Mining agent transcripts: filter boilerplate before ranking by keyword

When reconstructing project history from Claude Code session .jsonl transcripts, raw keyword counts mislead because global instructions and hook feedback repeat the term in every session. Parse each line as JSON and keep only user-typed strings and assistant text blocks, excluding <system-reminder> and command wrappers, before ranking files. Also check the tool state dirs (config/state/logs under ~/.config/<tool>) first - they often answer status questions faster than any transcript. For huge (50MB+) files, extract only substantial assistant text blocks from the target date window instead of reading sequentially.