Dated source reference. Current release status: handoff state. Full original relative paths are in the developer archive.

Ordered material-cache predictions

native-material-cache-sequence.hpp composes the existing write and flush models from one initial cache snapshot. Each subsequent action consumes the previous prediction; it never refreshes a changed cache from native memory. The result owns its register ranges, values and per-action vertex-shader invalidation markers. It contains no native calls or executable authority.

The supported actions are the three initialized cached writers (rows 0, 4 and 8) and the complete mask-helper flush. An absent mask differs from an empty mask stream. Invalid rows, malformed streams, unknown actions, inconsistent payloads and exhausted bounds refuse the entire sequence and clear all output, including when earlier actions were valid. Bounds are 4,096 actions, 65,536 emission records and 262,144 words across retained inputs and emissions.

JZ passes 178,872 Windows x86 checks, including the existing 512-transition native flush baseline and 768 interleaved write/flush actions. The latter run original cached-writer instructions and the complete native flush helper against one persistent native cache image per sequence, comparing ordered emissions, all active cache arrays, shader invalidation and LastError after every action. Driver constant calls and flush dispatch destinations are controlled; the render-state dispatch table is deliberately zeroed to isolate the cache effects. This does not qualify real render-state helpers or drivers.

The same portable sequence and failure tests pass 11,867 checks under local ASan/UBSan. Whole-packet integration, remaining callbacks, resource/source and context contracts, native ordered-effect acceptance and drawing authority are still required. JZ is a standalone component receipt; its pass is not inherited by the JY combined candidate frozen before this addition.