Do you have a reference for the claim that realloc() is unusable for the reason you say?

Many users of floating-point are not using floating-point in hard real-time contexts. Even soft real-time contexts can be fine with Ziv-style successive approximations. Correctly rounded library functions have intangible benefits beyond precision (portability, for instance), but if you don't value these benefits, you can always use inaccurate and always fast functions. You don't even need to invent your own: just rip off the first stage of a multi-stage Ziv implementation.

I don't think any of this is the reason why “people don't like doing floating-point calculations”.