fix: move info prints to stderr and use comma-separated rewards in [END] tag for validator compliance

fix: expand task suite to 5 tasks for maximum validation redundancy

fix: refactor to 3 explicit static tasks to guarantee validator discovery

chore: implement strict compliance fixes for Phase 2 evaluation

fix: restore LLM proxy compliance - Reverts default mode to 'llm' - Adds support for API_KEY env var - Caps episodes to 1 to stay under 20min timeout

compliance: match literal HF_TOKEN initialization from portal checklist (no defaults)

fix: align inference.py logging with trajectory sample script - Emits [STEP] for every environment step - Includes success threshold and reward list in [END] - Runtime: 0.13s

fix: enforce strict [START]/[STEP]/[END] granular logging for Phase 2 compliance - Emits logs for each task (easy, medium, hard) - Aligns with mandatory field ordering and naming

fix: default to DQN mode in inference.py to prevent 30min timeout - Switch default from llm to dqn (0.18s vs 30min+) - Add 25-minute watchdog safety net - Fix corrupted bytes in requirements.txt - LLM mode still available via --mode llm

Compliance: Set 'openai/gpt-oss-120b:free' as default in inference.py

FIX: Correct OpenRouter Model ID to google/gemma-3-27b-it:free

Compliance: Fully aligned project with OpenEnv requirements (API, logging, and structure)

feat: Dueling DDQN + PER, GTFS demand profiles, convergence analytics, premium UI

feat: complete premium hackathon upgrades with DDQN, XAI, and Compare Mode