diff --git a/CLAUDE.md b/CLAUDE.md
index f07c105..22e19a3 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -246,10 +246,20 @@ Details and the recovery ladder are in `gc_bba_test/README.md`. Key points:
   veth and the frames go nowhere).
 - **Stop-and-wait is load-bearing** — see the RBFI note; one frame in flight
   means the ring is never contended.
-- `bba_tx_bytes()` streams in 64-byte chunks, ONE EXI transaction each. Writing
-  byte-by-byte cost a full EXI_Lock/Select/Deselect/Unlock per byte — ~60
-  transactions per ACK — and that latency, not any transmit failure, is what
-  made ~⅓ of blocks retry (`txto=0` on hardware: TI always set).
+- `bba_tx_bytes()` streams in **8-byte** chunks, one EXI transaction each.
+  Writing byte-by-byte cost a full EXI_Lock/Select/Deselect/Unlock per byte —
+  ~60 transactions per ACK — and that latency, not any transmit failure, is
+  what made ~⅓ of blocks retry (`txto=0` on hardware: TI always set).
+
+  ⚠️ **The chunk size is bounded by `tx_bytes`, which is 16 deep.** Within one
+  transaction the bytes arrive back-to-back at EXI speed, far faster than
+  `TXFrameDrain` starts consuming, and `BBARegisterFile` silently drops bytes
+  when `w_rdy` is low because SPI cannot back-pressure. A 64-byte chunk
+  overran the FIFO, the frame went out truncated and the host saw no valid ACK
+  at all. Byte-by-byte only worked because the per-byte EXI overhead happened
+  to pace the drain. 8 is half the FIFO depth, so it cannot overflow even if
+  the drain consumes nothing for a whole transaction. **Do not raise it
+  without deepening `tx_bytes` or giving the TX path real back-pressure.**
 
 ### ⚠️ OPEN: ring descriptor sometimes lands at page offset +1
 
diff --git a/gc_bba_test/source/bba.c b/gc_bba_test/source/bba.c
index 8640913..bd21a70 100644
--- a/gc_bba_test/source/bba.c
+++ b/gc_bba_test/source/bba.c
@@ -125,21 +125,31 @@ int bba_tx_bytes(const u8 *buf, u32 len)
 {
 	u32 off = 0;
 
-	/* Stream in 64-byte chunks, ONE EXI transaction each.
+	/* Stream in 8-byte chunks, ONE EXI transaction each.
 	 *
-	 * This used to write byte by byte with bba_w8(), which means a full
-	 * EXI_Lock / EXI_Select / header / EXI_Deselect / EXI_Unlock per BYTE --
-	 * about sixty transactions to send one 60-byte ACK.  That latency is why
-	 * roughly a third of netboot blocks were retried: the ACK was always
-	 * transmitted (txto=0 on hardware, so TI always set) but did not get back
-	 * to the host inside its timeout.  The data phase of a register-space
-	 * write runs until CS, so there is no reason to pay the per-byte cost.
+	 * Writing byte by byte with bba_w8() costs a full EXI_Lock / EXI_Select /
+	 * header / EXI_Deselect / EXI_Unlock PER BYTE -- about sixty transactions
+	 * for one 60-byte ACK.  That latency is why ~1/3 of netboot blocks were
+	 * retried: the ACK always reached the W5100 (txto=0, TI always set) but
+	 * missed the host's timeout.
 	 *
-	 * 64 is bba_write()'s internal buffer limit, not a protocol limit. */
+	 * ⚠️ BUT THE CHUNK SIZE IS BOUNDED BY tx_bytes, WHICH IS 16 DEEP.
+	 * BBARegisterFile silently drops bytes when tx_bytes.w_rdy is low,
+	 * because SPI cannot back-pressure.  Within one transaction the bytes
+	 * arrive back-to-back at EXI speed -- far faster than TXFrameDrain can
+	 * start consuming -- so a 64-byte chunk overran the FIFO, the frame went
+	 * out truncated, and the host saw no valid ACK at all ("no ACK after 20
+	 * tries").  Byte-by-byte worked only because the per-byte EXI overhead
+	 * happened to give the drain time.
+	 *
+	 * 8 is half the FIFO depth: it cannot overflow even if the drain consumes
+	 * nothing for the whole transaction, while still cutting a 60-byte ACK
+	 * from ~60 transactions to 8.  Do not raise it without deepening
+	 * tx_bytes or giving the TX path real back-pressure. */
 	while (off < len) {
 		u32 n = len - off;
 		int rc;
-		if (n > 64) n = 64;
+		if (n > 8) n = 8;
 		rc = bba_write(REG_TXDATA, buf + off, n);
 		if (rc < 0) return rc;
 		off += n;