feat: classification persistence via classifications.jsonl

[ccage-simp]

Description

This PR implements "Classification Persistence" to protect classification data if the SQLite database is dropped, and to allow classifications to be synced across machines.

It introduces a secondary source of truth: ~/.ft-bookmarks/classifications.jsonl.

The Use Case: Multi-Machine Syncing

Users who sync their ~/.ft-bookmarks/ folder across computers via Git currently lose their classifications because bookmarks.db is ignored. Running ft index on a second machine builds a fresh database with zero classifications, leading to redundant LLM usage to re-categorize the same data.

The Implementation

1. The Save Hook: Modified classifyWithLlm and classifyDomainsWithLlm to append a ClassificationRecord to classifications.jsonl whenever a batch successfully returns from the LLM.
2. The Self-Heal Hook: Modified buildIndex in src/bookmarks-db.ts to read classifications.jsonl into memory prior to rebuilding. Unclassified records that have a backup in JSONL are automatically hydrated with their categories and domains.
3. The Export Utility: Added ft classify --export to allow existing users to dump their current SQLite classifications into the JSONL backup file.

This ensures that classifications are preserved during index rebuilds and can be tracked in Git alongside bookmarks.jsonl for easier cross-machine syncing.