Consider the following pipeline:

[
  {$documents: [{a: [1, 2, 3]}]},
  {$project: {"a.b": 1}},
  {$addFields: {"a.c": "why?"}}
]

Inclusion $project on a sub-field of an array will remove all scalar values, making array “a” empty.
$addFields on a sub-field of an array will add

to every element of the array, but array is empty, so the expected answer would be

[{a: []}]

However, we get

{ a: [ { c: 'why?' }, { c: 'why?' }, { c: 'why?' } ] }

because $project actually doesn't remove scalars, but replacing them with MISSING values.

This is a bug. Especially given it behaves inconsistently with reading from disk,
because this pipeline will return correct result.
Consider a collection test with a single document

db.test.aggregate([{$project: {_id: 0, "a.b": 1}}, {$addFields: {"a.c": "why?"}}]).

It will return correctly, because if Document has a backing BSONObj, it will use fast-path projection that actually removes elements from the array.