maybe can i suggest another approach?
the idea:
1) launch a vram fill command with one solid color at full screen. (high speed)
2) wait for a scanline begin
3) issue, with a regular interval R18 updates for exactly no more than an entire scanline, the regular interval could be 14 z80 cycles, so a bunch of out (c),a with c=0x98
4) if there is a safe period, maybe the distribution of corrupted pixels can give us some hint about the problem.
For example, if we see a even distribution for the entire scanline we know that there is no good interval to do this R18.
Maybe there could be some area where corruption do not occour opening the door to some ideas
