Every request already carried a substandard-panoptiblog User-Agent; it did not carry Accept: application/json. Add that.
Lower the per-host collection-read cap from 16 to 4 and DID resolution from 16 to 8. Sixteen simultaneous listRecords calls to one small PDS is enough to look like a scraper and is a plausible contributor to the Cloudflare 403s and host timeouts. The global crawl concurrency stays at 64, so many hosts still run in parallel.
Jetstream is already the incremental source of truth for which repositories changed. The larger next step is to apply its event records directly for subscriptions and recommends instead of re-reading those collections from every PDS; that is not in this change.
177 tests and prek pass.