Scope the primary-key -> _id rename to the outermost document

The MongoDB convention of renaming the schema's primary-key field to
`_id` was being applied recursively. Once `Conversions.from_ecto_pk/2`
descended into a nested map (for example an `embeds_one` sub-document),
the parent schema's `pk` value was still threaded through, so any
nested field named `:id` was also rewritten to `:_id`.

That meant an embedded value like

    %{jurisdiction: %{id: "MD", title: "Maryland"}}

was stored as

    {jurisdiction: {_id: "MD", title: "Maryland"}}

even when the embedded schema declared its own `@primary_key {:id, :string,
autogenerate: false}`. Queries that filter on `jurisdiction.id` (the
common shape for apps that put a string identifier on a sub-document)
return nothing, and the wire format diverges from what other clients
(Ruby driver, Node driver, etc.) write for the same model.

The pk -> _id rename is only meaningful at the top of a document. Below
that, the field names are user-controlled and should be passed through
unchanged. This patch makes the recursive call inside `document/2` and
`document/3` pass `nil` for the pk argument, so the rename only fires
for the outermost document.

`Conversions.to_ecto_pk/2` (the read path) is left as-is for now. Its
rename of nested `"_id"` -> `Atom.to_string(pk)` only matters when a
sub-document actually contains an `_id` field, which is uncommon in
hand-written documents (Mongo only auto-adds `_id` to top-level docs).
A symmetric read-side change can follow separately if there's appetite.

Author: scottmessinger

lib/mongo_ecto/conversions.ex

@@ -74,13 +74,25 @@ defmodule Mongo.Ecto.Conversions do
 
   defp document(doc, pk) do
     map(doc, fn {key, value} ->
-      pair(key, value, pk, &from_ecto_pk(&1, pk))
+      # Recurse with `nil` so the parent schema's primary key isn't
+      # accidentally rewritten when it appears inside a nested document
+      # (e.g. an `embeds_one` whose embedded schema has its own `:id`
+      # field). Without this guard, an embed like
+      #
+      #   %{jurisdiction: %{id: "MD", title: "Maryland"}}
+      #
+      # gets serialized as `jurisdiction._id`, which breaks queries that
+      # filter on `jurisdiction.id`. The pk → _id remap is only meaningful
+      # at the top of the document.
+      pair(key, value, pk, &from_ecto_pk(&1, nil))
     end)
   end
 
   defp document(doc, params, pk) do
     map(doc, fn {key, value} ->
-      pair(key, value, pk, &inject_params(&1, params, pk))
+      # Same rationale as `document/2`: scope the pk → _id rename to the
+      # outer document. See the comment in `document/2`.
+      pair(key, value, pk, &inject_params(&1, params, nil))
     end)
   end