A better question
So instead of asking:
Is declarative better than imperative?
it seems more useful to ask:
How do you structure a system so that data drives the output, shared behavior
is centralized, while variation is handled cleanly and locally?
When a problem has that shape:

  *   declarative processing is a natural fit
  *   layered specialization becomes highly valuable
  *   XSLT aligns closely with the architecture

________________________________
Closing thought
The real strength of XSLT is not captured by small, one-off comparisons.
It becomes visible when:

  *   structured input drives output
  *   a shared core must support multiple specialized outputs
  *   variation must be introduced without duplicating logic
In those cases, the issue is not whether the language is declarative in
theory.
The issue is whether the architecture is aligned with the problem.
XSLT provides a model for that alignment:

  *   data-driven processing
  *   centralized default behavior
  *   local layered specialization
That is why the question "declarative or imperative?" is too small.
________________________________
Many thanks again to Ken Holman and others for helping clarify these ideas. I
found the distinction between data-driven processing and layered
specialization particularly illuminating.
Best,
Roger