Invalidations apply asynchronously, so the entry an event carries can be
a stale snapshot by the time it reaches the handle: a local flush may
install newer state while the event sits in the queue, and the flush's
own event is dedup-suppressed, so rolling the handle back would never
heal. The apply loop has already ordered the event and any later state
into the local store, so resolve the refresh there first; fall back to
the event entry only for read-through directories, where the store
holds nothing and the event is the freshest ordered information
available.
Chris Lu
3696790a9d