Implement bounded repository navigation with repo-lens

- Add map, find, and around commands with repository root discovery
- Stream literal ripgrep searches with match counts and result limits
- Add direct subprocess execution and explicit error returns
- Disable C++ exceptions and add navigation tests
- Support CMake installation and document builds and deployment
- Add a reusable agent skill and document skill discovery

AGENTS.md

@@ -3,7 +3,7 @@
 - Keep engineering simple: prefer deterministic local queries, bounded output, composable commands, and explicit context limits. Orchestrate proven tools before reimplementing them; semantic tooling must remain optional.
 - Avoid speculative infrastructure, dependencies, and abstractions. Introduce complexity only for a demonstrated need; source and Git are canonical.
 - Write clear, maintainable native C++. Follow surrounding code and existing formatting/tool configuration. Do not invent detailed conventions; document project conventions as they actually emerge.
-- Use trailing return types where applicable. Declare `execute` overrides `[[nodiscard]]` and `noexcept`, matching the interface. Keep `noexcept` contracts honest: recoverable execution failures must be returned as exit codes rather than escaping exceptions.
+- Use trailing return types where applicable. Declare `execute` overrides `[[nodiscard]]` and `noexcept`, matching the interface. Keep `noexcept` contracts honest: exceptions are disabled; use explicit status returns and non-throwing filesystem overloads for recoverable failures. Do not use `try`, `catch`, or `throw`.
 - Start with `docs/PROJECTMAP.md`. Prefer available symbol lookup for known symbols, textual search for known text, and sparse semantic bookmarks for architectural concepts without suitable symbols. Search, narrow, inspect a bounded range or containing symbol, then expand only as needed. Justify whole-file reads.
 - Keep context small: summarize large result sets, report limits and omitted results, and inspect callers, references, tests, or related code only when relevant. Never claim semantic classifications that the underlying evidence cannot establish reliably.
 - Inspect the working tree before editing; preserve existing human changes. Keep changes focused and avoid unrelated cleanup.