FIX: Parse request body directly in /step override

FIX: Override /reset and /step to use global env when no session_id

FIX: Add global environment singleton fallback for session-less requests

TRIGGER: Force HF Space rebuild to deploy session state fixes

FIX: Remove custom /reset override to enable session state management

FIX: Add middleware to wrap direct /step payload format for validator compatibility

SIMPLIFY: Override /reset to create fresh environment directly, avoiding original endpoint call

FIX: Properly type request parameter in /reset endpoint as fastapi.Request

CRITICAL FIX: Replace /reset middleware with direct endpoint override - ensure base fields included

ENHANCE: Update reset middleware to include base Observation fields (done, reward, metadata) at top level

FIX: Add middleware to correct /reset response schema - add 'info' field, remove reward/done

CRITICAL FIX: Graders now return FLOAT ONLY (not tuples) for validator contract compliance - extracted float from raw grader tuples

Add comprehensive runtime grader validation endpoint (/validate-graders) to audit actual grader outputs under validator-style conditions

Enhanced /diagnostics: validate manifest import paths with importlib (mirrors validator behavior)

Add startup and diagnostics logging to debug task loading

Simplify metadata task schema to minimal fields (id, has_grader) for validator compatibility

PROPER FIX: Remove default /metadata route before registering custom override with tasks

CRITICAL FIX: Add /metadata endpoint with task enumeration for validator discovery

Add debug endpoint to verify deployed grader fix

Phase 2: Benchmark Quality Upgrades - Progressive graders, realistic multi-bug tasks, ground_truth exposure