That's a good point. The reason for this is that the padj comes from null hypothesis testing then BH correction, while the shrinkage comes from a Bayesian procedure to get a reliable effect size, and the shrinkage procedure tends to be more aggressive at shrinking potential noise than the testing procedure. One way to think about it is that 1 out of 10 of the FDR set could have LFC=0. The shrinkage procedure will tend to put genes that are compatible with LFC=0 to 0, and it's more aggressive at this than the padj generating procedure. If you want something like a padj but that is in agreement with the shrunken estimate, we have followed the Stephens 2016 paper and added the capability to compute s-values and attach them to the results. You just need to set svalue=TRUE when you run lfcShrink, and then you will have an aggregate posterior probability to work with. Instead of bounding the rate of LFC=0, it bounds the rate of false signs (e.g. LFC is estimated as positive, but it's truly negative). I tend to use lower thresholds for svalue than for padj, so I might consider a 0.01 or a 0.005 svalue cutoff (the maximum s-value being ~0.5 corresponding to just guessing the sign of the LFC).

This means the counts for nearly all genes are compatible with LFC=0. Likely small sample size? If you want unshrunken LFC just go with the output of results().