You mentioned 8 days, but your model is for 4 lags (time window of 5: current and previous 4). Is that what you want. You specified a cubic polynomial for the parameter constraints, which is just one degree less than the maximum (i.e., with no constraints). I think you are not gaining the advantages of PDL regression. With only 46 observations, I think you have an over-parameterized model. I have used PDL regression a lot, and you need a lot of data points to reliably estimate the parameters and determine the time length and polynomial order.
Your nlag option is fine for a 1-st order AR residual term.
Your seasonality term is probably OK, but it doesn't give you much flexibility. You might have to look into using a smoothing spline, but this can't be done within PDLREG. You would have to do this in another procedure, store the residuals, and then use PDLREG on the residuals. I suggest you check out the article by Schwartz (Epidemiology 11: 320-326 [2000]). Also check out Madden and Paul (Phytopathology 100: 1015-1029 [2010]).