last30days v4.0.2

Fixes from a Grok Bot run on Microsoft (#1449), measured on the same run before and after:

Prediction-market odds are real again. An outcome with an empty order book no longer reports a meaningless midpoint, and a set of mutually exclusive ranges that still adds up to far over 100% is hidden. Microsoft's weekly close went from 46% / 44% / 40% to 14% / 8.5% / 8.0%.
Hiring Signals shows different roles under each theme. Themes read the job title, department, and the role's own description, not the company text every posting repeats. Repeated examples on Microsoft went from 3 to 0.
The top 8 stories cover more than one event. One story takes at most 3 of the 8 slots, each searched angle gets a slot, and old evidence stays behind current stories. Microsoft's top 8 went from 8 outage stories to the outage plus AI and earnings.
Top voices are the accounts people engaged with, not the ones that posted most. A headline reposter with ten zero-like posts no longer outranks the company's own account.
Source health tells the truth. A YouTube bot check gets its own advice instead of "upgrade yt-dlp", the host plan no longer adds sources that cannot run, and Techmeme searches only the research window.
Company resolution stops getting rate-limited by Wikidata (#1447).

What's Changed

fix(company): send Wikimedia a contact User-Agent so Wikidata stops rate-limiting by @mvanhorn in #1447
fix(polymarket): price outcomes from the book and drop impossible negRisk sets by @mvanhorn in #1450
fix(render): rank top voices by engagement, not post count by @mvanhorn in #1452
fix: bot-check aware YouTube health, runnable-only plan backfill, dated Techmeme search by @mvanhorn in #1453
fix(hiring): theme on role text, not shared boilerplate; distinct examples by @mvanhorn in #1454
fix(clusters): story-diverse order of top clusters by @mvanhorn in #1451
chore(release): bump version to 4.0.2 by @github-actions[bot] in #1458

Full Changelog: v4.0.1...v4.0.2