fix(evolution): restore GEPA skill evolution on dspy 3.2

[A-KH17]

What does this PR do?

Fixes seven defects that made the Phase-1 GEPA skill-evolution pipeline unusable on the current dspy release (3.2.1). Depending on where execution stopped, users got a silent fallback to MIPROv2, a hard crash, or — worst — an "evolved" skill byte-identical to the baseline (the optimizer never touched the skill text; see fix 6).

Verified end-to-end with these fixes applied: a 150-rollout GEPA run on github-code-review (hermes-agent @ 9acc4b4, moonshot/kimi-k3 optimizer, moonshot/kimi-k2.7-code eval) completed 21 iterations, improved valset 0.474 → 0.579, and produced a genuinely mutated skill (used downstream in NousResearch/hermes-agent#69641).

The bugs

1. dspy.GEPA(max_steps=…) no longer exists in dspy 3.2 (budget args are auto / max_full_evals / max_metric_calls) → TypeError → silent MIPROv2 fallback. Now passes max_full_evals=iterations.
2. GEPA requires an explicit reflection_lm in dspy 3.2; none was passed → construction raised → fallback. Now passes reflection_lm=dspy.LM(optimizer_model, temperature=1.0, max_tokens=32000) (per dspy's own error-message guidance).
3. Metric signature mismatch. dspy 3.2 GEPA requires a 5-arg metric (gold, pred, trace, pred_name, pred_trace); skill_fitness_metric takes (example, prediction) → TypeError → fallback. Wrapped it; the wrapper returns a plain float because dict/Prediction returns break dspy.Evaluate's sum() aggregation.
4. The MIPROv2 fallback itself was broken — optuna is required but undeclared. Added to dev extras.
5. Constraint validation always failed skill_structure because it checked the frontmatter-less body for YAML frontmatter (warned at baseline, hard-stopped the evolved skill). Now validates the reassembled full text.
6. The core flaw: the skill was never optimized. SkillModule passed the skill text as an input field (skill_instructions=…); GEPA mutates signature instructions and demos, never plain inputs — so the skill body stayed frozen and optimized_module.skill_text returned the original text. Now the skill text is installed as the predictor instruction via with_instructions(...), and extraction pulls the optimized instruction from predictor.predict.signature.instructions.
7. skill_fitness_metric crashed on list rubrics — dataset generators may emit expected_behavior as a list of bullet strings (observed with moonshot/kimi-k2.7-code), and the keyword-overlap code called .lower() on it. Lists are now coerced to text.

Changes Made

• evolution/skills/evolve_skill.py — fixes 1, 2, 3, 5, and the extraction half of 6
• evolution/skills/skill_module.py — fix 6 (skill text as optimizable instruction)
• evolution/core/fitness.py — fix 7
• pyproject.toml — fix 4 (optuna in dev extras)

How to Test

1. pip install -e ".[dev]" on a clean env with dspy 3.2.x.
2. python -m evolution.skills.evolve_skill --skill github-code-review --hermes-repo <hermes-agent checkout> --optimizer-model <model> --eval-model <model> --iterations 1
3. Expected: GEPA runs (no "falling back to MIPROv2" message), constraints pass, and output/…/evolved_skill.md differs from baseline_skill.md with a holdout comparison printed.

Notes

• No behavior change to the intended design — this restores the documented GEPA pipeline (README: "GEPA Optimizer ◄── Execution traces") on current dspy.
• Fixes were developed against dspy 3.2.1; the metric wrapper and with_instructions approach follow dspy's public API.

[TurtleMcTurtle]

Review: PR #155 — Most complete GEPA fix

This is the most comprehensive fix among the overlapping PRs (#153, #155, #159, #167). It addresses all three layers of the GEPA compatibility bug:

1. SkillModule fix (root cause) ✅

• Moves skill text from InputField to Signature.instructions via .with_instructions(skill_text)
• This is critical: GEPA optimizes signature instructions, NOT input fields. Without this, GEPA runs but never actually mutates the skill text.
• Uses ChainOfThought (preserves CoT reasoning) vs fix: make DSPy skill evolution persist optimized instructions #153 which downgrades to bare Predict

2. GEPA constructor fix ✅

• max_full_evals + reflection_lm with temperature=1.0 and max_tokens=32000
• 5-arg metric wrapper (_gepa_metric)
• Removes the silent MIPROv2 fallback (good — it was masking the real failure)

3. Evolved body extraction ✅

• Reads optimized_module.predictor.predict.signature.instructions instead of optimized_module.skill_text
• This is the fix for evolved skills coming out byte-identical to baseline

4. Validator fix ✅

• Validates evolved_full (with frontmatter) instead of evolved_body — same fix as Validate the reassembled skill, not the bare body #161
• Also validates baseline with skill['raw'] instead of skill['body']

5. Bonus: list type coercion in fitness metric ✅

Missing vs other PRs:

• No custom API support (fix: GEPA optimizer compatibility + custom OpenAI-compatible API support #167 has this)
• No LLMJudge feedback for GEPA reflections (Fix GEPA integration silently falling back to MIPROv2 on DSPy >=3.1 #159 has this — GEPA benefits from textual feedback, not just bare floats)
• No overfit/collapse guard (fix(skills): extract evolved instruction with overfit/collapse guard #146 has this)
• Adds optuna to dev deps (needed for GEPA)

Suggestion: This should be the base PR for the GEPA fix. Cherry-pick the LLMJudge metric from #159 and the overfit guard from #146 to make it bulletproof. The custom API support from #167 should come as a separate PR.