final push

feat(grpo): enhance action response parsing by removing reasoning blocks and refining regex handling

feat(grpo): update max completion length and refine prompt handling for improved evaluation

feat(grpo): adjust training parameters and disable thinking mode for consistent action calls

feat(grpo): enhance training dynamics with new replay policies and update state steps

feat(grpo): update max sequence length and refine prompt formatting in training scripts

refactor: remove SakhaEnvWrapper class and streamline reward function in GRPO training script

refactor: simplify GRPO training script by removing CLI presets and syncing with notebook implementation

feat(eval): consolidate eval_harness into eval_policies, add LLM policy support

refactor(eval): extract shared eval constants and policies into eval_common.py

feat(plots): add plot generation script and training evidence plots

feat(train): add CLI args (learning-rate, batch-size) and full-shift defaults

feat(eval): add reproducible eval harness with comprehensive metrics

data(fixtures): capture pre-migration golden reward fixtures for parity testing

chore: pre-merge cleanup for colab-training branch

feat: add GRPO training script with mode presets, eval split, checkpointing

scripts: add GRPO training demo for Sakha env

fix(eval_policies): ensure output directory exists before writing JSON

feat(inference): adapt to new action system

initial cook