I transcode everything to unicode internally, then retranscode to utf-8 on output. (At first I thought I didn't need to touch any encoding which didn't support Han characters, but & escapes may need to be transcoded.) Figuring out the original encoding can be painful:

• Sometimes in http content-type header text/html; charset=
• Sometimes in meta tag, e.g.

  <META http-equiv="Content-Type" content="text/html; charset=EUC-JP">.

  Should appear as early as possible in HEAD element. (I currently do a regex search for this pattern before transcoding.)
• The html spec says that HTTP headers should take precedence over META tags, but in my experience, when they're both present, HTTP is wrong. (It's much easier for people to edit their html than to understand how to use .htaccess, or reconfigure their web server.)
• But: people will type anything into their META tags, just like any other html. I've seen content-type "text-html", a UTF-8 BOM prefixed to an html file with a meta GB2312 declaration, and UTF-8 data with a META UTF-16 declaration.
• Similar problems occur with rss/atom feeds and xml encoding declarations vs. http charset declarations: see http://www.intertwingly.net/blog/2004/02/13/HTML-charset-vs-XML-encoding on Sam Ruby's blog, and links therein.