compaction-survivallisted

# Compaction Survival A long autonomous run loses most often to this — the context that held the plan gets compacted or reset, and the next turn resumes from a summary that dropped the load-bearing detail. The defense is an anchor on disk that the run re-reads and rewrites as it goes, so the plan lives in a file, not only in the context window. This is a **flexible** skill: the anchor's schema and update cadence adapt to the task. What stays firm is small — the anchor is the single source of truth for run state, re-read at the start of each turn and updated before the state it describes can be lost. ## The anchor One file, at a stable path the run can find again after a reset. It has two tiers, split by a literal `<!-- anchor:tail -->` marker line: above it the live **HEAD** — the only part the re-injection hook emits — and below it the **TAIL**, which stays on disk. A marker-less anchor still injects whole, but then a long run's live state is whatever the 8K bound keeps. HEAD — bounded, rewritten in place: - **Mission** — the goal in a sentence or two, the hard constraints, and any user instruction that constrains *mechanism* rather than outcome, quoted in the user's own words with a stable id. Paraphrase is where an order dies: once the wording is gone a substituted mechanism reads as a design choice, and a conformance check has nothing to point at. - **Plan pointer** — where the full plan lives (a separate doc), so the anchor stays a cursor, not a second co