Two client-side costs, both measured against live relays rather than
argued about.
The dial pacer spaced every attempt to one relay a quarter second
apart. A cold start opens tens of sockets, so the later ones waited
seconds before their first byte - the client log shows it plainly,
tcp_ms climbing past eight seconds while the network round trip is
seventeen milliseconds. The premise for paying that was arriving as
separate clients rather than as one burst; it does not survive what
was measured since. The filtered network keys on the shape of the
hello and the name in SNI, not on arrival timing, and twenty-four
completely unpaced dials reached resPQ 24 out of 24 with the whole
batch done in 170ms. Spacing drops to 50ms - still distinct
arrivals, a ramp that costs a second instead of six. The failure
backoff, which is what actually protects a relay that swallows the
overflow, is untouched.
Nagle was never disabled. mtproto writes a short request and waits,
which is the exact case it penalises. Both transports now set
LowDelayOption in the connected handler - not in the constructor
next to the buffer sizes, because Qt applies socket options through
the socket engine and that does not exist until the connection is
up.
Guards for both, with the numbers next to the constant so the next
person to raise it argues with the measurement.
7b3bd451f5