OK, let’s do some math.
The Massberg compressor uses coef = e^(-1/(Fs*time)). Rearrange it as time = -1/(Fs*log(coef)). Applying it gives us (att, rel) = (16.5041, 119.966) for Mac and (att, rel) = (16.9997, 122.56) for Win. The values seem to be about half the given attack and release times, so let’s multiply by 2, i.e. go from a 63% to a 86% timings definition. After rounding to integers we now have (att, rel) = (33, 240) for Mac and (att, rel) = (34, 245) for Win. In other words we found the definitions mac_coef = e^(-2/(Fs*time)) and win_coef = e^(-2/(Fs*(time + shift))) with shift_att = 1 and shift_rel = 5.
My inital question can then be rephrased:
Are these different formulas intended or a bug?
With the win_coef formula bgvw, bgva and bgvx are consistent regarding timings, so I guess mac_coef is wrong.