MongoDB $match stage after $replaceRoot - mongodb

Question

Consider the following.
I have documents in which I save versions of fields.
Fields in the original document can null and I only save fields that actually changed.
For example:
{
_id : abc,
state : 'NEW',
field2 : 'WILL NOT CHANGE',
field3 : null,
lastChange : 1236547
changes : [
{
state : 'WORKING',
field3 : 'SOMETHING'
},
{
state : 'DONE'
}
]
}
In my aggregation pipeline I build the latest version of the document by merging all versions with the document root. This means I avoid complex querys in which I search the changes array and find the latest version of every field I want to find and I benefit from the side effect that the pipeline will return the latest document version as a document. The pipeline below ends at the point where my problem is. In the application there are more stages following that do counting and grouping etc.
Mongoose: meldungs.aggregate([
{
'$match': {
'$and': [{
lastChange: {
'$ne': null
}
},
{
lastChange: {
'$gte': ISODate('2017-03-01T23:00:00.000Z')
}
},
{
lastChange: {
'$lt': ISODate('2017-03-02T22:59:59.999Z')
}
}]
}
},
{
'$replaceRoot': {
newRoot: {
'$arrayToObject': {
'$reduce': {
input: '$changes',
initialValue: {
'$filter': {
input: {
'$objectToArray': '$$ROOT'
},
as: 'field',
cond: {
'$and': [
{ $ne : [ '$$field.v', null ] },
{ $ne : [ '$$field.k', '_id' ] }
]
}
}
},
in: {
'$concatArrays': ['$$value',
{
'$filter': {
input: {
'$objectToArray': '$$this.changes'
},
as: 'field',
cond: {
$ne : [ '$$field.v', null ]
}
}
}]
}
}
}
}
}
},
{
'$match': {
state: { '$eq': 'DONE' }
}
],
{})
This is working great and the new document looks as expected:
{
state : 'DONE',
field2 : 'WILL NOT CHANGE',
field3 : 'SOMETHING',
lastChange : 1236547
}
My problem ist that the matching stage that follows the merge/reduce fails to 'query' the new document. Instead the old field values seem to be used.
state: { '$eq': 'DONE' } // should return the document, but doesn't
state: { '$eq': 'NEW' } // should NOT return the document but does
My best guess is that this has something to do with pipeline optimization? Meaning the optimizer moves the $match statement into the first stage, or something like that.
Does anybody had this problem as well and knows a solution?
Udpate 1
After more testing it seems that using $$ROOT as the initialValue of the reduce is part of the problem. If I use an empty array as the initialValue I can $match the reduced fields correctly.