feat: eval harness scaffold (runner, judge, report, models) + 1 example golden case + nightly workflow_dispatch

[constk]

Problem

The eval flywheel is the harness's most distinctive feature for LLM-driven systems. Documenting the pattern without scaffolding code leaves users to reinvent it.

Proposed solution

Port src/eval/{__main__.py, runner.py, judge.py, report.py, models.py} from Teller. Generalise the LLM client via the LLM_PROVIDER config from #18 (no Azure/OpenAI lock-in). Add eval/golden_qa.json with ONE trivial example: {"id": "echo-hello", "question": "echo hello", "expected": "hello", "tolerance": "exact_match"}. Add eval/test_golden_qa.py runner using @pytest.mark.eval. Port .github/workflows/eval-nightly.yml BUT replace schedule with workflow_dispatch-only. Document opt-in (adding LLM secrets, re-enabling schedule) in docs/EVAL_HARNESS.md (#25).

Acceptance criteria

• uv run pytest eval/ runs the example case and passes (no LLM call needed for exact_match).
• eval-nightly.yml only triggers on workflow_dispatch.
• Inline comment in the workflow explains how to flip on schedule: once secrets are configured.
• src/eval/judge.py LLM-judge interface is provider-agnostic (calls a LLMClient Protocol).

Priority rationale

High: the eval harness is the hook that makes this template stand out for LLM-system engineers.

Depends on

#18, #19

[constk]

Merged in PR #60.