I wonder how common were full game VMs in the 90s. For a game older than myself, wouldn't a VM layer incur a great performance penalty on PCs from that time?
The trick to these earlier VMs, from the Infocom Z-Machine and Wizardry's interpreted Pascal code, through SCUMM, Sierra AGI and SCI, Another World, the Horrorsoft games, etc., is that they recognized that the games they were making were primarily going to be "content-delivery mechanisms": lots of text and graphical assets, driven by relatively simple computations: the authoring constraint is only related to the hardware in terms of I/O and data compression. So the code that was being run by the interpreter was mostly run-once "initialize the scene" and then some animation timers.
The opposing idea is represented more by arcade gaming, and later, stuff like Doom and Quake: The game is relatively intimate with the hardware in what it simulates, while the kind of definition that makes up a scene is more on the order of "put a monster here and a health pickup there", which aligns it towards being map data, instead of scripted logic.