the feed button to see the text"). Verified against `drivers/usb/class/usblp.c`: `write()`
+returns at URB *submission* (not completion), only ONE write URB is in flight at a time (the next
+write EAGAINs until it completes), and `usblp_release()` — i.e. our `close()` — **kills in-flight
+URBs**. The printer drains bulk data at PRINT speed (tiny internal buffer on these clones), so
+closing right after the last accepted write cancels the still-transferring tail — exactly where
+the feed + `GS V` cut bytes live. Kernel-accepted ≠ printer-received.
+
+Fix: the one-URB rule makes acceptance of write N a **completion certificate for write N−1**. So
+`writeAllUsb` now writes the payload's FINAL BYTE alone: when that 1-byte write is accepted, every
+byte before it is physically in the printer; a short drain pause (`USB_DRAIN_MS` 300 ms) covers
+the lone final-byte packet, then close is safe. (usblp also implements `poll(POLLOUT)` as the true
+completion signal, but Node cannot poll an arbitrary char-device fd without a native dep — the
+hold-back + drain gets the same guarantee for all but the final byte, whose packet the printer
+ACKs immediately after having just freed its buffer.)
+
 > Driver-choice note for this clone: the ICS XP-K200L does NOT serve the Rongta `/prn_stat.htm`
 > status page (checked on hardware at 10.0.10.11 — print socket 9100 open, status page absent),
 > so on NETWORK the honest driver is **cashino** (reachability-only monitoring); under `rongta`

wiki/log.md

@@ -2485,3 +2485,14 @@ Lab bench (USB printer test, no relays on hand) hit a SECOND printer/relay coupl
 blocks every non-access category while zero controllers exist. Printers are now exempt there too
 — the binding fix removed the requirement inside the form; this removes the gate in front of it.
 A controller-less box can configure + test a printer.
+
+## [2026-07-07] update | USB truncation, mode 2: close() kills the in-flight usblp URB
+
+Lab hardware test of the chunked-write fix STILL truncated (slip stopped mid-sentence, no cut,
+text hidden until the feed button). Root cause verified against kernel usblp.c: write() returns at
+URB submission; one URB in flight; usblp_release (close) kills it; the printer drains at print
+speed — so the accepted-but-untransferred tail (incl. feed+cut, always the last bytes) died at
+close. writeAllUsb now holds back the FINAL byte as its own write — usblp's one-URB rule makes its
+acceptance a completion certificate for everything before it — then drains 300ms for that single
+packet before close. Tests updated (+ final-byte-alone assertion). [[printer-usb-transport]] has
+the full kernel-level account.