Feature: Skill display-only output (shown to user, excluded from context)

Problem

Skills often include reference materials (help text, cheatsheets, formatted examples) that benefit the human but don't need to be in the AI's context window. Currently, everything a skill outputs consumes context tokens.

Use Case

A tmux skill wants to show a nicely formatted help reference to the user:

• User asks "tmux help"
• Skill loads and Claude shows formatted help text
• User reads it, gets value
• But Claude already knows tmux - the help text didn't need to be in context

This wastes tokens and duplicates information Claude already has from the skill's instructions.

Proposed Solution

Allow skills to mark content as display-only via frontmatter or inline directive:

Option A: Asset-level control in frontmatter

---
name: my-skill
assets:
  - path: help.md
    display: user-only  # shown to user, excluded from context
---

Option B: Inline directive in skill content

<display-only>
## Quick Reference
| Command | Description |
| ------- | ----------- |
| `tt`    | Toggle tmux session |
...
</display-only>

Relationship to Other Issues

• Different from #15718 (MCP tool ANSI formatting) - this is about skill-controlled content, not MCP tool output
• Addresses the general problem of display/context separation for skill authors

Benefits

• Skills can include rich reference materials without context bloat
• Better UX for help/documentation-style skills
• Skill authors have explicit control over what's contextual vs display-only
• Reduces token waste when Claude already has the knowledge from skill instructions

Prior Art

• Amp has "incognito shell mode" ($$) where shell output is displayed but excluded from context
• OpenCode maintains separate "UI summaries" vs "compaction summaries"
• This would be the first tool to give content authors (skill writers) direct control over display/context separation