perf-investigatelisted

Performance work is an extend op-cell: add measured capability without guessing, stacking changes, or laundering anecdotes into evidence. The invariant is: every hypothesis cites source evidence, every optimization has a baseline delta, and every decision is recorded in `.outline/perf/`. ## When to Apply / NOT Apply: performance investigation; unexplained latency; throughput or memory regression; profiling a named scenario; creating a versioned baseline; deciding whether an optimization is worth keeping. NOT: wrong-output bugs; speculative micro-optimizations; no reproducible command; no success metric; user only wants a quick code review; workloads too short to measure without variance control. ## Workflow Use artifacts only under `.outline/perf/`: - Ledger: `.outline/perf/investigations/<id>.md` - Baselines: `.outline/perf/baselines/<version>.json` - Profiles: `.outline/perf/profiles/<id>/` - Experiments: `.outline/perf/experiments/<id>/` ### 1. Setup — scenario, command, version, quote Require all three before measuring: - scenario: the exact workload or user journey - command: executable benchmark or repro command - version: baseline label, commit label, release label, or user-provided slug Record the user's problem statement verbatim. Do not paraphrase it. Create the ledger immediately with the phase order and empty evidence slots. ```sh mkdir -p .outline/perf/investigations .outline/perf/baselines .outline/perf/profiles .outline/perf/experiments ``` Ledger i