fix: persist initial context (distractors/pressure) across all episode turns

feat: extend environment with level=3, resistance reward ±0.2

feat: extend reset() to support level=2 with distractor context

feat: add 429 retry wrapper to grader semantic check

Phase 2.5: multi-turn episodes, bug fixes, dataset cleanup

Phase 2 complete: Level 1 env runs locally, tests green, 100-question dataset

Phase 1 complete: schemas, reward design, project scaffold