Profiling showed TerminalTexture.upload() was the bottleneck: - Before: per-pixel setPixelRGBA loop = 2.5ms for 984K pixels (15% frame budget) - After: MemoryUtil.memCopy to NativeImage backing memory = 0.95ms (5.7%) The remaining 0.95ms is the GL texture upload (texture.upload()) which is an irreducible GPU operation. Implementation: - NativeImageAccessor mixin exposes the private `pixels` field (native ptr) - TerminalTexture caches the pointer at construction - upload() does a single memcpy from the direct ByteBuffer to the NativeImage - Falls back to per-pixel loop for non-direct buffers Also: ScreenGroup caches min/max bounds in constructor (eliminates 6 stream pipelines per getSubRegion call per frame). Profiling data (4x3 grid, 1440x720 terminal): - pollPty: 0.01ms (negligible) - getPixelData (Rust render): 0.38ms - texture upload: 0.95ms (was 2.5ms) - Total per-frame terminal cost: ~1.3ms (was ~2.9ms) 46 GameTests + 52 Rust tests, all passing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>