elemMatchProjection.js incorrectly assumes that shell generates monotonically increasing _id values

The elemMatchProjection.js integration test has various assertions which sort by _id, and depend on this sort enforcing a particular ordering. The test, however, does not supply explicit _id values. It instead relies on the shell generating ObjectIds, and it assumes that these ObjectIds are monotically increasing.

This is not a sound assumption. As documented here, the least significant three bytes of an ObjectId consist of an incrementing counter which is seeded with a random value. If the uint32_t seed is close to 0xffffffff, then during the execution of the test, the counter will roll over to zero. If this occurs, the sort order of the _id values will not match the shell's insertion order, causing the test to fail.

The test fails consistently after building the server with the following patch:

diff --git a/src/mongo/bson/oid.cpp b/src/mongo/bson/oid.cpp

index 26d90d97ba..00f0159f1d 100644

--- a/src/mongo/bson/oid.cpp

+++ b/src/mongo/bson/oid.cpp

@@ -54,7 +54,7 @@ OID::InstanceUnique _instanceUnique;

 MONGO_INITIALIZER_GENERAL(OIDGeneration, MONGO_NO_PREREQUISITES, ("default"))

 (InitializerContext* context) {

     std::unique_ptr<SecureRandom> entropy(SecureRandom::create());

-    counter.reset(new AtomicUInt32(uint32_t(entropy->nextInt64())));

+    counter.reset(new AtomicUInt32(uint32_t(0xffffff35)));

     _instanceUnique = OID::InstanceUnique::generate(*entropy);

     return Status::OK();

 }

This replaces random seeding of the ObjectId counter with a hardcoded value of 0xffffff35, designed to demonstrate the problem.