fix(agent-surfaces): apply the danger policy to markdown and MCP, escape llms-full metadata

[mrgoonie]

Closes #91

Changes

• New shared markdown/listing-text.ts: isDangerListing and inlineText.
• /skills/<slug>.md and MCP get_package now omit the SKILL.md body for risk_label = 'danger' (warning on .md; contentIncluded: false plus contentOmitted on MCP), matching llms-full.txt.
• llms-full.txt (and the llms.txt index and .md header) flatten name, author, category and description to one line and escape a leading #, > and backticks.
• docs/api-reference.md: documents the policy and that unknown (pre-scanner) rows, like caution and safe, keep their bodies.

Tests

• pnpm typecheck: pass
• pnpm test: 94 files, 2053 passed, 2 skipped
• pnpm --filter web build: pass
• New tests cover danger omission on all three surfaces, unknown/caution keeping bodies, and metadata injection (newlines, leading #/>, fence-opening backticks).

🤖 Generated with Claude Code