I'm pretty hesitant about this approach - it adds a lot of complexity to serialization (which is already overdrawn on the complexity budget) for a relatively narrow use case, and it also seems likely to be a significant performance hit because of the double serialization required for the length calculation.

I'm assuming your current approach is to use bson::to_vec and then write that into the destination buffer via bytes::BufMut::put - has profiling shown the allocation or copy there to be substantial cost?