v0.16.0 - Phase 2 (cloud, CSOT, load balancers)

Phase 2 of the roadmap is done.

CI tests covering standalone, replica set, sharded, and load-balanced topologies.

Test matrix (GitHub, 766 examples, 0 failures)

Topology	Time	Pending
standalone	23.86s	219
replica set	3:36	104
sharded	7:05	112
load-balanced	2:58	137

Added

• auth: SASLprep (RFC 4013) for SCRAM-SHA-256 passwords. Printable ASCII is unchanged (no extra allocation). Usernames are not prepared.
• uri: timeoutMS (CSOT deadline), srvMaxHosts, srvServiceName. mongodb+srv URI validation for loadBalanced, replicaSet, and directConnection. timeoutMS=0 means infinite. Negative timeoutMS is rejected.
• sdam: SRV polling fiber for mongodb+srv:// on Sharded or Unknown. Adds and removes mongos hosts. Does not run in load-balanced mode. Monitor hellos keep the last 10 RTT samples for CSOT min RTT.
• csot: Remaining timeoutMS minus min RTT is sent as maxTimeMS. Server code 50 (MaxTimeMSExpired) becomes Error::Timeout. timeoutMS on the client, database, collection, CRUD, indexes, GridFS, commit_transaction, and abort_transaction. timeoutMode (cursorLifetime / iteration) on find, aggregate, listCollections, and listIndexes. Tailable awaitData sends maxTimeMS on find and caps getMore with maxAwaitTimeMS. Change streams use iteration timeouts and resume in the same next call. GridFS uses one deadline for the whole upload/download (including listIndexes before the first write). Retryable reads/writes retry until timeoutMS (or forever if timeoutMS=0). Wait-queue timeout is Error::Timeout when timeoutMS is set. wTimeoutMS is omitted when timeoutMS is set. Official CSOT UTF is 28 files.
• run command: Database#run_command and #run_cursor_command. The caller's document is copied. Not retryable. Database read/write concern is not applied. timeoutMode and cursorType follow CSOT. batch_size, comment, and max_time_ms go on getMore only. Official run-command UTF is copied.
• load balancer: Do not pre-create minPoolSize sockets. Require serviceId on hello. Pin the TCP socket for a transaction and for an open cursor. Unpin returns the socket to the pool. Pool generation is per serviceId. Wait-queue timeout lists cursor / transaction / other in-use counts. poolClearedEvent includes serviceId. Command events include serviceId. CMAP connectionReadyEvent and connectionClosedEvent with a reason. UTF assertNumberConnectionsCheckedOut. Official load-balancer UTF runs (except the JSON skipReason file and one per-serviceId test that needs two mongos from HAProxy).
• testing: UTF ops iterateUntilDocumentOrError, iterateOnce, createFindCursor, runCursorCommand, createCommandCursor, GridFS upload / delete / rename, dropIndex / dropIndexes, assertNumberConnectionsCheckedOut. Official GridFS, collection-management, index-management, run-command, and CSOT JSON. CLAM redacted-commands.json and CRUD create-null-ids.json now run. GitHub uploads tmp/utf-timing.log. GitHub load-balanced runs full crystal spec (HAProxy via spec/support/run-load-balancer.sh).
• change streams: watch sends comment on aggregate and getMore (string or document). showExpandedEvents and fullDocumentBeforeChange go on $changeStream. A labeled getMore error resumes with a new aggregate (getMore is not a retryable read). maxAwaitTimeMS must be less than timeoutMS when both are set.

Changed

• Cursor finalize no longer sends killCursors or touches the pool (GC thread). Call #close, #each, or a block find.
• Load-balanced topology does not start monitor sockets. Sessions are always supported in that mode. LoadBalancer always allows retryable reads and writes (hello fields stay unset without monitors).
• getMore is not retried as a retryable read. Overload retry still runs.
• timeoutMode iteration starts a fresh timeoutMS on each next / try_next (change streams still own the deadline for resume).
• HAProxy multi-mongos uses roundrobin and health-checks the normal mongos ports (not the PROXY v2 ports).

Fixed

• Upsert reply with _id: null deserializes. Duplicate-key write errors expose code.
• GridFS rename errors when the file id is missing. UTF downloadByName honors revision.
• UTF matcher: $date canonical vs relaxed, $$type int/long with $numberInt.
• Find omits tailable and awaitData unless they are true. Sending tailable: false broke Versioned API strict (crud-api-version-1-strict.json).
• A runCommand reply with a cursor id pins the load-balanced socket (the raw BSON is scanned for cursor.id).
• UTF createCollection stores the collection entity and sends capped / size / max.
• runCommand does not send $readPreference to a standalone, even when the caller asked for a non-primary mode.
• getMore and killCursors do not get readConcern / afterClusterTime (explicit causal sessions).

Test info

Sharded takes more time than replica set because change-stream iterate on mongos often waits about 2 seconds. Load-balanced 2:58 is a full crystal spec (not the old official-LB-UTF-only job). It has more pending files, so it is faster than replica set and sharded.