feat(agents): add task quality loop auditor

[irvineoy]

Summary

• add the repository-level quality_loop agent, defaulting to Codex
• require GitHub authentication, repository write access, Issues, a clean checkout, and Git author identity before task work starts
• run validator -> one repair attempt for blocking failures -> exactly one optimization iteration -> centralized evaluation -> independent Codex review
• record WARN findings without repair and create fingerprinted, idempotent issues for unrepairable tasks
• harden reproducibly easy (median >=5x) standalone baselines and targeted cases only behind fresh validation and dual original/candidate correctness gates
• keep GitHub credentials outside the GPU container, mount the main checkout read-only, verify the accepted diff on the host, and bundle task commits into one draft PR
• add resumable state, MI300/MI355X examples, Docker/Makefile entry points, documentation, and tests

Validation

• python3 -m unittest discover -s tests -p 'test_quality_loop.py' -v — 15 passed
• pinned ROCm Docker image: python -m pytest -q tests — 81 passed, 3 subtests passed
• bash tests/test_docker_benchmark.sh — passed
• make docker-smoke — passed on MI355X/gfx950
• real Codex task-validator smoke on three representative tasks:
  • hip2hip/gpumode/GELU — PASS
  • triton2triton/vllm/triton_rms_norm — PASS
  • flydsl2flydsl/fp8_gemm_4wave_kernel — WARN only (reference timing methodology); compilation/correctness/kernel timing passed

Dependency

This draft is based on #75 so its diff contains only the quality-loop implementation. Retarget it to main after #75 merges.