docs: correct the tile-stream wire overhead and test counts

harness/mame/NETWORK_REPORT.md

@@ -5098,7 +5098,7 @@ records what was run, on what, and — in the same breath — what was **not** r
 
 Three things were run:
 
-1. the host tests for the pipeline (`harness/udpstream/tests/`, 137 passing);
+1. the host tests for the pipeline (`harness/udpstream/tests/`, 139 passing);
 2. `make_tiles.py` over a deterministic 30-frame clip, producing the committed
    cache `harness/udpstream/tiles/` (dictionary + index/delta stream);
 3. the **MAME replay**: the same cache played through the payload's own RX ring
@@ -5235,7 +5235,7 @@ tile_check: PASS
 ```
 
 Each clip frame is delivered as the datagrams the sender would send (a keyframe
-or delta is one datagram at 1,450 bytes), the harness waits for the ring to
+or delta is one datagram — the full frame's VS payload is 1,296 B, 1,310 B on the wire; 1,450 B is only the maximum chunk), the harness waits for the ring to
 consume it, and then dumps the 160 active bytes of every screen row to
 `vram-replay-<frame>.bin` (1,024 x 160 = 163,840 B a frame). The checker
 decodes the stream independently, renders each frame through the dictionary and
@@ -5308,7 +5308,7 @@ RESULT DRY-RUN: 6300 frames (105 full, 6195 delta, 0 unchanged),
 
 that is **105 keyframes and 606 KB for 3.5 minutes**, 96 B/frame, which is the
 host-side budget the soak needs to beat on the wire (FCS, preamble and IFG will
-add ~7%).
+add about a quarter: 4 (FCS) + 8 (preamble) + 12 (IFG) = 24 B per datagram, and 6,323 datagrams carry 606,384 B of payload, so ~151,752 B of framing on top — the wire total is closer to 758 KB.)
 
 ### 76.7 The real clip: the command, untested here