Trying to replace HTML by cloning nodes but getting strange results

Question

There were a couple of issues with your code (or maybe a little more).

Chrome detects word-boundaries in its own way, so \b does not work as expected (e.g. a . is considered part of a word).
You were using the global modifier which returned the indexes of all the matches it found. But when handling each match, you modified the content of child.data, so the indices that referred to the original child.data were rendered useless. This problem would only come up whenever there were more than 1 matches in a single TextNode. (Note that once this error caused an exception to be raised, execution was aborted, so no further TextNodes were processed.)
The acronyms were searched for (and replaced) in the order of appearance in the acronym list. This could lead to cases, where only a substring of an acronym would be recognised as another acronym and incorrectly replaced. E.g. if ERA was seached for before ERA+, all ERA+ occurrences in the DOM would be replaced by <abbr ...>ERA</abbr>+ and would not be recognised as ERA+ occurrences later on.
Similarly to the above problem, a substring of an already processed acronym, could be subsequently recognised as another acronym and pertially replaced. E.g. if ERA+ was searched for before ERA the following would happen:
ERA+
-> <abbr (title_for_ERA+)>ERA+</abbr>
-> <abbr (title_for_ERA+)><abbr (title_for_ERA)>ERA</abbr>+</abbr>
Your one-letter "acronyms" would also match characters they shouldn't (e.g. E in E-mail, G in Paul G. etc).

(Among many possible ways) I chose to address the above problems like this:

For (1):
Instead of using \b...\b I used (^|[^A-Za-z0-9_])(...)([^A-Za-z0-9_]|$).
This will look for one character that is not a word character before and after our acronym under search (or settle for string start (^) or end ($) respectively). Since the matched characters (if any) before and after the actual acronym match need to be put back in the regular TextNodes, 3 backreferences are created and handled appropriately in the replace callback (see code below).

For (2):
I removed the global modifier and matched one occurrence at a time.
This also required a slight modification, so that the new TextNode, created with the part of child.data after the current match, is subsequently searched as well.

For (3):
Before starting the search and replace operations I ordered the array of acronyms by decreasing length, so longer acronyms were search for (and replaced) before sorter acronyms (which could possible be a substring of the former). E.g. ERA+ is always replaced before ERA, IP/GS is always replaced before IP etc.
(Note that this solves problem (3), but we still have to deal with (4).)

For (4):
Every time I create a new <abbr> node I add a class to it. Later on, when I encounter an element with that special class, I skip it (as I don't want any replacements to happen in a substring of an already matched acronym).

For (5):
Well, I am good, but I am not Jon Skeet :)
There is not much you can do about it, unless you want to bring on some AI, but I suppose it is not much of a problem either (i.e. you can live with it).

_{(As already mentioned the above solutions are neither the only ones available and probably nor optimal.)}

That said, here is my version of the code (with a few more miror (for the most part stylistic) changes):

var matchText = function (node, regex, callback, excludeElements) {
    excludeElements
            || (excludeElements = ['script', 'style', 'iframe', 'canvas']);
    var child = node.firstChild;
    if (!child) {
        return;
    }

    do {
        switch (child.nodeType) {
            case 1:
                if ((child.className === 'sabrabbr') ||
                        (excludeElements.indexOf(
                                child.tagName.toLowerCase()) > -1)) {
                    continue;
                }
                matchText(child, regex, callback, excludeElements);
                break;
            case 3:
                child.data.replace(regex, function (fullMatch, g1, g2, g3, idx,
                                                    original) {
                    var offset = idx + g1.length;
                    newTextNode = child.splitText(offset);
                    newTextNode.data = newTextNode.data.substr(g2.length);
                    callback.apply(window, [child, g2]);
                    child = child.nextSibling;
                });
                break;
        }
    } while (child = child.nextSibling);
    return node;
}

var abbrList = Object.keys(acronyms).sort(function(a, b) {
    return b.length - a.length;
});
for (var i = 0; i < abbrList.length; i++) {
    var abbrev = abbrList[i];
    abbrevSearch = abbrev.replace('%', '\\%').replace('+', '\\+').replace('/', '\\/');
    console.log("Looking for " + abbrev);
    var regex = new RegExp("(^|[^A-Za-z0-9_])(" + abbrevSearch
                           + ")([^A-Za-z0-9_]|$)", "");
    matchText(document.body, regex, function (node, match) {
        var span = document.createElement("abbr");
        span.className = "sabrabbr";
        span.title = acronyms[abbrev].replace('&#39;', '\'');
        span.textContent = match;
        node.parentNode.insertBefore(span, node.nextSibling);
    });
}

For the noble few that made it this far, there is, also, this short demo.