The S3 's are widely all over the place, almost impossible to get few "quad matched sets" to do a Lightspeed properly with close to zero volume low volume, unless they are forced matched with servo feedback arrangement and by over driving some which "raises the distortion substantially" to meet the others which will kill them over a very short period.
Actually the SR3 are better in my experience than the SR2S for matching. There is no need either, to use servo feedback arrangements.
Keeping current 3ma for the whole circuit is the key. Sweeping between lowest volume and highest should achieve variance of just 2.78ma to 3ma , for the entire circuit, encompassing the load of 4x NSL32SR3
There is no need to go any higher at all, which keeps distortion extremely low, and achieves much better stability with cell resistance.
R(on) of 20ma per device, or anywhere near that figure, is thus neatly avoided .