Townwick Q4: the intake round, and the questions the sections were waiting for

What shipped

The research corpus closed with a line that was really a brief: “A pipeline that adds section types without adding these intake questions will produce empty new sections.” Q4 is that gap closed.

The questions ship as per-vertical interview kits, scaffolded with pnpm interview <slug>. There are 27 of them across a core kit and three vertical kits, and every single one carries a citation back to a named research finding plus a marker showing which config path it feeds. Nothing is in there because it seemed like a good question to ask.

The extraction path is a strict prompt template. It renders in the Prompt Lab or by hand, the operator runs it in their own model session, and the output gets pasted into the Console’s validating editor. Then docs/07, the runbook that ties all of it to the consent rules: blank beats guessed, quotes are verbatim or absent, and [off] never reaches the page in any form.

The worked example is the fictional showcase barbershop interviewed inside its own fiction. A filled questionnaire produced by the real scaffolder, and a hand-authored example output that the gate holds to the extraction contract. Same discipline as every other showcase fixture, except now the grounded-extraction property is machine-checked rows instead of a habit.

Decisions

Zero-key was the call that shaped everything else. The extraction runs in the operator’s own session, so this repo makes no LLM call and holds no API key. The consuming machinery already existed, since the Console’s C3 editor validates a pasted config.yaml against the real schema before anything gets committed, so the only genuinely missing pieces were the questions and the prompt. Adding a key would have bought a billing surface and broken the C4 gate’s no-network proof, all to automate a step a human should be reading anyway. The packages/ai stubs are untouched and still wired for the day that changes.

The scaffolder refusing to overwrite is a small thing that matters. Once clients/<slug>/interview.md holds a conversation with a real person in their own words, “regenerate” means “throw their answers away”, so pnpm interview exits 2 instead. There’s no --force. A flag that deletes a conversation is not a flag I want in the tool.

And the grounded-extraction property is a gate row rather than a hope. The prompt says never invent, but rules without a gate are comments. The check borrows pnpm verify’s own properNounRuns and normalizeForMatch, so “proper noun” means the same thing at intake as it does at go-live. Five negative rows each fire on a mutated temp copy, and each mutation is proven to have actually mutated by byte comparison first.

What broke

The full-battery run surfaced something worse than a bug in this milestone: main’s CI has been red since the CJ flagship rebuild (07bcadb and 4515df3).

Five older gates, phase-1, phase-3, sections, motion and trust, hardcoded the seed client’s original state: stock photos, classic variant, no sections block, reveal off. The rebuild moved all of it. Triage says none of them is a product regression, they’re all gates pinned to a fixture that legitimately changed.

The sharpest case is a good demonstration of how brittle a string-matching gate can be. The motion gate has a row asserting the .member press-state is still present, and it trips on the string .member.is-pressed sitting inside a CSS comment in the reveal stylesheet, which reveal: true newly pulled into the seed build. So the gate is passing on a comment.

Reconciliation is its own follow-up round: gates re-pinned to temp-copy fixtures, negative proofs re-run. Q4’s own gates are green and touch none of the five, which is the only reason this shipped rather than stopping.

Numbers

Intake gate is 81/81. Console-c4 went to 114/114 once its arrays gained the fourth template. Tests are 362 plus 38 plus 71, typecheck 0, validate 6/6, and pnpm verify on cjs and fern-and-fade returns byte-identical verdicts to pre-Q4.

The negative proofs are the part worth reading. A name planted in the real example fixture failed the full gate at 77/81 with the plant named, and the collision between that plant and negative row (a)’s own target got caught by the mutation-must-mutate guard, which is the guard doing exactly what it’s for. Deleting intake-extract.md failed the edited c4 gate on all five touched rows. Both restored byte-exact, sha-verified.

Next

The output-quality programme is done. Next up is the CI reconciliation round to un-red main, then phase 4, billing, which is the last numbered phase in docs/06.