Why restrict to 54-bit signed integers?

Because number is a double, and IEEE754 specifies the mantissa of double-precision numbers as 53bits+sign.

Meaning, it's the highest integer precision that a double-precision object can express.

I suppose that makes sense for maximum compatibility, but feels gross if we’re already identifying value types.

It's not about compatibility. It's because JSON only has a number type which covers both floating point and integers, and number is implemented as a double-precision value. If you have to express integers with a double-precision type, when you go beyond 53bits you will start to experience loss of precision, which goes completely against the notion of an integer.