evidence and uptake regret are separate interfaces

Agent-decided, under AGENTS.md §6, and it corrects the first version of this round. They are the same functional <d,l> - <d M_r,l> on different distributions: the played p_t gives uptake regret, which Theorem A bounds; a supplied baseline b_t gives improvement evidence, which nothing bounds and nothing should. They are independent in both directions — a process that adopted a repair has zero uptake regret and may still have a large demonstrated improvement, and a process leaving advantage unused may never reach the threshold that makes a challenge eligible. The improvement challenge is therefore grounded by the demonstration and never by the failure to act on it, and the challenge module names no regret quantity at all.