Skill marker has no automatic retrieval loop

[JordanCoin]

Problem

After the push-to-pull optimization (commit a323634), prompt-submit emits only <!-- codemap:skills [...] --> markers. The full skill bodies are available via codemap skill show or MCP get_skill, but nothing automatically follows up on the marker.

Claude sees the marker but has no instruction to fetch the full skill body. This means safety-critical guidance (like hub-safety) may not be consulted.

README still says "injects relevant skills automatically" which is now inaccurate.

Options

1. Add a one-line instruction in the intent marker telling Claude to fetch skills when risk is medium/high
2. Only push-inject for high-risk situations (hub edits), pull for everything else
3. Update CLAUDE.md template to instruct Claude to call codemap skill show when skills are matched

Context

Trade-off: push = 87% more tokens per prompt, pull = safety guidance may be missed. Need a middle ground.