Narrow pointwise segmentation to temporary identities

[PraneethMerugu]

Summary

• keep graph-shaped specialization in the small owner that extracts temporary-field identities
• pass only the runtime temporary-identity set into the reusable pointwise segmentation law
• preserve the existing KernelAbstractions execution path, ordering, forwarding, and materialization semantics

Compiler evidence

Fresh Kaimon/KaimonCompilerTools measurements show that the prior 551-statement segmentation body specialized on complete validated-binding tuple shapes. The new graph-aware wrapper is 41 statements, while the 512-statement semantic body is reused across 4/8/16-stage graphs for each legitimate empty or UUID identity-set family. Complete planning and warm execution allocations are byte-identical to baseline.

Validation

• LocalMath focused CPU: 135/135
• LocalMath behavioral CPU: 1,942 passed; the sole package-suite error is the baseline-identical Aqua traversal failure for KernelAbstractions' optional EnzymeCore weak dependency
• LocalMath real Metal: 592/592 on KA 0.9.42, GPUCompiler 1.23.0, Metal 1.10.0 with scalar indexing disabled
• CorePotts CPU canaries: 17,658/17,658
• Potts CPU canaries: 99/99
• CorePotts Metal checkerboard: 83/83
• CPU/Metal continuation: 116/116

No cache, lookup table, second graph representation, alternate executor, or model-specific kernel is introduced.