Sums of fixed effects not agreeing

[jacoblefler]

I'm using data from the 2017 Commodity Flow Survey to run a gravity-style regression of trade flows on log distance and origin-industry and destination-industry fixed effects.
I run ppmlhdfe wgt_obs l_dist, absorb(n_orig_naics n_dest_naics, savefe) d, where wgt_obs is observed trade flow in pounds, l_dist is log distance, and n_orig_naics n_dest_naics are origin-industry and destination-industry fixed effects.
This produces new variables _ppmlhdfe_d __hdfe1__ __hdfe2__. However, when I add __hdfe1__ and __hdfe2__ manually, I find that their sum is very close to _ppmlhdfe_d for most observations, but very far off for a few observations.