annoying bug... Optimization was originally intended for the following insight:

if you have a unresolved read and a write w that is mod ordered after whatever that read sees, then any writes that are mo past w should not send their future values back to the unresolved read.  So we begin by looking for operations that happen after the unresolved read and if they are a write we have our w, and if they are a read then its reads_from is our w.
The case I missed was we don't want to consider the write that the read we want to send the value to is currently reading...  It won't cause a mo problem because it will be gone if we send a future value back to that read.