[bench] Run ProgramBench live — first 10 instances, GLM-5.1 + qwen3-coder

[0bserver07]

Goal

Drive a small live ProgramBench run end-to-end to validate the W13-B1 (orchestration adapter) + W14-8 (inference loop) work against a real model on real task instances.

Scope

• First 10 ProgramBench instances (smallest / fastest)
• Two model configurations to compare:
  • GLM-5.1 via Ollama Anthropic-compatible bridge
  • qwen3-coder-30b (local)
• Use the swe-agent.md preset already in chimera/agents/presets/

Prerequisites

• linux/amd64 Docker (Colima or Docker Desktop with emulation on macOS, native on Linux)
• pip install programbench (upstream grader)
• ProgramBench Docker images (pulled on demand by the harness)
• CHIMERA_PROGRAMBENCH_LIVE=1 env to enable the live test gate

Acceptance criteria

• Live run completes against ≥ 5 of the 10 instances (some may time out — acceptable)
• Pass/fail numbers logged + committed to data/programbench-live-N.jsonl
• If failure rate is unexpectedly high, file follow-up issue with diagnosis
• Update docs/benchmarks/programbench-howto.md with the "first live run" recipe

Files

• chimera/eval/benchmarks/programbench.py (already exists)
• chimera/agents/presets/swe-agent.md (already exists)
• tests/eval/test_programbench_inference.py (already exists, gated)

Estimated effort

Medium — most of the code is done. Risk is Docker setup + per-instance time (5–30 min each, so 10 instances ≈ 1–5 hours).