Emit per-task inference logs and restore 3-task spec

Rebuild task setup with 4 enabled graded tasks and multi-task inference

Add validator compatibility shims for inference and app entrypoint

Refactor inference.py to run single task per execution (validator calls once per task)

Align inference logging with strict START/STEP/END validator format

Clamp logged task scores to strict (0,1) range

Add robust fallback findings to keep task scores in valid range

Route inference calls through injected API_BASE_URL/API_KEY

Make inference.py fail-safe to avoid phase-2 crashes

Update grader and inference

Initial commit: MetaXScalar project