This includes two features that could be divided to separate stories. Both are intended to make the editor suggestions appear to have no latency:
Saved translations from the current editor window are immediately available as suggestions.
keep a local cache of recently saved translations
apply the suggestion search to that cache at the same time as it is sent to the server
cache results will likely be available before the server results, just display whatever is available as it arrives
merge the server and cached-based results, inserting the cache-results in an appropriate place based on their score.
should server results should win over cache results in a tie? May need some user testing or consultation to answer this
suggestions for nearby phrases are pre-fetched and cached so they can be shown as soon as the phrase is selected
optimal pre-fetch strategy should be determined by experiment
minimize cases where results are not instantly available
minimize cases where results are stale by the time the user selects the phrase
results that are old are "stale", but can still be used for initial suggestion display while the search repeats
because the TM results are not likely to change much over several minutes
most of the changes are likely from the current user in the current session, so the first point above will mean stale search results with cached recent translations are usually identical to updated search results