Research leftover questions, and open pages that actually answer

[devbadya]

A 0.8B model will invent a capital, a boiling point, or a current office-holder rather than admit it does not know. Research already existed for the shapes the skill author wrote; everything else was left unaided. Greetings were never the problem — leftover questions were.

Merged with main after #47 landed. Wikipedia-alongside-web, retries, inflections, and the Amtsträger incumbent-line come from there. This branch adds when to research, query focusing, and Wikipedia short-description ranking.

When it researches

Routing gains a fourth stage, after triggers, the keyword index, and carry-over:

• A leftover factual question (What is the capital of France?, Warum ist der Himmel blau?, What is 32 fahrenheit in celsius) now loads research-question.
• Greetings and small talk do not: wie geht's dir, wie gehts dir, how are you, Was machst du heute?, Erzähl mir einen Witz.
• Skills that already matched still win. Weather, arithmetic, lookup-term, memory and the clocks are unchanged.
• A follow-up like and in Lisbon? still carries the resident skill rather than becoming a search for Lisbon.

This is isResearchable in code, not a broader trigger. A regex on the skill itself cannot hold the exclusions without stealing greetings the way who is once stole Who is that?.

What one research call does now

1. Narrows the query the way weather narrows a place: What is the capital of France? is searched as capital of France. Searching the whole sentence ranked a page about questions (and WhatsApp) ahead of Paris. Why and how keep the shell, because sky blue ranks a colour swatch and why is the sky blue ranks the physics. Language is taken from the original question, so a stripped German who still hits German Wikipedia.
2. Searches the live web and Wikipedia together (MediaWiki is free of the reader budget; skipped when Wikipedia is already the provider).
3. Puts the Wikipedia hit first when one appeared. MediaWiki's own ranking still puts a list and capital punishment ahead of Paris; the article whose short description is the query (Paris: "Capital of France") takes that free slot, and a Wikimedia list is pushed behind it. Then one hit per remaining site, ranked by whether the snippet already bears on the question — including an exact-phrase bonus, so capital of France beats capital punishment in France. Login walls and download_pdf galleries go last.
4. Reads three in parallel. A 429, a Sucuri/Cloudflare interstitial, or a page with no prose is replaced from the remaining hits (up to five reads) rather than quoted. The search snippet only stands in when nothing else could be opened.
5. Scores passages with question-shell words dropped and inflected forms counting as the same term, so Bundeskanzlers answers Bundeskanzler. A one-line Wikipedia claim that names the incumbent is kept even when the lead is a definition of the office.

lookup-term is unchanged — a name still does not need three sources.

Eval

• What is the capital of France? now expects research.
• Wie geht's dir? and How are you? expect no tool.
• Who wrote Dune? and What's the population of Tokyo? already routed to research-question; the scenarios now expect research rather than web_search.

Verification

pnpm check after merging main — 1,284 tests. New cases cover the question stage, the greeting exclusions, query focusing, Wikipedia short-description ranking, snippet ranking, login demotion, German inflection, Wikipedia merge, firewall replacement, and the Amtsträger incumbent-line.

Live Wikipedia (CORS * from the Pages origin, extra description/pageterms props): searching capital of France and reranking by short description puts Paris first. Searching the unfocused full question still ranks Closed-ended question / WhatsApp — which is why focusing is not optional.

The model itself is not exercised here: this environment has no GPU adapter. Whether the skill actually reaches research still needs /?eval in Chrome on a real GPU.