RSI: implement adaptive multi-trajectory search

[w4ffl35]

Limitation\n\nThe RSI data model records compute policies such as independent trajectories, planner-executors, and critic-retry, but the controller currently runs one mutation/evaluation path per candidate.\n\n## Proposed direction\n\nImplement at least one adaptive policy that launches independent candidate trajectories, uses evaluation evidence to allocate follow-up compute, and records the decision path in the archive and reports.\n\n## Acceptance criteria\n\n- A configured policy produces more than one independently evaluated trajectory when budget allows.\n- Selection records why additional compute was allocated or stopped.\n- The behavior is bounded by explicit cost, time, iteration, and concurrency limits.