feat: hybrid memory search with FTS5 + vector merge

[bug-ops]

Summary

• Add SQLite FTS5 virtual table (messages_fts) with auto-sync triggers for keyword search over messages
• SemanticMemory::recall() now runs both Qdrant vector similarity and FTS5 BM25 keyword search, merging results with configurable weights (default 0.7 vector / 0.3 keyword)
• When Qdrant is unavailable, falls back to FTS5-only keyword search instead of returning empty results
• New vector_weight/keyword_weight config in [memory.semantic]

Changes

File	Change
migrations/008_fts5.sql	FTS5 virtual table + backfill + INSERT/UPDATE/DELETE triggers
sqlite/messages.rs	keyword_search() method with optional conversation filter
semantic.rs	Hybrid recall with weighted merge, with_weights() constructor
config/types.rs	vector_weight/keyword_weight fields on SemanticConfig
main.rs	Wire config weights to SemanticMemory::with_weights()

Test plan

• 4 new keyword search tests (basic, filter, no match, limit)
• 4 new hybrid recall tests (FTS5 fallback, filter, no match, limit)
• All 1435 existing tests pass
• Clippy clean, fmt clean

Closes #372,Closes #373,Closes #374,Closes #364
Epic: #364