1.) We do have a focus point option for autocomplete, with either a soft (will sometimes show you results far away if we think they're really relevant) or a hard (only stuff nearby) radius that can be set. Look at the boundary.circle and focus.point parameters in our /v1/search endpoint docs (https://github.com/pelias/pelias-doc/blob/master/search.md). As you probably realized tuning this for general use is actually quite difficult, but if you're running pelias yourself there's nothing stopping you from tuning to your particular use case. But if you're running your own instance you can also just add a filter by state, since we already have a state field in all our documents.
2.) That's right, even for our datasets that have polygon data (only OSM and Quattroshapes), we just take the center point. This is something we want to add in the future, but it may take a few months at least, since we don't currently have much infrastructure for handling polygon data. We also have another awesome team at Mapzen that is working on a gazetteer project called Who's On First(http://whosonfirst.mapzen.com) that should really help by offering concordance between datasets.
3.) This one's interesting, we should be able to fix it right away because NYC data is generally quite good. Can you check if your building is in OpenStreetMap(http://openstreetmap.org) or OpenAddresses (http://openaddresses.io has a download for the USA, then find the file for NYC)?