agentrecap threads.csv: classify tool calls via event_kind and tool_input

agentrecap (npm, v0.13) has no content filter; only --with-transcript adds message/reasoning/tool text to threads.csv, and --dir/--since/--until filter by scope only. To analyze usage by purpose, keep rows with event_kind=tool_call and parse tool_input for file_path (Edit/Write) or command (Bash). tool_input is often truncated JSON, so json.loads fails on ~40% of Edit/Write rows; extract fields with a regex instead. Token usage lives on rows with usage_canonical=True and can only be attributed per thread, not per tool call, and is dominated by cache reads. user_prompt rows often start with harness-injected context (<ide_opened_file>, skill bodies), so skip those when looking for the real first ask.