final push

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

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

test(rubric): add golden parity tests for rubric migration validation

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

test(inference): add prompt profile and rank_candidates tests

feat(tests): update for new action system

fix: heuristic takeover when llm fails, step reward/penalty derived from grader, some other fixes

initial cook