Static binary translation without heuristics—how practical is it really?

There's been interesting progress lately in the field of static binary translation, with some researchers claiming they can translate entire binaries deterministically without relying on heuristics. The idea is pretty appealing: convert compiled code from one architecture to another in a fully predictable, reproducible way.

But I'm curious what the real-world trade-offs look like. Traditional binary translation often uses heuristics to guess where code boundaries are, identify data vs. instructions, and handle indirect jumps. If you remove those entirely, you gain certainty and reproducibility—but what do you lose? Does the output code run faster or slower? Is the translation time reasonable for large binaries?

I imagine this approach might work beautifully for simple, statically-compiled binaries, but what about more complex real-world scenarios? Position-independent code, obfuscation, dynamically linked libraries, or binaries with embedded data—are these edge cases handled elegantly or do they fall outside the scope?

Also, I'm wondering about the practical applications. Who actually needs this? Reverse engineering? Legacy system migration? Security auditing? Are there specific use cases where deterministic translation matters more than speed or accuracy?

Would love to hear from anyone working with binary translation tools, whether you've encountered heuristic-based limitations firsthand or if you see the appeal of a fully deterministic approach.