I am not sure the wavelength calibration of the MILES stars is all that accurate. I remember seeing some apparent discrepancies when I was trying to do some radial velocity work. Do your individual errors for the lamp lines look good particularly those bracketing the H alpha line? (from the report that ISIS generates in the “Go” window. )
I would not worry about adjusting the Grism further if ISIS is finding the lines to do the wavelength calibration OK (The RMS error should be low, <0.5A and there should be no obvious larger errors in any individual line as reported by ISIS)
The kink in the instrument response of the ALPY around 4000A is real and often seen, though the cause is not clear. See my recent post here for example.