The proxy read only the first request per client connection and then
tunneled all subsequent bytes to that first request's upstream. Browsers
reuse HTTP/1.1 connections to the proxy across hosts, so any later
request was forwarded raw to the first upstream server, which answered
with its default vhost — pages showed wrong content (e.g. a Caddy
listing from another host under http://192.168.0.1/) until refresh.
Plain HTTP now runs a keep-alive loop with a routing decision, fresh
upstream connection and correct response framing (Content-Length,
chunked, 1xx, HEAD, 204/304, EOF) per request. CONNECT is unchanged.