Nightly memory tidy-up: merge the same fact saved in two languages #353

Follow-up to #351. To avoid losing memories, a Hindi fact is never skipped as a copy of an English one, so the same fact can be stored twice, e.g. "माया शाकाहारी है" and "Maya is vegetarian".

Want: the nightly memory tidy-up (dreaming) finds such pairs and merges them into one memory, keeping the clearer wording and both sources.

• Candidate pairs: same person (the cross-script name matching from fix(memory): never merge a Hindi memory into a fact about someone else #351) and close meaning. Use the multilingual embedding model (bge-m3) when it's set; otherwise ask the fast model directly whether the two say the same thing.
• Merge only when the model is sure they say the same thing, never when one adds a detail (Saving a new preference can overwrite a different one #257's rule) or they conflict (that's the contradiction path, Memory: catch contradictions in other languages and long sentences #95 and feat(memory): catch contradictions in other languages and long sentences #338).
• Keep it reversible: the merged-away memory goes to the review history with "merged into …", and the user can undo.
• Works with qwen3:8b; keep the prompt short.

Done when: with mixed Hindi and English facts about Maya, a real tidy-up run merges "माया शाकाहारी है" with "Maya is vegetarian", leaves different facts alone, and the merge can be undone.