Minor nit: append(yamlFiles, regoFiles...) may reuse yamlFiles's backing array if it has excess capacity, silently clobbering it. filepath.Glob returns a tightly-sized slice in practice so this is safe today, but it's a fragile assumption. Consider:

ruleFiles := make([]string, 0, len(yamlFiles)+len(regoFiles))
ruleFiles = append(ruleFiles, yamlFiles...)
ruleFiles = append(ruleFiles, regoFiles...)

This silent continue was fine for YAML (where non-ruletype YAML files legitimately live in directories). For .rego files the assumption is much stronger -- any .rego file in a testdata directory is almost certainly meant to be a ruletype, so swallowing the parse error gives no signal when metadata is malformed. At minimum a log.Printf or zerolog warn here would save a lot of head-scratching.

The zero UUID is a fine sentinel, but uuid.UUID{}.String() is a somewhat roundabout way to write it. It constructs a value type just to call String(). The literal "00000000-0000-0000-0000-000000000000" (used in the deleted YAML file) is clearer and avoids the uuid import entirely:

rt.Context.Project = ptr.Ptr("00000000-0000-0000-0000-000000000000")

The previous code guarded this with if rt != nil. ReadResourceTyped can't actually return (nil, nil) today (it always returns a pointer), so removing the guard is technically correct -- but it's a non-obvious invariant. A short comment here would prevent someone from reflexively restoring the guard in a future cleanup:

// rt is never nil when err is nil; ReadResource always returns a pointer value.

Trailing space on a blank comment line (# instead of #). The parser handles it fine, but OPA convention (and most editors) use a bare # for blank comment lines. Worth trimming.