docs(predict-rlm): split rlm agent skills

[magix022]

Rationale

The repo-scoped rlm skill had grown into a mixed workflow for building downstream RLMs, contributing to predict-rlm/RLM-GEPA, and scaffolding GEPA optimization. That made implicit skill selection too broad and loaded unrelated guidance into tasks that only needed one workflow.

Summary

• Narrow rlm to downstream PredictRLM/RLM package design and build workflows.
• Add standalone rlm-gepa and predict-rlm-contributor skills with one-level references and UI metadata.
• Move detailed API, project-layout, GEPA, data/scoring, and contributor guidance into focused reference files.
• Update README and RLM-GEPA docs to route users to $rlm, $rlm-gepa, or $predict-rlm-contributor based on intent.

Test Plan

• uv run python /Users/emile/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/rlm
• uv run python /Users/emile/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/rlm-gepa
• uv run python /Users/emile/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/predict-rlm-contributor
• git diff --check
• git diff --cached --check

[magix022]

Closing because the split-skill implementation is outdated and is being replaced by a fresh skill-update branch.