Add golden dataset & evaluation framework for agent performance (#308)

[sudo-ss1]

Add golden dataset & evaluation framework for agent performance (#308)

Closes #308.

What & why

Adds a golden dataset of labeled résumés plus an evaluation runner so the agent's
scoring can be measured and compared across models (the issue's motivation:
understanding performance when swapping in ad-hoc models).

It measures at the evaluator boundary: each case supplies a structured
résumé (JSONResume) + optional GitHub data, and the runner compares the
evaluator's output against human-assigned bands and per-category tolerance
ranges. This isolates scoring quality (what changes when you swap models) from
PDF parsing, extraction, and live GitHub calls. All committed résumés are
synthetic (no PII, no real GitHub accounts).

What's included

• evals/golden/*.json — 9 synthetic labeled résumés (3 strong / 3 medium /
  3 weak, incl. edge cases: OSS-only, production-only, tutorial-projects-only).
• evals/run_eval.py — runs each case through the real evaluator; reports band
  accuracy, per-category within-range rate + bias (calibration), Spearman rank
  correlation, and cross-repeat stability; writes a model-stamped JSON report.
  Flags: --repeat, --filter, --out, --min-band-accuracy (regression gate).
• evals/metrics.py — pure, offline metric functions.
• evals/tests/ — offline unit + smoke + dataset-integrity tests (no LLM/network).
• score_utils.py — extracted score aggregation so the CLI and the eval compute
  the total identically (behavior-preserving refactor of score.py).
• README updates + evals/README.md (format, metrics, how to add cases).

Usage

python -m evals.run_eval                 # score the dataset, print metrics
python -m evals.run_eval --repeat 3      # average runs to smooth LLM noise
LLM_PROVIDER=gemini DEFAULT_MODEL=gemini-2.0-flash python -m evals.run_eval
python -m pytest evals/tests             # offline tests, no backend needed

Testing

• 54 offline tests pass; Black clean.
• Refactor verified behavior-preserving (CLI still prints the same total and cap
  warning for a sample evaluation).
• Not run against a live LLM in CI (requires an Ollama/Gemini backend); the tests
  run fully offline via a dependency-injected evaluator.

Notes for reviewers

• Follows CONTRIBUTING: Black-formatted, offline smoke tests exercising each
  stage, issue referenced. No prompt changes, so no before/after prompt examples.
• The seed set is intentionally small and designed to be grown by contributors;
  evals/README.md documents the case format and how to add more.