plan-fixlisted

# Plan fix Read the paper source in full, treat the bundle's annotations as a single editorial brief (one diff may satisfy several marks), and emit a JSON plan describing the minimum coherent set of edits. The desktop renders a sibling Markdown projection from the JSON for the user to read; **do not emit Markdown yourself** — the structured JSON is the contract. Do not write to any source file in this skill. ## Pacing rule — emit phase markers BEFORE deep reasoning Each `[obelus:phase] <name>` marker (listed below) must be emitted **on the assistant's first text output of that phase, before any large thinking block or any tool call**. Do not pre-think the entire phase before emitting the marker. Tool calls are cheap; thinking blocks are not — and the desktop's stopwatch and jobs dock both depend on the marker being the first thing the model produces when a phase starts. A 30k-character thinking block before the first `[obelus:phase]` of a phase is the single most expensive failure mode of this skill. Equivalent rule for tool-heavy phases: the first action in **Locating the source span** is a `Read`, not a thinking burst. The first action in **Stress-test** is the `Task` (subagent) call, not a thinking burst. ## Reference modules — `Read` is forbidden until the owning phase begins The detailed sections for the optional sweeps, HTML rules, and worked examples have been moved to standalone files under `refs/` (sitting next to this `SKILL.md` at `<plugin>/skills/plan-fix/re