coco_mem_fetch.sv compared read_count against WORD_END_COUNT on the same
114 MHz edge that took SDRAM_VID_ACK and updated VID_READ_OFFSET, the fetch
address; that path missed timing (-0.7 ns). read_count only changes on the
ack, so the compare is now registered in state_start with identical results.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>