fix: use LLM judge feedback for skill fitness

[steezkelly]

Summary

Fixes #12 by wiring the existing rubric-based LLM judge into the skill optimization metric:

• makes skill_fitness_metric(...) use the currently configured DSPy LM as the primary judge
• returns dspy.Prediction(score=float, feedback=str) so GEPA can use reflective feedback instead of generic score-only feedback
• preserves the previous keyword-overlap score as an offline/rate-limit fallback with explicit feedback
• keeps holdout aggregation numeric by coercing metric results with float(...)
• adds regression tests for judge scoring, fallback scoring, and empty-output scoring

Root cause

The codebase already had LLMJudge, but the actual optimizer metric still used only keyword overlap. That produced a narrow and easily gamed objective and deprived GEPA of actionable feedback.

Test Plan

• RED first: pytest tests/core/test_fitness.py -q failed before _score_with_llm_judge existed
• pytest tests/core/test_fitness.py -q
• pytest -q
• runtime probe: no configured LM returns dspy.Prediction(...), float(prediction) works, and fallback feedback is populated
• static added-line security scan: clean
• git diff --check

Result: 142 passed, 11 warnings (DSPy deprecation warnings only).

Closes #12

[steezkelly]

Closing this PR in favor of consolidated PR #68. Local integration found real helper-block overlap in evolution/skills/evolve_skill.py across the stack, and #68 preserves local test evidence: targeted stack tests 41 passed; full suite 164 passed; GitHub checks were absent on the split PRs. Review #68 instead.