fix(evolve): GEPA never executed and evolved output always failed validation

[acidsound]

Summary

Three bugs in evolve_skill.py meant the Phase 1 pipeline could never produce a valid evolved skill:

1. Validation ran against a body-only artifact

The evolved text (evolved_body) is the skill body with frontmatter stripped, but skill_structure requires YAML frontmatter (---, name:, description:). Every evolved variant failed this check by construction — no improvement could ever deploy.

Fix: validate the reassembled file (reassemble_skill(frontmatter, body)).

2. GEPA never actually ran

dspy.GEPA(metric=..., max_steps=iterations)  # TypeError: unexpected kwarg

Current DSPy's GEPA has no max_steps; every run hit the except block and silently fell back to MIPROv2 — despite GEPA being the project's headline engine. Also:

• GEPA's metric signature is (gold, pred, trace, pred_name, pred_trace), not DSPy's 3-arg form → adapted via wrapper
• GEPA raises without a reflection_lm → wired to optimizer_model

3. Evolved text was read from a field GEPA never mutates

optimized_module.skill_text is untouched by optimization; GEPA evolves the predictor's signature instruction. Verified on real runs: baseline/evolved files had identical md5 while reporting +0.006/+0.024 "improvements" (judge variance). Now extracted via named_predictors().

Verification

• Full test suite: 145 passed
• Real run (hermes-agent skill, 3 iterations): holdout 0.599 → 0.628, size 12,598 → 3,027 chars, constraints PASS, non-empty diff present
• Before fix: same run reported fake improvements with md5-identical output

Notes

• The MIPROv2 fallback remains as a safety net but no longer triggers.
• Related: computer-use SKILL.md (18.5KB) exceeds the 15KB size constraint at baseline — separate concern, not addressed here.