fix: load_skill handles dir paths; validate evolved_full not body

[shilei103-afk]

Summary

Two bugs prevented the self-evolution pipeline from completing:

Bug 1: load_skill() receives directory path, reads empty content

In evolution/skills/skill_module.py, find_skill() returns a directory path (e.g. /skills/github/github-code-review/). load_skill() tried to read this path as a file directly, got empty content, and all constraint checks failed silently.

Fix: detect when path is a directory and append /SKILL.md.

Bug 2: validate_all() called on evolved_body instead of evolved_full

In evolution/skills/evolve_skill.py, constraint validation was called on evolved_body (markdown body only, no frontmatter), so skill_structure constraint always failed because it checks for YAML frontmatter --- and name: field.

Fix: validate evolved_full (frontmatter + body) instead.

Bonus: holdout eval crash now caught

DSPy JSONAdapter crashes when the skill outputs bash commands. This is now wrapped in try/except so constraints-passed skills are always saved.

Testing

• Dry run passes: ✓ skill_structure: Skill has valid frontmatter (name + description)
• Optimization completes with all constraints passing
• Output saved to output/github-code-review/<timestamp>/

[steezkelly]

I did a quick pass over this because it overlaps with the local/fork fixes from #38/#49.

Useful parts:

• The evolved_full validation change is the right root-cause fix for Bug: validate_all checks evolved_body instead of evolved_full, causing all evolved skills to fail skill_structure #34/Constraint validator rejects every evolved skill: checks frontmatter on body-only text #11. Validating body-only markdown will always trip skill_structure because the YAML frontmatter has already been stripped.
• Passing baseline_text=skill["raw"] is also more consistent if the candidate being checked is the full reassembled skill.

One thing to double-check before merge:

• On current upstream main, find_skill() appears to already return the concrete SKILL.md path, not the skill directory. So the load_skill() directory fallback is harmless defensive hardening, but I could not reproduce the claim that upstream find_skill() itself returns a directory path.

Potential concern:

• The broad except Exception around holdout evaluation may make the pipeline report/save a candidate with 0.0/0.0 scores if the real evaluator is broken. That is useful for avoiding artifact loss, but it weakens the deploy/quality signal. I would consider splitting that into a separate PR or recording holdout_evaluation_failed: true in metrics so downstream automation does not treat the candidate as cleanly evaluated.