Hindi memories get merged into the wrong facts on the default local models #339

Seen in the real test for #338, on main as well. When Hindi facts are saved with qwen3:8b and nomic-embed-text:

• "प्रिया पुणे में रहती है" (Priya lives in Pune) and "प्रिया अब दिल्ली में रहती है" (Priya now lives in Delhi) were filed as updates to the English "Riya lives in Pune".
• A Hindi "Karan went to Goa" was skipped as a duplicate of a Kabir fact.

nomic-embed-text scores Hindi text close to unrelated English and Hindi facts, and the small model then chooses UPDATE or skip.

Ideas: require the same subject (the #257 same-thing guard) before an UPDATE or duplicate skip across scripts, and consider a multilingual embedding model option (e.g. bge-m3 or multilingual-e5) in the hardware recommendations.

Done when: those Hindi facts are saved as separate memories with qwen3:8b, checked in a real run.