perf(ai-sdk): embed query once across router fan-out + fold umlauts in intent/concept matching

Authority Router re-embedded the query per collection (6x); on dev the embed
endpoint (OVH) is remote so that was 6 round-trips = 7-12s per /retrieve. Embed
once, reuse via ctx across the concurrent per-collection searches.
DetectIntent + ConceptNorms now fold ae/oe/ue/ss so ASCII (Pruefe) and umlaut
(Pruefe) inputs both match.

ai-compliance-sdk/internal/ucca/authority_router.go

@@ -74,6 +74,10 @@ func (c *LegalRAGClient) Retrieve(ctx context.Context, query string, topK int) (
 		}
 	}
 
+	// Embed the query ONCE and stash it in ctx so the concurrent per-collection searches
+	// below reuse it instead of each re-embedding (was N remote round-trips on dev/OVH).
+	ctx = c.withQueryEmbedding(ctx, query)
+
 	out := make([][]LegalSearchResult, len(collections))
 	var wg sync.WaitGroup
 	for i, coll := range collections {