Add hybrid banked PID+FF topology controller

[teerthsharma]

Summary

This PR adds hybrid_banked_pid, a PID-style feedback controller with feedforward preview and deterministic topology-signature coefficient banks.

The controller remains runtime-light: no MPC, no online optimization, and no model inference inside the controller. Offline tuning is compressed into a coefficient JSON.

Method

The controller combines:

• 2DOF PID feedback
• target/feedforward preview terms
• bounded action smoothing
• p-adic-style coefficient bank lookup
• topology-signature bank selection from warmup trajectory features

Topology is used as a deterministic regime fingerprint, not as an unbounded control law. If no exact/topology bank applies, the controller falls back to the global PID+FF coefficients.

Validation

Verified locally on the official-style 5000 segment evaluation:

• hybrid_banked_pid: mean total_cost = 82.6889
• lataccel_cost mean: 1.1928
• jerk_cost mean: 23.0483

Eval command:

python -u eval_s3_poly_pid.py \
  --model_path ./models/tinyphysics.onnx \
  --data_path ./data \
  --num_segs 5000 \
  --controllers hybrid_banked_pid \
  --workers 4 \
  --output ./runs/real_tail_bank_blend14_top300_faithful/best_filtered_first1000_eval_5000.csv

Notes

This is intended as a compact PID-style submission under the challenge threshold, not as a claim of leaderboard-best performance.