When xsl:number is used with level="any" and a from pattern, and no node matches the from pattern, the spec is reasonably clear that the root node of the tree should be treated as if it matched the from pattern, and that numbering should therefore be from the root:

"Let matches-from($node) be a function that returns true if and only if the given node $node matches the pattern given in the from attribute, or if $node is the root node of a tree."

However, quite a few tests in the test suite break if you implement this rule as written. An example is numbering18. These tests are written on the assumption that if no node matches the from pattern, then no number is allocated.

Martin Honnen reports on xsl-list that Saxon and Intel are following the behaviour shown in tests such as numbering18, not the behaviour described in the spec. We therefore need to decide whether to endorse the spec as written, or whether to change it to match the behaviour of the tests.

Changing the spec might not be trivial to do, and could end up being messy,  because the definition in question affects all values of "level". My preference would be to change the tests, and regard these products as being in error.

The XSLT 1.0 spec is hopelessly ambiguous in this area. However, it is very likely that these tests were carried forward from 1.0 tests and reflect the behaviour of at least some XSLT 1.0 implementations.