perf: apply TDS performance optimisations from C# port

- TDSLogin7: negotiate 32KB packet size (was 4096) — reduces packet count ~8x
- sendPacket: coalesce all TDS fragments into single ByteBuffer, one writeAndFlush
  instead of N channel.write() calls per large request
- TDSDecoder.decodeNbcRow: replace readBytes() heap copy with zero-copy readSlice()
  and getInteger(at:) for null bitmap access

Mirrors the same three fixes applied to CosmoSQLClient-Dotnet (commit 04ad1f5)
which yielded 1–21% throughput improvement over Microsoft.Data.SqlClient.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: vkuttyp

Sources/CosmoMSSQL/MSSQLConnection.swift

@@ -751,16 +751,22 @@ public final class MSSQLConnection: SQLDatabase, @unchecked Sendable {
     }
 
     private func sendPacket(type: TDSPacketType, payload: inout ByteBuffer) {
-        // Encode TDS packet(s) and send each packet individually.
-        // SQL Server can reject multi-packet TDS messages sent in a single write.
+        // Coalesce all TDS packets into a single ByteBuffer and flush in one syscall.
         let payloadSize = payload.readableBytes
-        let maxBody = 4096 - TDSPacketHeader.size  // 4088
+        let maxBody = 32768 - TDSPacketHeader.size  // 32760
+
+        // Pre-compute total wire size so we allocate exactly once.
+        let fullPackets = payloadSize / maxBody
+        let remainder  = payloadSize % maxBody
+        let numPackets = fullPackets + (remainder > 0 ? 1 : 0)
+        let wireSize   = payloadSize + numPackets * TDSPacketHeader.size
+        var wire = channel.allocator.buffer(capacity: wireSize)
 
         var offset = 0
         var packetID: UInt8 = 1
         while offset < payloadSize {
             let chunkLen = min(maxBody, payloadSize - offset)
-            let isLast = (offset + chunkLen) >= payloadSize
+            let isLast   = (offset + chunkLen) >= payloadSize
             let totalLen = UInt16(chunkLen + TDSPacketHeader.size)
 
             let header = TDSPacketHeader(
@@ -769,21 +775,14 @@ public final class MSSQLConnection: SQLDatabase, @unchecked Sendable {
                 length: totalLen,
                 packetID: packetID
             )
-            var pkt = channel.allocator.buffer(capacity: Int(totalLen))
-            header.encode(into: &pkt)
-            pkt.writeBytes(payload.getBytes(at: payload.readerIndex + offset, length: chunkLen)!)
-
-            if isLast {
-                // Last chunk: writeAndFlush flushes all buffered writes in one TCP segment.
-                // Fire-and-forget — the server can't reply until it receives this, so reads
-                // will naturally follow the write through the event loop's FIFO ordering.
-                channel.writeAndFlush(pkt, promise: nil)
-            } else {
-                channel.write(pkt, promise: nil)
-            }
+            header.encode(into: &wire)
+            wire.writeImmutableBuffer(payload.getSlice(at: payload.readerIndex + offset, length: chunkLen)!)
+
             packetID = packetID == 255 ? 1 : packetID + 1
             offset += chunkLen
         }
+        // Single writeAndFlush → one TCP segment for all packets.
+        channel.writeAndFlush(wire, promise: nil)
     }
 
     /// Receive one complete TDS message via the async stream bridge handler.

Sources/CosmoMSSQL/TDS/TDSDecoder.swift

@@ -158,15 +158,16 @@ struct TDSTokenDecoder {
 
     private mutating func decodeNbcRow(_ buf: inout ByteBuffer) throws {
         let bitmapLen = (columns.count + 7) / 8
-        guard let bitmapBytes = buf.readBytes(length: bitmapLen) else {
+        // Use a zero-copy slice instead of readBytes() to avoid heap-allocating a [UInt8].
+        guard let bitmapSlice = buf.readSlice(length: bitmapLen) else {
             throw TDSError.incomplete
         }
 
         var values: [SQLValue] = []
         for (i, col) in columns.enumerated() {
-            let byteIdx = i / 8
+            let byteIdx = bitmapSlice.readerIndex + i / 8
             let bitIdx  = i % 8
-            let isNull  = (bitmapBytes[byteIdx] >> bitIdx) & 0x01 == 1
+            let isNull  = ((bitmapSlice.getInteger(at: byteIdx, as: UInt8.self) ?? 0) >> bitIdx) & 0x01 == 1
             if isNull {
                 values.append(.null)
             } else {

Sources/CosmoMSSQL/TDS/TDSLogin7.swift

@@ -9,7 +9,7 @@ import Foundation
 struct TDSLogin7 {
     // Fixed header fields
     var tdsVersion:        UInt32 = 0x74000004   // TDS 7.4
-    var packetSize:        UInt32 = 4096
+    var packetSize:        UInt32 = 32768
     var clientProgVer:     UInt32 = 0x07000000
     var clientPID:         UInt32 = UInt32(ProcessInfo.processInfo.processIdentifier)
     var connectionID:      UInt32 = 0