I wonder if you could build some sort of deliberate reflector into a microphone housing to generate comb filtering that an analytic technique (or ML) could derive spatial data from. I know in the limit this is possible, since there's dummies used for binaural recording or testing earbuds, but could an intentionally designed mechanism accomplish the same thing in a smaller package?

It'd be neat if you could then encode the spatial information and use an HRTF on the destination receiver to appropriately place the sound in the listener's soundstage.