MCP server follow-up improvements

[st0012]

From PR #586 review — deferred suggestions and ideas for future iterations.

New tools

• resolve_at_position — Given file + line + column, return the declaration being referenced. The most powerful navigation primitive for LLM debugging workflows. (Morriar)
• find_method_references — Expose graph.method_references() as a name-based search for method call sites (e.g., "find all calls to validate"). Not fully resolved, but very useful. (Morriar)
• get_diagnostics — Expose graph.all_diagnostics() to return parse/type errors. Lets an LLM understand what's broken before editing. (Morriar)

Tool enhancements

• Include method parameters in get_declaration — MethodDefinition#parameters() data exists but isn't serialized. LLMs need signatures without reading source. (Morriar)
• Reverse index for find_constant_references — Build declaration_id → Vec<const_ref_offset> at startup for O(k) lookups instead of O(n) scans. Addressed by using declaration.references() which already tracks resolved references per declaration. (Morriar, vinistock)
• Audit codebase_stats tool usefulness — Is aggregate codebase info useful to agents, or only to tool developers? (vinistock)
• Type-safe Location to prevent double to_presentation — Follow up similar to PR Create DisplayLocation #607 to make it impossible to call to_presentation() multiple times and accidentally bump line/col values. (vinistock)

Code organization

• Split server.rs into separate files per tool handler — Prevents the file from growing too large and enables unit testing each handler independently. (vinistock)
• Add unit tests alongside integration tests — Current tests all spawn the server and do real IO. Add unit tests for tool logic, keep a few integration tests. (vinistock)