feat: dual-mode — MCP server and CLI skill in one file

Expose the same 66 Gitea tools two ways from server/gitea_mcp.py:
- MCP server (no subcommand -> mcp.run()), as before
- CLI: `<tool> --flags` prints one JSON object {success, data/error}

A new @tool decorator registers each function in both FastMCP and a TOOLS
registry; a CLI dispatcher builds argparse from each function's signature and
wraps results in the repo's JSON+success contract. Adds `list-tools` for
discovery. No tool logic duplicated.

- Rewrite skills/gitea/SKILL.md for CLI-first usage (MCP kept as an option)
- Sync version to 0.3.0 across plugin.json, pyproject.toml, SKILL.md
- Add tests/unit/test_cli.py (dispatch routing, flag typing, errors, main)
- Document dual-mode in README.md and CLAUDE.md
- Ignore coverage artifacts

Tests: 119 passed, 93% coverage.

README.md

@@ -2,6 +2,26 @@
 
 Full Gitea platform management — 66 tools for repositories, issues, PRs, branches, files, releases, webhooks, and organizations.
 
+## Two ways to run (one source file)
+
+`server/gitea_mcp.py` is both an MCP server and a CLI. The 66 tool functions are
+shared; only the entrypoint differs:
+
+- **MCP server** — registered via `.mcp.json`, tools appear as `mcp__gitea__<tool>`.
+  This is how you connect it to an MCP-aware client.
+- **CLI skill** — `uv run server/gitea_mcp.py <tool> --flags` prints one JSON
+  object on stdout. This is how Cline (and the `skills/gitea` skill) uses it.
+
+```bash
+# CLI examples
+uv run --project . server/gitea_mcp.py --config ./config.json list_instances
+uv run --project . server/gitea_mcp.py --config ./config.json get_repo --owner andrey --repo infra
+uv run --project . server/gitea_mcp.py list-tools     # self-describing catalogue
+```
+
+CLI output contract: exactly one JSON object — `{"success": true, "data": …}` or
+`{"success": false, "error": …}`. Check the `success` field, not the exit code.
+
 ## Setup (one-time)
 
 ### 1. Edit config.json