Eval protocol for a 0.5B coder LoRA (no LLM-as-judge in serve)

[YauhenBichel]

Constraint

PythonVibeGuard is deterministic. We will not add a second model in scripts/serve.py to "score" drafts. Eval can live in docs/ or a script that never ships in the sidecar.

What we do not have yet

• A held-out prompt list (not in all_pairs())
• A pass/fail rubric: runs, type-checks, harness pass, length
• A control: base qwen2.5-coder:0.5b with the same system prompt, no LoRA

Propose a protocol

Keep it cheap: 10–20 prompts, Mac or Linux, Ollama optional. If you need a GPU cluster, say so, but the default should run on a laptop.

Related: #8 (guard evasion), #9 (data vs style prior).