In the emulator

The pictures on this site are not captures: tools/pantallas.py builds them by running in Python the cartridge's own steps. Looking at them is not enough to take them on trust, so they are compared byte for byte with the real VRAM.

How the real VRAM is taken

tools/omsx_vram.tcl lets the game run in openMSX and, at several points, dumps the 16 KB of VRAM, the eight VDP registers and the variables that say what was being drawn:

"C:/Program Files/openMSX/openmsx.exe" -machine Philips_VG_8020 \
    -cart yiear.rom -script tools/omsx_vram.tcl

It sets no breakpoints. The whole game runs inside the interrupt at 0x4043 and a bp there drowns the emulator, so the dumps go by emulated clock. It also carries a real-time watchdog: a broken script cannot hang the emulator indefinitely.

The result

Comparing the VRAM pantallas.py builds with the one openMSX holds 120 seconds into a game:

table	bytes	differing
COLOUR (0x0000–0x17FF)	6,144	0
PATTERNS (0x2000–0x37FF)	6,144	0
NAMES (0x3800–0x3AFF)	768	43
SPRITE PATTERNS (0x1800–0x1FFF)	2,048	520

The two big tables come out exact. What does not match is what the game rewrites every frame and the build has no business placing:

• in the name table, the scoreboard figures, the opponent's name and the tiles the fighter is drawn with at that instant;
• in the sprite patterns, the player's pose for that frame, redrawn over the figures the startup left there.

The buffer trap

The sprite patterns and part of the tiles are reused several times per frame. If a dump comes out "all different except one thing", the fault is not the format: it is the moment it was taken.

That is why the dump worth comparing is one taken with the fight running and settled, not one from the startup or from a transition between scenes.

A mistake only the comparison caught

The mirror loop at 0x4C46 was read at first as though the source advanced a third each turn. The renders came out plausible — a temple showed up — and only the VRAM comparison brought it down: the source does not advance, and each turn the cartridge does pop de / ld hl,00800h / add hl,de / ex de,hl / pop hl, which adds 0x800 to the destination and recovers the starting HL.

With that fixed, the pattern table went from 776 differing bytes to zero.