Correlation from the live canary (last 24h, same probe):

So: strfry shows up at BOTH ends of the distribution. Software is not the discriminating variable in my data — deployment is. The relays that serve every check are curated/paid operators with rate limits that fail open; the dark ones are free-tier instances that queue or drop under load (yabu.me and nostr.land both ACK then never serve). If you want a strfry-specific note: its LMDB backend and write path are not the failure point I observe — capacity decisions are. A strfry relay with modest limits and a persistent disk has been the best performer in my set (nos.lol).