Research a question across three independent sites

[devbadya]

A 0.8B model will not search, pick a result, and read three pages inside four tool rounds. In practice it searches four times and never reads. research does the fan-out in code, the way weather already does.

This is the cut that was missing from #38: three independent sites, not five (#32) and not every web_search (#34). Wikipedia pages skip the reader, so a typical call spends one reader request on the search and two on the other sites.

What one call does

1. Searches (German questions still hit German Wikipedia / DuckDuckGo kl=de-de from Make web research follow the question's language #38).
2. Reorders so each site appears once — investor.nvidia.com and nvidianews.nvidia.com count as one, bbc.co.uk and theguardian.co.uk as two. Wikipedia, where every hit is one site, refills with further articles instead of collapsing to one.
3. Reads three in parallel. A 429 or a firewall page becomes that result's search snippet rather than taking the other two down with it.
4. Returns quoted passages, scored lexically against the question, plus today's date. The digest is capped at 4,000 characters.

research-question offers only research. lookup-term (Was ist Stripe?) still uses web_search / read_page — a name does not need three sources.

Verification

pnpm check — 1,050 tests — after rebase onto current main. 40+ of those cover site diversity, boilerplate filters, passage selection, snippet fallback, and the whole call over a stubbed network.

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