umerkay adds Olostep web research endpoints behind optional dependency

The implementation spans six files: backend/src/routes/research.ts (the router), backend/src/lib/olostep.ts (the SDK wrapper), and plumbing changes to package.json, package-lock.json, index.ts, and .env.example.

The olostep package lands in optionalDependencies and the lib file loads it with a one-shot import("olostep") that caches the result - null if the module isn't installed, the module if it is. Missing-key errors and missing-package errors both surface as "Error: ..." strings rather than thrown exceptions. That pattern costs almost nothing to replicate and means the feature can ship to a codebase before everyone has an Olostep account.

The four endpoints map to Olostep's four API surfaces: scrape (Markdown, images stripped), search (AI-powered answers with source URLs), crawl (async multi-page, returns only a crawl ID), and map (URL discovery with optional query). All four require auth and validate the URL scheme.

A few gaps to weigh before pulling. The /api/research mount prefix breaks the pattern of every other router in the file. The crawl endpoint returns an ID but there is no status or result endpoint yet. No rate limiting and no audit logging on outbound paid API calls. The Olostep Format enum and client.scrapes/client.answers/client.crawls/client.maps shapes are used directly, with no provider interface in between.