retro-improvelisted

# Retro-Improve -- single-writer self-improvement Two learning loops writing the same file produce conflicting edits. One writer per artifact set. This skill is that single writer for `managed-by: sefi-agents` files. User instructions always override this skill. All factual output follows the anti-hallucination skill: cite or mark UNKNOWN, never guess. agentic-signals: goal_intake, refusal_gate, verification, loop_discipline, close_out ## HARD GUARDS You may edit only files whose frontmatter contains `managed-by: sefi-agents`. You may not create new skills without a human-approved entry in `inbox/`. If `improvement.enabled` is false in `sefi.config.yml`, output the proposed diff to `state/retro-<date>.md` and stop. `managed-by` files are installed once per user, not per project, so the single-writer invariant above holds only within one install: on a shared install, this project's metrics would rewrite agents every other project loads, and every local guard here would still pass. That is what `improvement.enabled: false` is for, and why `/sefi:init` sets it on any install serving more than one project. Never edit host-runtime memory, user config, or other plugins. ## Inputs (the scorecard) Read `state/retro-ledger.md` FIRST, before selecting any target. It is this loop's memory of its own edits, and without it the loop cannot tell that it already edited a file last week, or that a human already rejected the exact proposal it is about to make again. Three rules bind (full