It is reproducible. It has to be for their feature to work. They're trying to find some way to communicate to people that the recording is to a ring buffer, and isn't durably saving everything it hears indefinitely. It's an unenviable position to have to find a way to sell your engineering that enables a fairly privacy-preserving implementation of a feature that intuitively feels like it depends on surveillance to a mass audience. Made especially difficult by the environment where most products really are surveilling you even when they don't need to.

Of course, Apple still collects more data than it needs to and abuses users in other ways, so I don't really feel bad for them.

So the conclusion would be that if the software can be modified, it has a writable storage media attached (not necessarily local) and an image sensor and/or mic, it is, was or will be a recording device at some point in time.

Sure. It's the transcript that is stored, isn't it?

could what it produces be used by someone else (lawfully or not) to invade my personal privacy?

that’s really the question to be answered here.