High-performance bounded survey-weight calibration: multiple solvers
(oris, oris_soft, raking, Sinkhorn, Greenkhorn, GREG, logit, Newton-KL)
sharing one C++17 core between R and Python, with min_weight/max_weight
bounds enforced on every returned weight.
On 100,000 observations across 5 margins, oris_soft calibrates in
0.0427s wall time at max margin error 3.35e-05, honouring the [0,3]
weight bound and retaining 67,489.4 effective observations — see
docs/performance.md for the full methodology, machine
provenance, and one-command reproduction.