fix: constraint validation target + GEPA compatibility with dspy 3.3.0

[luckystar2026]

Title: fix: constraint validation target + GEPA compatibility with dspy 3.3.0

Branch: fix/gepa-compat-and-constraint-validation

Summary

Four small fixes in evolution/skills/evolve_skill.py discovered while integrating on Windows with dspy 3.3.0 / gepa 1.x:

1. Constraint validation was checking the wrong text (always FAILED)

_check_skill_structure validates a full SKILL.md (expects --- frontmatter + name + description), but evolve_skill.py passed the bare body:

# before
baseline_constraints = validator.validate_all(skill["body"], "skill")
...
evolved_constraints = validator.validate_all(evolved_body, "skill", baseline_text=skill["body"])

A bare body never starts with ---, so every run — including perfectly valid ones — reported skill_structure FAILED and the evolved skill was saved as evolved_FAILED.md.

# after
baseline_constraints = validator.validate_all(skill["raw"], "skill")
...
evolved_constraints = validator.validate_all(evolved_full, "skill", baseline_text=skill["raw"])

2–4. GEPA API drift (dspy 3.3.0)

The repo targets an older GEPA API. On dspy 3.3.0 all three of these throw, the except Exception silently falls back to MIPROv2 (the fallback message is easy to miss in rich-console logs), and GEPA's main path never runs:

• max_steps=N → max_full_evals=N (TypeError: GEPA.__init__() got an unexpected keyword argument 'max_steps')
• Metric protocol: dspy 3.3.0's GEPA requires a 5-arg metric (gold, pred, trace, pred_name, pred_trace). skill_fitness_metric takes 3 — added a thin adapter:

  def _gepa_fitness_metric(gold, pred, trace=None, pred_name=None, pred_trace=None):
      return skill_fitness_metric(gold, pred, trace)

• reflection_lm is mandatory (AssertionError: GEPA requires a reflection language model…):

  reflection_lm=dspy.LM(eval_model, temperature=1.0, max_tokens=4096),

Verification (Windows, dspy 3.3.0, DeepSeek, arxiv skill)

• pytest tests/ -q: 144 passed / 1 failed (test_resolve_expands_user_home — pre-existing POSIX path assumption, unrelated)
• GEPA main path now runs: 9 iterations, 68/75 rollouts, dspy.teleprompt.gepa.gepa: Iteration N: New subsample score … visible in logs
• All 4 constraint gates pass (size ≤15KB, growth ≤+20%, non-empty, structure)
• Holdout behavior score 0.431 → 0.564 (+30.7%)

Note

This PR fixes the pipeline so evolution runs and gates correctly, but skill text itself still never changes on disk — that is a separate design issue in SkillModule (skill_text is an input field, not the optimized instruction). Tracked in issue #172.