perf(row-reduce): complete Phase 4 evidence

[kim-em]

Summary

• add compiled mode-1 benchmarks for all ten advertised HexRowReduce executable operations
• add dense and rank-deficient rational families, exact-hash FLINT RREF/nullspace anchors, and CI/release-manifest integration
• record clean scientific evidence, comparator ratios, and samply profiles; promote HexRowReduce to Phase 4

Complexity contract

All ten direct registrations use the strongest ordered evidence mode. The public RREF, rank, span, and nullspace wrappers declare cubic fixed-aspect Gauss--Jordan cost; prepared echelon span queries declare quadratic matrix--vector work; prepared reduced-nullspace construction declares the documented cubic loose bound while profiling separates it from elimination. These models and input dimensions match the HexRowReduce SPEC, and each registration passes the mode-1 normalized-constant/slope checks.

Verification

• lake build (10,506 jobs)
• lake build hexrowreduce_bench
• .lake/build/bin/hexrowreduce_bench verify (39/39)
• python3 scripts/check_phase4.py
• release manifest, DAG, manual split, trust-surface, and Mathlib-free checks
• oracle unit tests and release-sync unit tests
• clean scientific, FLINT comparator, and two samply profile runs

Closes #9811

[kim-em]

Closing after premise validation found asymptotic implementation gaps that make the proposed Phase-4 claims unsound. The concrete source and measurement evidence is recorded on #9811. This PR must not merge by weakening the models or tolerances.