[arclabs561]

Replace locally authored unsafe PQ/NEON kernels and benchmark allocator wrappers, and enforce forbid(unsafe_code) across targets. Safe stats_alloc instrumentation retains process-wide diagnostics; acquired bytes count allocations plus positive realloc growth.

This development branch pins Pulp to 601de0c071ba32df3fd7ff33fc4218c9bd712aba and Innr to 392736772688262ce5f89f51db6b08a8167b8262. A root Innr override makes Qntz and benchmark dependencies exercise that same safe port. Rust 1.94 is required by the pinned x86 backend; CI and docs match. The LID example uses the standard golden-ratio constant with identical f32 bits.

Local QA, all-feature check, strict all-target Clippy, docs, benchmark compilation and matched Rust 1.94 ARM/x86 library checks pass. The full library suite passed 614 tests with 5 existing measurement skips on the preceding snapshot; the final executable delta is a bit-identical example constant. Final exact-head CI 39844248983 passes all 13 jobs, including Python, semver, feature coverage, regressions, ARM, x86, MSRV and ANN quality/smoke checks.

Keep this draft: Git dependencies still gate registry publication. This root patch validates development composition; dependency internals retain unsafe and unrelated paste remains through Simba.