build-settle-the-numberlisted

# Build: /settle-the-number ## What the generated skill does, concretely An analyst gets the "why don't these match?" Slack message and runs: ``` /settle-the-number conversion rate: marketing weekly dashboard vs finance board pack ``` The skill traces both numbers through the semantic layer to the dbt SQL, finds the hop where they diverge, and returns something like: > **Both numbers are correct answers to different questions.** Marketing's 4.2% counts sessions with a checkout event (`stg_web_sessions`, refunds included). Finance's 3.1% counts completed paid orders net of refunds (`fct_orders`, `is_refunded = false`). The definitions diverged in PR #482, eight months ago. > **Divergence:** `models/marts/fct_orders.sql:34` vs `models/staging/stg_web_sessions.sql:19` *(quoted side by side)* > **Recommendation:** these need two names. Draft dictionary entries attached. Ten minutes, evidence included, no meeting required. Your Phase 1–3 work below is what makes this output possible on *their* stack instead of a generic guess. You are about to generate a `/settle-the-number` skill customised to this company's data stack. Complete all three phases first. ## Phase 1 — Inspect 1. **Map where definitions live.** In order of authority: - A metric dictionary or semantic-layer definitions (dbt `metrics`, Lightdash `meta` blocks, LookML measures) - dbt model SQL (the real definition, whatever the docs claim) - BI-layer custom fields and table calculations (where definiti