spec+plan: header-only frame CRC (the body is covered by the Ethernet FCS)

docs/superpowers/plans/2026-10-30-vt1200-stream-transport-and-row-strips.md

@@ -286,7 +286,18 @@ git commit -m "feat(stream): frame parser shared by host tests and payload"
 - Modify: `harness/udpstream/vidstream.c`, `harness/mame/vidstream.lua`
 
 **Interfaces:**
-- Produces: `vram_blit_b_(src_bitaddr, daddr_xy, dydx)` — `PIXBLT B,XY`, `SPTCH=0`, `PSIZE=1`, `CONVSP/CONVDP=0x14`, `CONTROL` raster op 0; reloads `B2`/`B7` per call.
+- Produces: `blit_row(src_bitaddr, daddr_xy, dydx)` — the C wrapper for `PIXBLT B,XY`, `SPTCH=0`,
+  `PSIZE=1`, `CONVSP/CONVDP=0x14`, `CONTROL` raster op 0, `B8=0`/`B9=0xFFFF` for the binary source;
+  reloads `B0/B1/B2/B3/B4/B7/B8/B9` per call.
+- **Ruling R13 (implemented):** the asm symbol takes **no arguments** — the 1987 compiler's frame
+  layout is undocumented here, so values are published in `.bss` words and the asm reloads them.
+  `blit_row()` is the only entry point; Tasks 6/7 call the wrapper, never the symbol.
+
+**Ruling R14 (CRC scope):** the frame CRC covers the **header only** (bytes 0..11). The body's
+integrity is the Ethernet FCS the LANCE already validates per packet, and truncation is caught by
+`payload_len` + descriptor count. Measured evidence: a payload-wide CRC costs ~139 instructions per
+byte and alone capped the frame rate (~1 fps), violating the design's no-per-byte-CPU rule.
+`frame.py`, `frame_parse.c` and `fp_parse` must all agree on header-only.
 
 - [ ] **Step 1: Write the failing test**