docs(skill): teach the lean search_code workflow

[artur-firsttouch]

Problem

The installed codebase-memory skill routes text searches to search_code, but it does not explain how to use the lean output controls now available on main. In practice, agents can respond to broad results by raising limit or combining mode: "full" with a high limit, producing unnecessarily large MCP responses. Literal alternation such as foo|bar is also easy to send without regex: true.

The current release, v0.10.8, does not expose the newer controls. Current main already implements the server-side pieces: max_output_tokens, independent ranked/raw/directory pagination, bounded source and raw-line expansion, semantic truncation metadata, and the literal-pipe warning. This proposal does not change the MCP API or handler behavior.

Proposal

Add a concise Code Search Workflow section to the installer-owned embedded codebase-memory skill in src/cli/cli.c:

1. Use search_graph(semantic_query=[...]) for broad natural-language discovery.
2. Use search_graph for known or partial symbols.
3. Use search_code for exact identifiers, literals, errors, and settings.
4. Start with compact or files, result_limit: 10, and max_output_tokens: 3200.
5. Narrow with path_filter and file_pattern before paging via returned offsets.
6. Treat omissions as a narrowing signal rather than immediately raising limits.
7. Reserve full for one to three expected hits and bound source_max_lines.
8. Require regex: true for patterns such as foo|bar.
9. Explicitly call out full + high limit as an anti-pattern.

Keep the always-loaded agent instructions unchanged so detailed usage guidance costs context only when the skill is loaded. Extend cli_skill_files_content to contract-test the key guidance.

Non-goals

• No MCP schema or behavior changes.
• No new include_tests or generated-file controls. Existing scoped search, ignore rules, and test-result ranking remain unchanged.
• No default changes or compatibility impact.

Would this focused skill-guidance change be acceptable?