Finally, are you sure it's the open/close overhead that would kill the naive approach? I'm with you on this, but the point is that neither of us can tell without measuring.

More or less. The reason I said "positive" is because a sizable portion of the site is a photo archive under one directory, which contains a whole slew of pages that present thumbnails -- they have little or no text in them. Of course, there is a little more involved than just "opening and closing" them, they must also be parsed to eliminate the html tags. Which that is an inescapable part of the deal. But if I exclude that directory -- which contains a negligible portion of the data -- the search is very very noticeably faster.

I also know from other directory tree stuff that even WITHOUT this parsing, a few hundred or thousand files spread across 10+ gigs is a LOT just to stat the files. Try "du / >tmp.txt" on your hard drive. It will take several minutes at least, and that is a C program just collecting file sizes. A site-search engine is not much good if it takes more than 5-10 seconds, methinks.