[M5][CONTROL] Implement adaptive-therapy and optimal-control research policies

[andreazedda]

Objective

Create a research layer for dynamic treatment policies that update actions as disease, toxicity, organ state, uncertainty, and resistance evolve over time.

Scope

• formalize state, action, observation, and constraint spaces;
• support open-loop schedules, feedback policies, model-predictive control, threshold policies, and other transparent baselines;
• represent dose, hold, restart, treatment sequence, modality switch, and supportive intervention as bounded research actions;
• include efficacy, residual disease, toxicity, immune recovery, function, feasibility, and uncertainty in the objective structure;
• enforce hard safety and feasibility constraints separately from soft utility penalties;
• compare adaptive policies with fixed schedules under identical model assumptions;
• evaluate robustness to state-estimation error, parameter uncertainty, delay, and model mismatch;
• preserve a strict distinction between research policy and clinical instruction.

Acceptance criteria

• Every policy has a complete state/action/constraint/utility specification.
• No action is generated outside declared bounds.
• Hard constraints cannot be traded away by utility weights.
• Adaptive and fixed-policy comparisons use identical initial conditions and model versions.
• Policy performance is stress-tested under uncertainty and structural alternatives.
• Instability, infeasibility, and policy-chattering are detected.
• Results include transparent baseline policies.
• Runtime language labels outputs as simulated research policies, not medical recommendations.

Non-goals

This issue does not deploy automated treatment control or establish clinical safety.