search: query-time candidate-set resolver over raw edges

[IceRhymers]

Part of #82. Depends on #83 and #84.

Summary

Build the query-time resolver that turns raw edge target names into ranked candidate-set payloads, in the shared service layer.

Scope

• Resolver module (e.g. app/search/references.py) joining edges.target_name against symbols.name: exact match primary, same-repo preferred with cross-repo fallback, ranking signals (kind match, same-file, same-repo). Use the two-query design from app/search/symbols.py — its docstring documents the SQLAlchemy auto-correlation hazard that a self-referencing join makes worse.
• Payload builders in app/service.py (both MCP and webui consume these): candidate sets with (repo, path, line, edge_kind, enclosing_symbol, candidates[]), truncated/truncation_reason flags, deterministic ordering, keyset-friendly.
• Branch scoping byte-identical to search_code: explicit branch: → branches @> ARRAY[:v]; otherwise the implicit default-branch conjunct (coalesce(default_branch, 'HEAD')).
• Quality measurement: record the unique/ambiguous/unresolved distribution on the indexed corpus against the deep-dive probe baseline (28.8% / 33.4% / 37.8% on the 4-repo sample).

Acceptance criteria

• Resolver returns ranked candidate sets; ambiguity is represented, never silently collapsed to one target.
• statement_timeout guard and truncation flags honored; ordering deterministic.
• Branch/default-branch scoping proven by tests matching search_code behavior.
• Resolution-quality distribution measured and recorded in the PR.

[IceRhymers]

Completed via child PR #93, squash-merged into the #82 integration branch (integration/knowledge-graph-reference-edges). The umbrella draft PR #89 remains open until the full epic is integrated and human-reviewed into protected master.