The following is just a few examples regarding "actual evidence for or against reliability in the first version of any new release", as requested. This is not meant to be a complete analysis, but rather a suggestion on what you might want to research.
You can google up "List of issues that are fixed by SQL Server 2008 Service Pack 1" and "List of issues that are fixed by SQL Server 2008 Service Pack 3" on MSDN website. Compare the number and severity of issues in both lists. IMO the first list is longer, and it has more items that could ruin my day, such as:
- Error message when you connect to a named instance of SQL Server on a client computer that is running Windows Vista or Windows Server 2008
- The Log Reader Agent skips some transactions when the Log Reader Agent runs to replicate transactions
- Error message when you run a query that involves an outer join operation in SQL Server 2008
- Error message when you perform an update or a delete operation on a table that does not have a clustered index created in SQL Server 2008
- A query that uses parameters and the RECOMPILE option returns incorrect results when you run the query in multiple connections concurrently in SQL Server 2008
Let us drill down one more level and consider just one command, the MERGE. It was released as part of SQL 2008 with several problems, described in the following links:
As such, at the time of the original release of SQL 2008, I decided against using MERGE. I am using MERGE a lot now, on 2008 R2, and I think it is a really great feature.
Edit: here is the list of defects in SQL 2012 that were recently fixed. Hope it helps.
Another edit: I have chosen MERGE for more detailed analysis, because it is a very important improvement. In fact, it is a major step in catching up with Oracle, and it does improve our productivity. As such, MERGE has been marketed a lot at the time of SQL 2008 release. Yet it was not completely ready to use in serious production systems when it was originally released, and there was no easy way to know it from the presentations/articles/blog posts and such.
Similarly, snapshot isolation is an awesome new feature which just works, but invoking scalar UDFs in CHECK constraints does not work in all cases and as such should not be used in production when we need data integrity. However, both new features were recommended in "What is new in SQL xxxx" presentations, as well as in books, articles etc, and with similar enthusiasm.
We need to be very careful with new features - not all of them are going to be useful/reliable/performant.