I'm a bit confused, I thought any string with only lowercase letters was "valid base64" (more precisely, I thought "valid base64" is equivalent to "string consists only of the 64 special characters we're using to represent digits 0-63").
It gets more complicated if you're using base64 to represent a number of bytes that isn't a multiple of 3, those would be unlikely to happen randomly. Those will usually end in a number of '=' signs to pad their length to a multiple of 4 and indicate how many bytes are missing. Although apparently there are also versions o base64 that don't include the padding.
As an example, Elixir's stdlib includes a base64 decoder; passing "padding: false" will give the "decode anything" behavior you're describing.
Try running `base64 -d <<< a` and it will fail, but `base64 -d <<< aaaa` works (4*6 == 24 bits, which is divisible by 8 and gives 3 bytes of output)