If you don't do this, your program may still work but you are wasting space. For instance, if the multiplication of your two 20.44 numbers does not overflow the 20.44 format, good for you, but it means that between the two operands, you were wasting 20 bits in total for leading zeroes that carried no information.
Consider floating-point as a way not to have to make all these decisions at compile-time and still use space extremely efficiently to represent as much precision as possible.
Really, what your probably asking for is decimal floating point. Which, has its own set of limitations when implemented in hardware.
And if you happen to have an application where what you really need is more precision there are a number of bignum libraries like GMP (https://gmplib.org/) which provide that. But, throwing more precision at a problem doesn't alleviate the need to do error estimation.
KSP scene has even a fun name for it, Deep Space Kraken [0].
This can be of course mitigated by subdividing the world (octrees and stuff) and stacking coordinate frames (though for some reason nobody seems to be doing that), but you said that floats are good "for arbitrary inputs". And if you have small and big floats in your computations, the result might quickly stop making any sense.
I admit I'm confused about the topic, any clarification is welcome.
[0] - http://wiki.kerbalspaceprogram.com/wiki/Deep_Space_Kraken
Fixed point won't save you from the space kraken. Instead of going so deep into space that consecutive numbers gradually become .1 apart, your ship will spontaneously explode a fiftieth of that distance out.
It's probably much easier to use two coordinate frames than to properly use fixed point numbers. Then a ship can run physics relative to its center of mass or a nearby planet/moon and have very high precision/accuracy.
There have been other approaches besides fixed that people have tried out, such as interval arithmetic and exact (rational) arithmetic, but they all have their own problems.