as for coefficients: I think FEG values (both filtering coeff and Q) uses the same dB logarithmic volumes as AEG but with 3 more LSB added. So in short - Neill was one bit wrong, it have 9 fractional bits and 4 exponential (not 8 and 5). In the case of Q it must be <<=6 so it become in same dB range as FEG level. this assumptions also kind of approved by old AICA implementation http://sabia.tic.udc.es/gc/Contenidos%20...CA_E.HTM#no6_17 , when at some stage of development Q register was full 13 bits, not just 5bits as in final version.
But... there is some problem in algorithm. by design it makes sound a bit off/quiet (then C 1 and Q 0 output becomes = input minus previous output), so with Feg_Lv 0x1FF8 and Q 4 output is not the same as input.
is there anyone good at DSP processing and may give advice to look at some another filtering algorithms/formulas ?