feedbacklisted

# feedback — learn from corrections, honestly Follow the shared contract in `../README.md`. "Learning" here means enriching the local knowledge base and example store (README's honest positioning) — nothing else. Promotion is **always human-validated**. ## Prerequisite A generated test book to compare against (`.qaia/testbooks/<US-ID>/`, from `testbook-generate`). If none exists — no checkpoint, no test book — say so and offer `us-ingest` to start the journey instead of asking for corrections that have nothing to be diffed against. ## Steps 1. **Collect.** Ask what the user changed or rejected in the test book (or diff the edited `.feature` files against the generated version if both exist). For each correction, capture: scenario ID, what was wrong, the corrected form, and **why** — the why is the valuable part. 2. **Classify.** Each correction is one of: - `business-rule` — the generation contradicted or missed a domain rule → candidate for `knowledge/`; - `style` — wording, structure, granularity preference → candidate for a project convention entry; - `one-off` — specific to this US, not generalizable → example only. 3. **Store examples.** Write each correction to `.qaia/feedback/examples/<US-ID>-<n>.md` with its classification and provenance. 4. **Propose promotions**: only when the same pattern appears in **≥ 2** stored examples, or the user explicitly asks for immediate promotion (single-criterion — the "states a reusable rule" shortcut promoted everything