A possible optimization would be to change internal representation on-the-fly for long-ish strings as soon as random accesses are observed. Guidance from experi

A possible optimization would be to change internal representation on-the-fly for long-ish strings as soon as random accesses are observed. Guidance from experiments would be required to tell where the right tresholds are. Also JavaScript implementations already do internal conversions between string implementations.