I've read that with extensive reading, it's best to already know something like 96-98% of the words in the text.  At this level you can infer the meaning of the words from context, and thus acquire new vocabulary without consulting dictionaries etc.

With Japanese, how does an extensive reader handle acquisition of words in kanji, when the readings may be unknown?  I have had all of the following experiences:

• I could not come up with any reading for the kanji (common for unknown 訓読み)
  
• I thought I could infer the reading of the kanji from a signal primitive, but I later found out that I was wrong.
  
• I thought I knew the reading of the kanji from another compound, but it turned out that the new word used a different reading.
  
• I thought I knew the word, but missed some subtlety of meaning or context that meant it was read differently (開く comes to mind, but this situation is fairly rare).
  

In particular I'm concerned about the middle two, since I associated an incorrect reading with the word, which could persist for a long time.  Lately I've been looking up at least the reading for every new compound I come across, but this slows down reading enormously.  How concerned should I be about associating the wrong reading with a compound, and having to correct it later?

Honestly... I wouldn't worry about your point 2/3 problems at all. On the other hand, I also wouldn't worry about lookups slowing down reading either.

Looking up every compound -does- slow down your reading, but it's also the only way to learn correct readings.

As for using guessed readings, at least for me, I don't find that a problem at all - I -know- that I'm guessing and so it doesn't reinforce in the same way as other kinds of misreadings. I'll often look up a word that I'm quite sure of the meaning from context but that I've been guessing the reading on... and having done that if I was right, everything is fine; if not, I might look it up once or twice more to correct my original misreading. Usually I'm right though at this point - it turns out that most of the 'hard' readings (irregulars, mixed on-kun, etc.) are actually common words. The bigger your base vocabulary and the more you're looking at rare words, the more accurate guessing becomes. Of course there are still -some- irregular readings, but they get fewer and farther between as you go.

If I don't know -any- reading for a character (related to your point 1), I'll always look it up regardless of context and regardless of how hard I'm trying to read 'extensively'. I simply can't read words if I don't have a sound to associate with them.

Your point-4 is -kind- of a concern, but ... not really. Sure, I've inferred that a word has some secondary meaning only to later learn that the two meanings have different pronunciations. But it doesn't really hurt you in any way to do that when you're reading or writing... the only problem that could occur is if you tried to speak the word with the wrong pronunciation before you learned the difference. Which seems pretty unlikely since most such cases involve relatively rare words. (あく・ひらく obviously being an exception as they are both quite common.)

Anyway, once I reached the point of knowing the vast majority of words on the page, I really have stopped worrying about lookups. If I want to know a word, I look it up. If I don't care, I use a guessed reading and keep going. It's really fine either way. A lot of the time my choice is honestly based on my mood - if I'm caught up in the plot I will plunge forward with guessed readings. If I'm less involved in the plot and more involved in the language and imagery and want to know every nuance then I'll look up -everything-, and maybe even beyond the dictionary if it seems worth investigating.

The truth is... once you reach this point, it doesn't really matter -how- you read as long as you read -a lot-. Of course, the less you look up the more important it is to have a lot of listening or conversation practice to get those correct pronunciations from -somewhere-, but if you haven't abandoned looking up words altogether you should be fine.

It may also depend on your type of memory. I would expect that someone with a strong visual memory could actually memorize the new words (kind of 読解-only) without even caring about how they are pronounced. Learning the readings could be taken care of separately, for instance through later Anki sessions.

I remember reading translations of Russian classical novels containing a lot of long unfamiliar character names which, even after hundred of pages, I would still have been unable to pronounce. Yet I could recognize very well whom a given sentence was about, just from the visual aspect of their names.

Edit: rephrased for clarity.

My approach to reading paper novels (which I own, not borrowed) is to mark unknown words with a pencil, and also make a mark at the top of the line which has an unknown word. So when I'm reading and I come across a word I'm not sure about, I just mark it a keep on reading. Later, I'll go through and look up the words I marked and write it down (I like to use physical notebooks). If I re-read the book, I can keep the "custom dictionary" notebook nearby to look up any words I don't remember.

It can be frustrating to see words and not know how to read them, but to a certain extent you just have to get used to it. It used to bother me a lot, but now, like I said, I just keep reading. If I'm reading a library book then I have to deal with not being able to mark and lookup those unknown words. When I think about it, I come across words I don't know in English a lot when reading literature, but I barely notice because I'm so used to reading past them.

Personally, I try to look up words that I haven't seen before or have forgotten, exactly for the reason that you state. Guessing has its limitations.