File import

Drag-and-drop upload → parse → map → filter → preview → import. UI lives at /import.

• 4 formats: CSV (RFC 4180, BOM, multi-line values), TSV, JSON array, JSONL (with # comments).
• Mapping: each file column → a table column + an optional SQL transform (UPPER(value), CONCAT(...)).
• Row filters: boolean expressions, so only matching rows get imported.
• On-duplicate: error, skip, update.
• max_rows and dry_run for safely testing a mapping before the real import.
• Embeddings at import time — import data and build vectors in one action.
• Saved profiles — mapping settings persist in <db>.fdb.import and get reused for repeated imports.

AI-suggested column mapping currently runs on name-matching heuristics; the fields for wiring in an LLM (AISuggestedMappings, AIPromptOverride) are already there for a future iteration.