This looks like a good optimization, it obviously needs more testing
because Linux always seems to find new broken hardware.  The areas
that need to be tested should include:
 * MacOs has a broken version of delayed ack that might cause
   HyStart to radically underestimate.

We tested with FreeBSD. I presume that it covers MacOS. We will look into 
that.

 * Applications that dribble out packets might get better (or
   worse) performance.  This include dumb web servers.

Not really. It does not affect those flows that do not use their full 
capacity of the network. HyStart kicks in only when the flow are using more 
than avail bw during slow start. If the applications do not pump data in, it 
would still be in slow start -- while in slow start, HyStart works like the 
traditional slow start (double cwnd per rtt).

 * Does this increase or reduce latency when using TCP for
   applications which never fill the congestion window? (games, financial, 
etc).

For the same reasons as above, No, HyStart does not engage in this 
situation.