test(memory/graph): regression test for entity-graph seed stopword filter

[skynetcmd]

Summary

Adds the missing regression test for the entity-graph seed stopword filter (shipped in #13 / dd6f1ed). That fix landed with a pre-registered metric but no test gated it — a §3/§11 (robustness / bench-discipline) behavior-baseline gap per docs/DESIGN_PHILOSOPHIES.md. This closes it.

Why it matters

The fix prevents sentence-initial Title-Case common words ("Can", "What", "How", "I") from being captured as entity-graph seeds, where "Can" LIKE-matched "Canada"/"Canva" and pulled spurious BFS neighbors into the top-k. Measured impact (from #13): −13.3pp → 0pp at session-hit-rate@k=5 on single-session-preference questions, +1.4pp @k=3 overall. Nothing locked that behavior until now.

What the test does

Exercises the real _ENTITY_MENTION_RE + _QUERY_STARTER_STOPWORDS from memory.graph (no DB / embedder), replicating the Step-1 seed-extraction loop, and asserts:

• common query starters (Can/What/How/I/…) are in the stoplist;
• "Can you recommend a show like Stranger Things" → seeds keep Stranger Things, drop Can (the canonical false-match regression);
• a query of only starters yields zero seeds (BFS short-circuits);
• a genuine Title-Case entity not in the stoplist survives.

Testing

• 16 tests, pass. No DB, no network — fast (§8).
• Mutation-verified: emptying the stoplist turns the suite red, so it genuinely gates the fix (§3).
• Runs clean alongside test_entity_graph.py (52 passed together); ruff clean (§13).

🤖 Generated with Claude Code