Force research on factual questions instead of letting the model guess

[devbadya]

A 0.8B model that skipped the research tool was still inventing facts. In the reported turn, research-question routed for who is the president of the USA, the model never called the tool, and the reply mixed Macron and Merkel from the previous chat. reviewAnswer cannot catch that: without a tool result there is nothing to check against.

This change takes the fetch out of the model's hands.

• When research-question routes, research runs in code before any generation.
• A complete new question is looked up as written. A follow-up that only names a place keeps the last office (und der von Frankreich? → Bundeskanzler Frankreich).
• The reply is the digest extract plus its source. If research has passages but no extract, a quoted passage is shown. If the lookup fails or finds nothing confident, the turn says so — it does not generate a guess.
• The model is not asked to write the fact on these turns.

pnpm check is green: 1364 tests.

This is the behaviour the 0.8B model cannot be trusted to do itself. Real generation still needs a GPU in Chrome or Edge; the tests cover the new path with a scripted research tool.