If the protocol breaks (is interrupted) after the server has signed and returned the refund transaction (T2) to the client, the server has to blacklist its input (the hash of T1) indefinitely, or am I overlooking something? Otherwise the client can replay T1 at a later time, after the locktime of T2 has passed, and refund to himself before the protocol is terminated. At least the server has to keep track of the timelock used in T2. What does the implementation do?
[..]
Now you feel betrayed by the merchant and publish the very first full reimbursement but the merchant has time until the timeout of A expires to publish any lower reimbursement that you signed and that overrules all prior Bs.
[..]
I think you were describing why the protocol is safe for the client (buyer), while my concern was whether it is safe for the server (merchant). Anyway, Mike made this clear to me now.