[Typed evaluation P2] Optimize executable decision programs with independent final assessment

[drewstone]

Parent: tangle-network/agent-eval#768. Roadmap item 14; M8. Owner: Runtime improve/strategy evolution with Eval comparison. Depends on #1296/#1297/#1298 and agent-eval#770/#771.

• Use existing improve() and strategy evolution to vary questions, evidence renderers, thresholds, models and branch policies, not only prose prompts.
• Execute exact candidate programs and retained state under comparable resources; account for search, evaluations, retrieval and failed/incomplete attempts.
• Re-score saved distributions for deterministic mapping changes, but rerun changed execution paths before claiming downstream improvement.
• Keep final cases and final judge feedback outside adaptive search; retain candidate/version lineage and require explicit promotion of the exact assessed program.

Acceptance: an executable policy beats an appropriate frozen baseline on untouched tasks with uncertainty and full costs reported. Work on working-evaluator quality is assessed separately from optimizing a worker against it. No new optimizer facade or automatic self-modification authority; a higher development judge score alone is not success.