The headings of three tables were off by one. Those at 0x5FEF, 0x60B6 and 0x4C93 were numbered as if the index started at 1. The RAM measured at the start of each act gave it away: the 0x5C8E script shows up in the horse act, not the balls, and the frame at 0x4CC2 is the trapeze's.
Type 0 is the trapeze, not a trampoline: the trampoline is just the floor.
Two errata that made the scenery repeat.0x458F reads through DE and writes through HL, the other way round from what it looked like. That is why 0x7477 copies 0x2AC0over0x2BB0, overlapping, and the horse act's scenery repeats every 240 bytes.
The "1P" blink does not look at the number of players. The ld hl,(0xE002) brings the flags in L and the frame counter in H; the bit 5,h flips the mask every 32 frames.
The table at 0x4CEA has ten entries, not five. Two are 0x0000, and two others point at the only pieces of code no instruction reached.
0x462F repeats a pattern by stepping back: [N][eight bytes] writes that pattern N times, because the loop goes back eight before repeating. Its caller does not load its address; it carries on from the HL the previous scene left behind.
Each act's VRAM is inherited from the menu: only the name table is cleared. Building it from scratch does not give zero bytes; starting from the menu does.
Four sprites per line, and the fifth is not drawn. On the trapeze six of them cross lines 48 to 52, and the VDP drops two. Without that limit the comparison against the picture does not reach zero; with it, 49,152 dots out of 49,152.
Seven orphan bytes: they disassemble into sensible instructions, but nothing reaches them.
It is the same code as Magical Tree, shifted. The fourteen-byte signature of the decompressor loop (0x45D1 here) shows up unchanged in RC-713, and with it the text-script interpreter and the scene dispatcher.