Keep researched names inside the country the question named

[devbadya]

Problem

Two live failures on wer ist der russische präsident / nein in russland:

1. Fake incumbent. research searched russische präsident and still opened with a German Amtsträger line (Friedrich Merz, Bundespräsident / Bundeskanzler). extractAnswer treated any Amtsträger ist … Name sentence — and any … ist … präsident sentence — as the answer, without requiring Russia in that source. wrong-fact then forced the model to repeat Merz.
2. Guessed follow-up. nein in russland kept the research skill (short fragment) and did not call the tool. The 0.8B model answered from training data: “I believe … Vladimir Putin … since October 2009”. It cannot know this without researching again.

Fix

• A nationality or country in the question must appear in the same source (title, URL, or passage) before a name is extracted. Amtsträger ist Friedrich Merz still answers Bundeskanzler; it does not answer russische Präsident. related() treats russische / Russland / Russia as the same place so Wikipedia ranking prefers Präsident Russlands.
• The research skill now has a German wer ist der russische Präsident? exemplar (so that shape is not always Merz) and a nein in Russland exemplar that calls research again with Präsident Russland.
• conversationTopic pins the last office onto that correction, the same way und der von Frankreich? keeps Bundeskanzler.

Tests

• extractAnswer rejects Merz for a Russia question, accepts Putin from Präsident Russlands, and prefers Putin when both pages are present.
• pickCandidates gives the Wikipedia slot to Präsident Russlands, not Bundespräsident.
• Topic + routing + eval runner cover nein in russland.
• New eval scenario: web-follow-up-correction.

pnpm check is green (1,346 tests). This does not claim the 0.8B model behaves correctly on a GPU; that still needs /?eval in Chrome or Edge.

Out of scope

No longer system prompt, no auto-learning from every question, no second model call per turn.