fix(query): prevent generic symbol query flooding via multiplicity penalty and seed deduplication

[devcool20]

Description

This PR resolves Issue #1766, which causes query results to be flooded when codebases contain multiple generic/homonymous names (such as Next.js route handlers named GET or POST, or framework entrypoints like index or handler).

Previously, when a query matched a generic word (e.g., "get"), multiple distinct nodes sharing the same generic label (e.g., GET()) would receive the full _EXACT_MATCH_BONUS and get selected as query seeds. Running a BFS traversal from multiple generic nodes ended up pulling in massive portions of unrelated code, burying the actual specific target.

---

Proposed Changes

1. Label Multiplicity Penalty (graphify/serve.py):

   • Computes and caches a map of normalized label counts (_label_counts_cache) on the graph G.
   • Divides the final matching score of any node by its label multiplicity (score /= multiplicity), ensuring generic framework convention names cannot easily overpower specific identifiers.

2. Seed Label Deduplication (graphify/serve.py):

   • Adds a seen_labels tracking set in _pick_seeds to filter out duplicate labels, guaranteeing that only a single representative node gets seeded for any given homonymous label.

---

Validation & Testing

• Scenario Mock Test: Verified using a Next.js route handler mock graph. Before the fix, querying "get users" seeded users_model, api_users_GET, and api_analytics_GET (flooding the BFS traversal). After the fix, the query correctly seeds only users_model and a single representative GET node (api_analytics_GET).
• Regression Testing: Ran the full serve test suite (pytest tests/test_serve.py). All 80 unit tests passed successfully with 0 regressions.

[safishamsi]

Thanks @devcool20 — the seed-deduplication idea is the right fix and I've landed it in 6ca8604 (ships in 0.9.14), crediting you.

I kept the seed dedup (with two tweaks: the dedup key is normalized so GET/Get/get collapse together, and the per-term guarantee loop honors the same cap), but dropped the second mechanism — the per-label multiplicity penalty (score /= multiplicity). That penalty lives in _score_nodes, which shortest_path and explain also use to resolve endpoints, so dividing scores there silently reweights path/explain endpoint selection (out of scope for #1766 and able to flip which node a path anchors on). The dedup alone bounds the flood. I also added the regression tests the PR was missing. Appreciate the fix.