[Shiqan]

Performance & Memory Allocation Improvements

This PR optimizes the IO, memory slicing, and bitstream parsing hot paths in the replay reader. By replacing reflection and allocating operations with Span<byte>, ReadOnlyMemory<byte>, BinaryPrimitives, zero-copy buffer slicing, and reusable object pools, end-to-end replay parsing is significantly faster and allocates ~34% to 38% less managed heap memory.

---

Benchmark & Real-World Results

1. Recent Replay Comparison 16.03 MB)

Metric	master (Baseline)	perf-improvements	Improvement
Heap Allocations	212.85 MB	140.32 MB	-72.53 MB (-34.1%)
Cold Parse Time	622 ms	462 ms	25.7% faster
Warm Parse Time	440 ms	404 ms	8.2% faster
Average Parse Time	555 ms	459 ms	17.3% faster
GC Gen 0 Pauses	10–11 collections	5 collections	-50% GC pressure

---

Summary of Changes

1. Unreal.Core (IO & Bitstream)

• Direct Span Primitives: Replaced BitConverter and sliced allocations with BinaryPrimitives.Read*LittleEndian directly over Bytes.Span.
• Zero-Alloc Enum Deserialization: Replaced Enum.ToObject(typeof(T), byte) with Unsafe.As<byte, T> and Unsafe.As<uint, T>, eliminating boxing and reflection.
• BitReader 64-Bit Word Shifting: Replaced single-bit loops in ReadBitsToInt and ReadBitsToLong with unaligned 64-bit word reads (BinaryPrimitives.ReadUInt64LittleEndian) and bitwise shifts (7.3x faster).
• Zero-Alloc Property Reading: Added FillBitsFrom to deserialize property bunches directly into internal reusable buffers instead of allocating a new byte[] for every property.
• Flat Array TempEnd: Replaced Dictionary<FBitArchiveEndIndex, int> with a fixed int[4] array for bunch temp ends (4.6x faster).
• String Prefix Stripping: Replaced LINQ .Where(...).Any() in RemoveAllPathPrefixes with an allocation-free ReadOnlySpan<string> loop (-95.4% allocations).
• Zero-Copy Memory Slicing: Added ReadMemory(byteCount) in FArchive and BinaryReader to slice underlying buffers without copying.

2. Unreal.Core (Replication & Guidance)

• Zero-Copy Raw Packets: Introduced ReceivedRawPacket(ReadOnlyMemory<byte>), eliminating hundreds of thousands of packet array allocations per replay.
• Bunch Reuse: Reused _currentBunch for non-partial bunches, saving ~350,000 bunch object allocations per parse.
• Lazy Vector Serialization: ConditionallySerializeQuantizedVector now avoids instantiating default new FVector() instances when no delta was serialized (-50% vector allocations).
• Direct NetGuid Cache Resolution: Added direct 1-step index lookup (_netFieldExportGroupFromIndex) to bypass 2-step dictionary lookups (2.1x faster).

3. FortniteReplayReader

• Modernized Chunk Decryption: Used File.ReadAllBytes with zero-copy chunk slices and in-place Aes.DecryptEcb over spans.

4. BenchmarkReader & Testing

• Modernized Benchmark Harness: Simplified Program.cs to top-level statements delegating to BenchmarkDotNet.Running.BenchmarkSwitcher. Allows interactively selecting any benchmark or passing CLI filters (e.g., -- 4 or --filter *BenchmarkEnumParsing*).
• Benchmark Suite: Added isolated BenchmarkDotNet classes for each distinct optimization.
• Regression Safety Net: Added xUnit performance and allocation tests in Unreal.Core.Test asserting functional parity and zero heap allocations (GC.GetAllocatedBytesForCurrentThread() == 0).