`duplicate-names` XSL transformation is too slow on large XMIR files

I'm getting performance warnings when running lints against eo-runtime:

[WARNING] XSL transformation 'duplicate-names' took 151ms, whereas threshold is 100ms
[WARNING] XSL transformation 'duplicate-names' took 158ms, whereas threshold is 100ms

What happens: For each named <o> the transformation checks:

<xsl:if test="preceding-sibling::o/@name = $x/@name">

This scans all preceding siblings for every element, making it O(k²) per parent where k is the number of named children. On objects with many children (common in generated XMIR from bytecode) this becomes the dominant cost.

What should happen: The transformation completes within the 100ms threshold. Using xsl:key to group siblings by name would allow detecting duplicates in a single pass instead of per-sibling scans.