iterative-polish-autolooplisted

# Iterative polish autoloop A single audit produces a verdict; a verdict is not a polished flow. This skill closes the loop: each iteration produces a measurable grade delta, one fix at a time, with backend-truth verification at every step. It exists because, without the loop, polish stalls at "B+ feels good enough", per-iteration fixes regress other screens, fixes batch into un-attributable PRs, composition-level bugs survive macro grading, and the UI ships claims the backend contradicts. **Purpose:** drive a flow from raw to award-quality through tight iterative loops. Each iteration: drive end-to-end on a clean fixture → capture every state → grade ruthlessly → pick ONE highest-ROI fix → apply + commit atomically → schedule next. ## Derive this project's inputs first Read the project at runtime to resolve these — do not assume: - **Fixture reset command** — the seed/cleanup that returns a known clean state. - **Flow capture harness** — the auto-driver (Maestro YAML chain / Playwright spec / a shell script) that runs the flow without user input. - **Reviewer agent** — the L1 grader for multi-screen flows (or the single-screen UX reviewer when drilling down). - **Backend-truth probe queries** — 3–5 project-specific SQL/API calls, each with a known-good shape; deviation = a bug. - **Semantic-count audit patterns** — where a rendered count can lie for this project (pipeline-subset vs user truth on generative surfaces; filter-scoped vs unfiltered counts on dashboards). - **