sabaridsnfuji/repro-near-optimal-and-efficient-first-order-algorithm-for-multi-task-learning-with-shared-r

Fix Claims 4/5/6: multi-seed averaging (C4), kappa-sweep for untested dependence (C5), and a real RNG-comparability bug where the new task silently changed across T (C6)

Publish reproduction logbook: Near-Optimal First-Order Algorithm for Multi-Task Learning

initial commit