perf(mosaic): skip proven regional depth reads

[Qubitium]

Summary

• add a per-owned-region normal-Z depth minimum for eligible Mosaic opaque prepared draws
• skip redundant per-sample depth rereads on proven passing spans
• reduce failed atomic queue claims after a Mosaic queue is exhausted
• preserve strict fallback and keep the change private to the Mosaic Vulkan ABI path
• document realistic Vulkan ABI workload evidence

This targets complex terminal, desktop, game, and editor-shaped streams rather than vkcube. vkcube is intentionally not optimized or used as evidence.

Why this is safe

The regional minimum is only used for aligned, little-endian, uncounted depth buffers and admitted fast Mosaic draws. Unsupported or strict/fallback work invalidates the summary. The second triangle in an admitted quad observes the first triangle's conservative update before using the fast path.

Evidence

Controlled four-physical-core runs using ZPU_MAX_THREADS=4 tools/limited-cpus.sh:

• vkQuake-shaped: approximately 6.03 ms -> 4.02 ms median across three smoke invocations
• vkQuake2-shaped: approximately 7.10 ms -> 4.24 ms median
• Serious Sam-shaped: effectively within host noise
• targeted candidate p50: vkQuake 4.54x vs per-draw, 2.79x vs legacy; vkQuake2 3.82x vs per-draw, 2.34x vs legacy; Serious Sam 4.29x vs per-draw, 2.10x vs legacy
• all 13 realistic workload checksums remain identical

These are host-specific measurements, not a universal 4x guarantee.

Validation

• zig fmt --check
• git diff --check
• ZPU_MAX_THREADS=4 tools/limited-cpus.sh zig build
• ZPU_MAX_THREADS=4 tools/limited-cpus.sh zig build test
• zig test src/vulkan/cpu_cube.zig -lc (25/25)
• zig test src/benchmark_vulkan_abi.zig -lc (27/27)
• zig test src/vulkan/driver.zig -lc (296/296)