[stamich]

Summary

Reduce fused-vector allocation by generating whole-run float32, float64 and mixed-precision loops. Prefer eager vector ufuncs for direct single-operation expressions, reduce small-expression eager setup, and add allocation and contraction benchmarks.

Add an experimental, explicitly closed NJ.parallel(workers) context for selected contiguous float32/64 unary and binary ufuncs, expressions and deterministic full-array sum. Confined memory, unsafe overlap, unsupported layouts and small inputs use existing caller-thread paths. Fixed sum chunks and merge order are independent of worker count. The parallel execution guide documents lifecycle, scope closure, nesting and numerical limits.

Validation

• JDK 25 local build, fused-provider tests and the fused-expression example passed before the final nested-context and interruption fixes; targeted parallel context, stress and expression tests passed afterward.
• Staged local artifacts passed separate classpath and module-path consumer checks, both with and without the vector provider.
• NumPy 2.5.4 verified all 4,245 generated oracle files; the generated parallel ufunc/sum fixtures and reference runner unit tests passed locally.
• A short JMH smoke run exercised parallel ADD, CHAIN and SUM. A separate single-fork shared-machine investigation found that generated loops substantially reduce large float32 fused allocation; these runs do not establish reference-machine performance.
• Current-head GitHub build, NumPy oracle, scalar, required-SIMD, native backend and SonarCloud checks passed at ac454b5.
• The branch contains 27 task commits with the requested author and author/committer dates.

Open evidence and readiness

E04 prepares the reference campaign; the user's Ryzen 9 5950X run has not been performed on this revision. Worker scaling, crossover thresholds, dot/vecdot reference measurements and the earlier 0.5.0 fusion/allocating/AUTO performance gates remain open. No target has been silently waived. See docs/benchmarks/M4-reference-protocol.md, docs/benchmarks/M4-performance-status.md and docs/milestones/M4-0.6-release-checklist.md.

Keep this PR draft until the outstanding reference evidence is reviewed. No tag or publication is implied.