Fix DSPy 3.2+ API compat (max_full_evals + metric kwargs) and 2 evolve_skill validator bugs

[0L1v3DaD]

Summary

Four small but high-impact fixes to make evolution.skills.evolve_skill
work end-to-end on a fresh install with DSPy 3.2+. Tested empirically with
a 5-iteration GEPA run on a real skill against ~30 evaluation examples.

Bugs fixed (one commit each)

#	File	Symptom before fix
1	evolution/skills/evolve_skill.py	TypeError: GEPA.compile() got unexpected keyword 'max_steps' — DSPy 3.2 renamed it to max_full_evals and added a required reflection_lm argument
2	evolution/core/fitness.py	TypeError: skill_fitness_metric() takes 2-3 args but 4 were given — DSPy 3.2 GEPA passes (gold, pred, trace, pred_name, pred_trace), the metric only accepted 3
3	evolution/skills/evolve_skill.py	Silent no-op: optimizer "succeeded" with full convergence reported but the saved file was byte-identical to the input. Cause: read optimized_module.skill_text (the input field) instead of optimized_module.predictor.predict.signature.instructions (the actual evolved prompt)
4	evolution/skills/evolve_skill.py	False-negative validation: skill_structure: missing YAML frontmatter even though the file written to disk had perfect frontmatter. Cause: validator.validate_all(evolved_body, ...) was called on the body-only string a few lines before reassemble_skill() prepended the frontmatter

Why this matters

The silent no-op (commit 3) is the worst of the four because it's invisible:

• No exception is raised
• All constraint gates pass (the unchanged baseline trivially satisfies size / growth / structure gates)
• The file written to disk is a valid skill, just identical to the input
• Significant token spend per run for zero learning

Once that's fixed, the validator wiring bug (commit 4) becomes visible
and easy to fix.

Verification

After all four commits applied:

• 5-iteration GEPA run completes without errors
• Real evolved prompt: ~6.5KB (vs ~211-char baseline prompt) — substantial growth in the evolved instruction prefix
• Skill body preserved verbatim — correct, GEPA never had access to mutate it
• 4/4 constraint gates pass: size_limit, growth_limit, non_empty, skill_structure
• Manual diff of evolved vs baseline confirms real behavioral changes (response shape, anti-pattern guidance, etc.)

Backward compatibility

All four fixes are additive or substitute-equivalent on the DSPy 3.2+ path:

• max_full_evals is the new name; max_steps was dropped, no compat shim possible
• reflection_lm is now required by GEPA; we now provide one explicitly using the existing config.optimizer_model
• The metric signature accepts the extra DSPy 3.2 kwargs but defaults them to None, so MIPROv2 / BootstrapFewShot (3-arg callers) still work
• The validator change just hands a strict-superset string to validate_all — all constraint checks remain valid

Test plan

Reproduced before and after each individual commit:

1. After commit 1: GEPA initializes without TypeError on max_steps
2. After commit 2: metric runs without TypeError on extra positional args
3. After commit 3: evolved file size differs from input file size (real artifact, not a copy)
4. After commit 4: 4/4 gates green on the real evolved artifact