Creating and mounting a loop-volume within the HGFS and then running NPM in there, would probably also be helpful. It certainly is when wrangling a directory with lots of tiny inodes over SMB. (It’s also what Apple themselves do for running Time Machine over SMB: they create a sparsebundle image on the remote and mount it locally, and write to that.)

(The efficient thing about this class of solutions, if you’re wondering, is that the client mounting the loop-image ends up owning + managing the loop-image’s internal filesystem’s metadata within its own local disk cache, such that it can coalesce filesystem metadata writes and only push a new copy of the disk-image blocks backing the filesystem metadata after a potentially-huge number of changes. With a network/host-guest filesystem, meanwhile, every filesystem metadata change must become its own synchronous message to the host, to be pushed to the host’s filesystem driver for linearization, so it can succeed or fail relative to other things going on within the host.)