Make web research follow the question's language

[devbadya]

Web research was answering German questions from English Wikipedia, dating "today's news" with nothing, and spending two of the reader's 20 requests a minute to open a Wikipedia page MediaWiki already serves with CORS.

This is a different path from #32 and #34, which fold five page-reads into one call and spend most of the keyless budget on a single question. Here the existing web_search / read_page pair stays, and the work that does not need the reader moves off it.

What one search/read now does

• German questions search German sources. Wikipedia uses de.wikipedia.org when the query is German (Wer ist…, Was kostet…, umlauts) and falls back to English if that edition has no article. DuckDuckGo gets kl=de-de. English who was Ada Lovelace is not mistaken for German, because bare was is also English.
• Wikipedia pages skip the reader. read_page of a *.wikipedia.org URL goes through MediaWiki (origin=*, verified from https://devbadya.github.io). The extract is truncated here to 8,000 characters. Wikipedia's own exchars parameter is capped at 1,200, which cuts off before the sentence that names a current office holder — on the live Bundeskanzler article, Friedrich Merz sits at character 1,539.
• Every search is dated. Results start Searched 2026-08-26 for "…", so "current" and "today's news" have a date without spending a tool round on current_time.
• The research skill teaches the missing step. The old exemplar only answered from a snippet. Wikipedia's lead about an office often does not name the person, which is the UN secretary-general eval. A second exemplar now opens that page; a third answers Wer ist der Bundeskanzler? in German from a German source, without translating the query.

Cost

A DuckDuckGo search still spends one reader request. Opening the Wikipedia hit spends none. The five-source research tools spend six.

Verification

pnpm check (936 tests) and pnpm build pass. New tests cover language detection, the German Wikipedia host, the English fallback, DuckDuckGo kl=de-de, MediaWiki read_page, and the date stamp.

Live against Wikipedia with Origin: https://devbadya.github.io: search for Bundeskanzler returns the German articles with Access-Control-Allow-Origin: *; reading Bundeskanzler (Deutschland) and keeping 8,000 characters includes Friedrich Merz; Who was Ada Lovelace still hits English Wikipedia.

The model itself is not exercised: this environment has no GPU adapter, and there is no CPU fallback. Routing and tool-use rates still need /?eval in Chrome on a real GPU.