Long Task Continuation

It's checkpoint-driven: you write intent files, baseline usage snapshots, todo lists, and drift checks to disk under a dated work directory. Before each slice you restate the goal and what you're changing. After each slice you log evidence and update state. The discipline is overkill for quick fixes but essential when subagents are involved, architecture is shifting, or you need to resume three sessions later without guessing where you were. It's not a planner or executor, just the scaffolding that makes long runs legible and recoverable.