[CanReader]

Pins the engine to engine-optimization (CanReader/SleakEngine#227), adds a repeatable benchmark, and fixes stutter while moving through the world.

Changes

• scripts/bench_run.sh: deterministic runs (capture mode: fixed 1/60 step, synchronous chunk loading, fixed seed) that keep each run's CSV, log and frame capture. Defaults to the real display; BENCH_HEADLESS=1 uses Xvfb, where present is a CPU copy so only the CPU/GPU columns mean anything.
• scripts/bench_compare.py: per-scenario mean ± stdev of FPS, frame time, CPU frame, GPU passes and culling, plus capture md5s for pixel-identity checks. Truncated runs are skipped.
• -world, --autofly and -campose now work in Release builds when --bench is passed (still Debug-only otherwise).
• --bench-postfx turns on SSAO, SSR and bloom so those engine paths can be measured.
• Column mesh merging reuses scratch buffers and copies each chunk in bulk instead of growing fresh vectors per rebuild. Most of the old cost was page faults on new memory. Output is byte-identical.
• Main-thread column rebuilds stop once they pass 2 ms in a frame (ChunkManager::UPLOAD_BUDGET_MS); the rest stay dirty for the next frame.

Results

Standing still: see CanReader/SleakEngine#227. Vulkan at 1600x900, render distance 12: about 212 → 220 FPS, CPU frame 2.05 → 1.62 ms, output byte-identical to staging.

Flying (--autofly 20, fullscreen 1920x1080, render distance 16, Release, before/after runs interleaved):

	Before	After
Average FPS	117	137
p99 frame time	23.0 ms	15.3 ms
Frames over 16.7 ms	4.5/s	0.9/s
CPU frame p99	15.5 ms	8.0 ms