I assumed it would follow my hint, and maybe error out at execution time if I wound up with some bad data and the index was missing some needed values.
The query optimizer will only use a filtered index in a query plan if it can guarantee (within its reasoning framework) that all possible matches can be served from the index. This is by design, to avoid the sort of runtime error you describe.
Failure to results in a NESTED LOOPS JOIN from my non-clustered index against a clustered index Key Lookup, presumably to grab the parentId. INCLUDING parent ID eliminates this, and leaves me with a nice non-clustered index scan.
This is a known current limitationknown current limitation. Adding the filtered column(s) to the key or include list is the standard workaround, and a current best practice for all sorts of semi-related reasons.
The FORCE ORDER, MERGE JOIN is definitely needed though.
Be extremely careful using hints (directives) like this unless you fully understand all the consequences.
FORCE ORDER in particular is an extremely powerful and wide-ranging hint, with a number of non-obvious side-effects including the placement of aggregate operators, and the order of evaluation of subqueries and common table expressions.
For the most part, you should try to write queries that provide the query optimizer with enough good-quality information to make the right decisions without hints. The hinted plan may be 'optimal' today, but it may not remain so as the data volume and/or distribution changes over time.