MCP: Fix get_declaration leaking internal Debug representation for Unresolved ancestors

[bnferguson]

Background

While working with the MCP Claude was complaining about debug output in get_declaration, digging into it it seems that when you have a ancestor that is unindexed (for instance, something defined in a gem), rather than say what the ancestor is rubydex prints the debug info. For instance for a class like class Dog < ActiveRecord::Base; end

you get:

      "ancestors": [{
        "name": "Unresolved(Name { str: Id { value: 13424087348773472403, _marker: PhantomData<...> }, parent_scope: Some(Id { value: 13486303262390686460,
  _marker: PhantomData<...> }), nesting: None, ref_count: 1 })",
        "kind": "Unresolved"
      }]

instead of:

      "ancestors": [{
        "name": "ActiveRecord::Base",
        "kind": "Unresolved"
      }]

The latter seems to make a bit more sense to me but I don't know if perhaps there's reasoning I'm missing for the former. So feel free to close this (obvs) if this is way off base and not in fact a bug or a fix that makes total sense.

Approach

Went with a very simple/minimal fix by reusing Graph::build_concatenated_name_from_name, which already existed in stats/orphan_report.rs for the orphan report. It walks the parent_scope chain off the Name and concatenates each segment into a qualified constant path.

Buuuut to call it from the MCP crate, its visibility moves from pub(crate) to pub.

The architecture docs note that model/graph.rs is where all the Graph-y public bits are so it seems like this my make sense to be moved over there. But in the interest of keeping this PR for a bug that may not even be a bug I kept the fix narrow. Happy to move things around.

There's a slight behavior change with this approach too. Previously, if graph.names().get(name_id) returned None, the ? would short-circuit filter_map and the ancestor would silently vanish from the response. After this change, that case renders as "<unknown>" (the existing fallback inside build_concatenated_name_from_name) and the ancestor stays in the array. TLDR; a missing name is now visible rather than swallowed. That seems preferable to me at first glance but may not be desired.

I do believe old swallowing behavior could be maintained pretty easily.

Testing

Added a regression test:

cargo test -p rubydex-mcp get_declaration_unresolved_ancestor_renders_qualified_name

Also if you want to test manually, index a workspace whose classes inherit from gem-provided base classes (any ActiveRecord model should do), and call get_declaration. You should no longer see the debug output but instead see ActiveRecord::Base

Notes

(For the cross linking context sickos out there [that's me])

• Distinct from Create TODODeclaration for unresolved constant references #707 (resolver creating TODODeclarations for unresolved references), that's about why something is unresolved; this is about how the Unresolved variant is serialized to the MCP client.
• Distinct from Singleton class ancestor chain is broken when intermediate classes have no singleton definitions #708 (singleton class ancestor chain correctness), different failure mode.
• Not currently tracked in MCP server follow-up improvements #608 (MCP server follow-up improvements).
• Related: index_workspace fails with FileError on the rails meta-gem #803 (index_workspace doesn't index gems) explains why these ancestors are unresolved

[bnferguson]

I have signed the CLA!