That’s a common case, which I even mention here: schools.rst · GitHub (linked in this thread: Can one design coding rules to follow so that `-ffast-math` is safe?).
You CAN do it your way, but then you have to go the extra length to ensure things behave correctly bit to bit, which I call the IEEE school of thought. This approach has pros and cons. If you want to use the other approach, then just adjust your epsilon to be larger than just nearest spacing, and things should be robust and fast (and accurate enough).