The fix is sensible if Unicode is to be supported but it worsens the time complexity since .at is linear time in the size of the base string rather than constant time. Consider extracting the Unicode characters into an array as "preprocessing".

I'm not convinced that's true — surely accessing an array index is constant time, just like accessing an object property. There's no reason for at to be iterating over the array, as arr.at(n) acts identically to arr[n] for non-negative integer values of n. Maybe I should just swap those out for square-bracket accessors as they'll never be negative anyway.

The only linear time parts I added were these:

 const baseOneCharacters = [...baseOneCharacterString]
 const baseTwoCharacters = [...baseTwoCharacterString]

But they only run once per function call, and in any case, the first argument is already iterated prior to being reversed.

@appgurueu Should I wait for the changes you suggested or good to merge?

Changing baseTwoCharacters.at(remainder) to baseTwoCharacters[remainder] uncovered a performance bug — value was often not an integer, because the division result wasn't being floored. charAt and at both silently floor their arguments, which obscured the bug:

'abc'.charAt(1.5) // b
;[...'abc'].at(1.5) // b

However, as a result, the loop was being iterated over many more times than necessary as value < 1 got repeatedly divided until it reached floating-point 0, with all the extra zero-characters finally being lopped off at the end with .replace(new RegExp(`^${baseTwoZero}+`), ''). Without that regex replacement, convertArbitraryBase('98', '0123456789', '01234567') had 358 leading zeros (i.e. 358 unnecessary iterations).

I've fixed that by using floor division and removing the regex replacement.