relay.damus.io is the least predictable relay we tested. It enforces a per-IP read budget: after a few quick requests its pages shrink to a handful of events, and a walk that runs right after another on the same connection can come back short (91 events, then 38, for the same filter). Paced, it still completes; a 5,000-event walk finished in 19 requests. Sometimes it says so (CLOSED "read bandwidth budget exhausted … (retry in 30ms)", or HTTP 503 at connect), but the shrinking pages and its tag searches give no signal. A #p search with no kinds returned only the newest 27 events over 60 days. Over an older slice, the same search with kinds 1 and 7 returned 253, and with limit: 500 (its own published max_limit) it returned nothing. A walk can’t tell that apart from a genuinely small result. Its NIP-11 publishes max_limit 500 and max_subscriptions 200, but not the read budget or the tag-scan behaviour, so those are limits we can’t plan around.