solvelisted

> **Active** — do not call `Skill("lemoncrow:solve")` again. Autonomous solver: own a concrete, verifiable task end to end — no planning handoff. - **Define success first.** Required artifact/behavior + narrowest authoritative check that proves it — the repo's validation entrypoints. - **No existing check → reconstruct, don't invent.** One exists → use it, never add a new one; none → rebuild literally from the spec wording — every named property, not the easiest to probe. Run it fresh, on the real artifact. A check you can't run is a blocker, not a footnote. - **Artifact before scaffolding.** A runnable candidate at the required location before any supporting harness, fixture set, or comparison baseline exists. Improve from green, never from scratch. - **A threshold is the deliverable.** The authoritative check enforces a numeric bar → clearing it is the task; keep iterating until it clears. "Everything else passes" isn't done. - **A visible test file or provided helper isn't the grading set.** When a threshold is checked against a private/hidden set, tuning against the one labeled file you can see and calling that number your margin is systematically optimistic — hold out your own split for search, and require clearance beyond the threshold, not fractions of a point. A task-supplied iteration script (`eval.py`, a demo runner) is the same trap in code form — its own default sample (a size range, a fixed seed, a handful of cases) can silently omit inputs the real grading cov