I've no got to the point where my XQuery Update implementation is passing almost all the published tests. That doesn't mean its complete, because the test suite doesn't cover certain areas, such as revalidation of updated documents. But it's looking good.
There were quite a few hassles getting corner cases to parse correctly. The XQuery grammar really is very fragile, and it's hard to make changes to a parser without breaking something. The particularly thorny point was the overloading of the "as" keyword in the insert and rename expressions, which has required changes to the Tokenizer that have side-effects on almost all other constructs using "as".
The other problem has been making the linked tree fully updateable. There were all kinds of hidden assumptions here about node identity and document order that have had to be revisited. Some of them still need to be addressed, for example I don't yet have an efficient way of representing the base URI of nodes. Currently a lot of properties (like base URI and in-scope-namespaces) are computed on demand rather than being physically stored in each node, and this makes for extra work when updating. There are also document-level indexes that need to be updated or invalidated. I was lucky that the test suite caught one or two of these, because it's only doing tiny queries on updated documents so a lot of errors could remain uncaught.
There's a nasty messiness in the spec concerning the rules for mixing updating and non-updating statements. For example, if one branch of a conditional is an updating expression, then the other must either be an updating expression, or (), or a call on error(). The problem with this is that you can only tell whether an expression is updating after you have bound function calls to the corresponding function declarations, and by that time Saxon has already done the first stages of simplification of the expression tree: for example it replaces a typeswitch expression by a corresponding structure of nested conditionals on-the-fly while parsing. If I can't get the rules changed in this area then it's going to be quite difficult to implement them precisely: it would mean, for example, that the expression tree has to retain information about otherwise redundant parentheses in the source query. Really this part of the language is badly designed; the rule on mixing updating and non-updating expressions should be phrased as a semantic rule not as a syntactic one. I'd be suprised if other implementors don't have similar problems, although some of them may be less agressive in the rewrites they attempt than Saxon is.