US-60 — A repeatable score for ai/query.py, run in CI #393

As a maintainer, I want the question-answering layer scored the same way every time, so a change that makes it answer wrongly fails CI instead of reaching a user.

Acceptance criteria

• A held-out question set for ai/query.py, scored with the same method as A held-out, repeatable score for the explanation layer (#346) #385
• The score runs in CI, which has neither Batfish nor Ollama
• "Answered and wrong" is counted separately from "refused", as docs/evaluation.md Part 2 does

Shape

Not a pipeline feature: AI layer and tooling.

Why this is in the milestone

The question layer has been measured once, by hand: 20 questions on 10 August (docs/evaluation.md, Part 2). It has never been pinned in CI.

Depends on #385: its method is reused, so it merges first.

Epic: Specialist pick (explainable and trustworthy AI for network security) | Size: estimated at sprint planning | Assigned: @patelankeet2

Specialist pick agreed by all four on 13 October 2026 (slate row 3). Acceptance criteria are drafted from the slate's wording; the owner refines them at sprint planning on 19 October.

Part of Phase 3 - market readiness.