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

Copied material effects

The newer mixed callback sequence predictor composes these material effects with root state callbacks, constant setup/maps and declaration/cache changes in one persistent copied state. Its independent x86 comparisons and typed input bindings retain the same admission boundary.

native-material-leaf-plan.hpp predicts all nine parameter leaves from copied values and an explicit environment. Calls contain owned constant values, texture backing identities, sampler changes and render-state changes. The prediction also owns the resulting native shadow bytes. It never dereferences a native address, invokes a driver or creates a physical owner.

KS passes 486,021 Windows x86 checks, including 13,824 persistent native actions across all nine leaves and three distinct device-selection branches. The actual ready-texture query, texture/sampler transforms and render-state helpers run; only driver destinations are controlled. Local ASan/UBSan passes 99,238 checks. The supported texture-bias arithmetic requires nearest rounding and normal or zero finite inputs/results. Subnormal, non-finite, overflow and underflow cases refuse. The prediction preserves its caller's floating-point exception flags. Native floating-point mode still requires independent qualification.

native-material-effect-plan.hpp composes the native ready-material order: cache flush, vertex/pixel binding, dynamic program and static program. Several materials share the preceding predicted state, with no native-state refresh. The cache arrays and full shadow must agree at entry and after each material. State row eight's unused register 103 overlaps the native mip-bias word; the fixture preserves that real alias instead of treating them as separate state. Bounds cover the whole sequence: 1,024 materials, 65,536 parameter records, 65,536 calls and 262,144 retained input/output words. Any refusal clears the complete prediction, including when earlier materials were valid.

KT2 passes 1,278,087 Windows checks, including 576 consecutive complete native callbacks. The original cache helper, shader resolution/binds, VM and nine leaves run against persistent state. Every call, shadow byte and shader-cache word matches the copied prediction after each callback. The driver remains controlled. KT's preceding failed freeze is preserved: the fixture originally wrote a subnormal integer bit pattern into the aliased mip-bias word and the prediction correctly refused it. The supported-value fixture correction does not expand floating-point coverage.

packet-material-effect-input.hpp connects retained snapshots to the pure prediction types. It checks packet/context identity, exact copied root/program records and payload spans, shader identities, texture backings and physical owners. KU2 passes 13,515 Windows and local sanitizer checks, including absent programs whose authenticated pointer words remain nonzero. Native code does not follow those unused words; nonempty programs still require exact spans. KV2 passes 10,795 checks of the actual batch observer's input preparation, retirement and cancellation. KV's failed Windows-macro compilation remains preserved.

The observer retains these inputs with their original batch until retirement. Its receipt describes input preparation only. Census address order is not native execution order. Complete packet command/effect composition, source/free/context closure, native effect admission and drawing authority remain open. All components above have callback_authority=false.