pvt-preplisted

# PVT Prep A PVT runs in a short window on the production database, and every statement there is run by someone else. The deliverable is therefore not a test plan — it is a **script set another person can execute in order without asking you anything**. Two constraints drive every decision: 1. **Window time is the scarce resource.** Data that could have been staged before the window must not be created inside it. 2. **Every SRE round trip costs the window.** Batch what SRE runs; hand them the fewest scripts that still keep the phases separable. ## Procedure 1. **Intake the cases.** Read the source the user names — CSV, markdown table, or a Jira story. Do not restate cases you cannot find; list them as gaps. 2. **Ground them.** `describe_tables` and `inspect_relationships` on every table a case touches. A case whose column does not exist is a finding, not a script. 3. **Check the blast radius.** `simulate_impact` on each table or column the scripts will write. PRD writes that reach an endpoint or job nobody expected are the ones that hurt. 4. **Group.** Collapse the cases into the smallest set of scenarios that can share one data setup (see below). 5. **Sequence.** Assign each scenario to a script in the numbered set, with an owner suffix. 6. **Verify.** Prove every `SELECT` you hand over with `run_sql`. Never run a write from this skill — the scripts are artifacts, not actions. ## Test case intake Accept a CSV or table with at least: case i