Reduce cognitive load in the sub-agent syntax side quest

[Copilot]

The sub-agent syntax side quest had drifted into a dense reference page: too many rules, too much exposition, and too little learner action. This change refactors it into a short repair exercise that teaches the same core constraints through one concrete workflow edit.

• What changed

  • Replaced the rule-heavy reference structure with a single broken-to-correct sub-agent example
  • Reduced scope to the three constraints learners actually need to internalize first:
    • valid sub-agent naming
    • supported sub-agent frontmatter
    • block placement at the bottom of the workflow

• Learning flow

  • Learners now fix one intentionally broken block in sequence instead of paging through multiple rule sections
  • The exercise ends by applying the pattern to their own Step 21 workflow, keeping the side quest anchored to a real artifact

• Content cleanup

  • Tightened explanations around compiler behavior for unsupported sub-agent fields
  • Clarified why ## headings terminate sub-agent blocks
  • Simplified the checkpoint to match the concrete edits the learner makes

• Example

  Write a daily issue digest.
  
  ## How to use this workflow
  
  Run it from GitHub Actions.
  
  ## agent: `issue-summarizer`
  ---
  description: Summarizes one issue in one sentence
  model: small
  ---
  
  Read one issue and return exactly one sentence.