https://github.com/gbmhunter/MFixedPoint looks like another potential option, but still has that performance penalty. I think that's an inevitability at this point. But if the game is optimized well enough then I guess the 5x to 10x performance penalty is worth it, it's not like we're in 3D. Potentially we'll reach playability on 020 / 386 level hardware and still have decent ish performance going up to the 060 / Pentium 1. I'll put together a stress testing environment to determine if this is the case.
If I do end up having to make both an FPU and fixed-point version of the engine, as a worst case scenario, then the demo record will need to contain some flags about the settings the engine was built with when recording the game session.