Fix README: remove submission badge, correct heuristic scores, fix lookup_gr description

Improve README for Phase 3 judging: architecture diagram, reward rationale, all 5 required sections

Fix test: update perfect score assertion from 1.0 to 0.999 after clamping

Fix: clamp grader scores to strictly (0, 1) — evaluator requires 0 < score < 1

Fix: align port to 7860, add warmup LLM call before environment interaction

Fix: use AWS ECR Public mirror instead of Docker Hub (evaluator cannot reach registry-1.docker.io)

Fix: pin Python base image to 3.11.9-slim-bookworm to resolve Docker Hub manifest error

Fix: remove HF_TOKEN fallback to stop bypassing evaluator's LiteLLM proxy

Nuclear fix: simplify credentials, add aggressive debug logging

Match exact credential pattern from dashboard checklist

Fix: use evaluator API_KEY before HF_TOKEN, create client inside main()

Match official stdout format + support IMAGE_NAME/ENV_URL/SPACE_URL

Update README baseline scores to actual measured values

Major upgrade: 4 invoice templates, 3-way matching, confidence scoring, 113 tests

Fix openenv validate: add uv.lock, server/app.py, project.scripts, openenv-core>=0.2.0

Fix: all POST endpoints tolerate empty/missing body

Fix: use openenv-core create_app framework for standard validator compliance

InvoiceAgent: OpenEnv environment for invoice processing