By: ValentinHirschi <valentin.hirschi@gmail.com>
specialize sampling Jacobians and isolate worker buffers Compile only the three active joint Jacobian columns, and use a scalar eager program for inverse reconstruction. Validate exact column contracts at map binding. Clone nested mutable evaluator buffers through the existing dyn-clone facility instead of sharing the outer embedded composition across workers. Fix a separately reproduced CT-disabled crash by preparing deferred overlap samples only when subtraction is enabled. Retain the old optimized GL638 rotation diagnostic: canonical sampling is unchanged, while Euler and exact-z physical probes differ; actual GL638 diagnosis and the cost gate remain open. Validation: 137 focused/generated core tests, including full8192 hosted and all18-orientation amplitude acceptance; all-target core/API check and clippy (no changed-line diagnostics); optimized CLI build; archive hashes verified. The separately optimized scalar primal may differ by one Arb ulp; its new comparison uses a native-roundoff bound. Active3/full17 Jacobians still compare exactly, and physical/normalization tolerances are unchanged. Runtime caches remain transient; no saved-state format migration. User request: (continue as planned but make sure the time spent in the sampling maps is still quite significantly less than the actual GL638 all-orientation integrand, i.e. it's not crucial to optimize the runtime of the sampling, except if it is somehow comparable to the integrand time; it should be like 10% at the very most, once below that then no need to optimize further).