feat(advisor): Authority Router — Advisor collection-agnostisch, KB-2026.1-Gewinn im Produktpfad

Der Advisor fan-outete bisher selbst ueber eine feste Liste expliziter Collections
(advisor-rag.ts) und umging damit das #61-Scope-Routing (das nur den Default-Pfad
routet) → der gemessene +28-Retrieval-Gewinn (CB-100: 53→81, 0 Regr) kam nie beim
Antwort-LLM an. Dieser Router zieht den Fan-out in die Retriever-Schicht:

- SDK: LegalRAGClient.Retrieve() + POST /sdk/v1/rag/retrieve {query, top_k} —
  fan-outet server-seitig ueber die Broad-Authority-Base + die KB-2026.1-Slice bei
  inKBScope, merge+dedup, sortiert nach Authority-Score (rerankByAuthority je
  Collection), top-K. Index-Warmup vor dem nebenlaeufigen Fan-out (Map-Race-frei).
  Per-Env via RAG_ROUTER_COLLECTIONS.
- admin: advisor-rag.ts ruft EINMAL /retrieve statt 6-fach expliziter Collections.
  Advisor ist collection-agnostisch (Vertrag Compiler→Collections→Retriever→Advisor);
  COMPLIANCE_COLLECTIONS/searchCollection entfernt.

Validierung: Go-Unit (Router-Selektion, dedup); e2e gegen dev-Qdrant (echter
Retrieve(), CB-100-Stichprobe stride 5): OLD-hit 11/20 → NEW-hit 15/20, GAIN 4
(alle DS-Guidance), REGR 0 — reproduziert den +28/0-Regr durch den Produktionscode.
TS-Tests auf den Single-/retrieve-Call angepasst.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

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

@@ -0,0 +1,114 @@
+package ucca
+
+import (
+	"context"
+	"os"
+	"sort"
+	"strings"
+	"sync"
+)
+
+// routerBaseCollections is the broad authority base the Authority Router fans out over. It mirrors
+// the Advisor's historical multi-collection set; the KB-2026.1 slice is added separately when the
+// query is in scope. Override via RAG_ROUTER_COLLECTIONS (comma-separated) per environment.
+func (c *LegalRAGClient) routerBaseCollections() []string {
+	if v := strings.TrimSpace(os.Getenv("RAG_ROUTER_COLLECTIONS")); v != "" {
+		var out []string
+		for _, p := range strings.Split(v, ",") {
+			if s := strings.TrimSpace(p); s != "" {
+				out = append(out, s)
+			}
+		}
+		if len(out) > 0 {
+			return out
+		}
+	}
+	return []string{
+		"bp_compliance_gesetze",
+		"bp_compliance_ce",
+		"bp_compliance_datenschutz",
+		"bp_dsfa_corpus",
+		"bp_compliance_recht",
+		"bp_legal_templates",
+	}
+}
+
+const routerPerCollectionTopK = 3
+
+// Retrieve is the Authority Router entry point: callers (the Advisor) pass ONLY a query and stay
+// collection-agnostic. The router fans out over the broad authority base and ADDS the KB-2026.1
+// slice when the query is in scope (inKBScope), then merges all hits, deduplicates, and returns the
+// top-K by authority score. This moves the former Advisor-side collection fan-out into the retrieval
+// layer (the "Retriever" tier of the quality pyramid), so the proven KB-2026.1 slice gain reaches
+// the product path without the Advisor knowing about individual collections.
+//
+// The merged set is ordered by the per-collection authority score that rerankByAuthority already
+// produced inside searchInternal — i.e. binding-vs-guidance ordering is preserved across the merge.
+// Per-collection failures (e.g. a collection absent on an environment) degrade gracefully.
+func (c *LegalRAGClient) Retrieve(ctx context.Context, query string, topK int) ([]LegalSearchResult, error) {
+	if topK <= 0 {
+		topK = 8
+	}
+
+	collections := c.routerBaseCollections()
+	if c.kbScopeRoutingEnabled && c.kbSliceCollection != "" && inKBScope(query) {
+		collections = append(collections, c.kbSliceCollection)
+	}
+
+	// Warm the full-text indexes sequentially first so the concurrent fan-out below only READS the
+	// shared textIndexEnsured map (the writes happen here, serialized) — closes the cold-start map
+	// race deterministically. Best-effort: a missing collection just stays un-indexed (hybrid then
+	// falls back to dense, or the per-collection search degrades to nothing).
+	if c.hybridEnabled {
+		for _, coll := range collections {
+			_ = c.ensureTextIndex(ctx, coll)
+		}
+	}
+
+	out := make([][]LegalSearchResult, len(collections))
+	var wg sync.WaitGroup
+	for i, coll := range collections {
+		wg.Add(1)
+		go func(i int, coll string) {
+			defer wg.Done()
+			if res, err := c.searchInternal(ctx, coll, query, nil, routerPerCollectionTopK); err == nil {
+				out[i] = res
+			}
+		}(i, coll)
+	}
+	wg.Wait()
+
+	merged := make([]LegalSearchResult, 0, len(collections)*routerPerCollectionTopK)
+	for _, r := range out {
+		merged = append(merged, r...)
+	}
+	merged = dedupResults(merged)
+	sort.SliceStable(merged, func(a, b int) bool { return merged[a].Score > merged[b].Score })
+	if len(merged) > topK {
+		merged = merged[:topK]
+	}
+	return merged, nil
+}
+
+// dedupResults removes duplicate passages that can appear when collections overlap, keeping the
+// highest-scoring occurrence. Identity = regulation_code + article_label + a text prefix.
+func dedupResults(in []LegalSearchResult) []LegalSearchResult {
+	pos := make(map[string]int, len(in))
+	out := make([]LegalSearchResult, 0, len(in))
+	for _, r := range in {
+		text := r.Text
+		if len(text) > 80 {
+			text = text[:80]
+		}
+		key := r.RegulationCode + "|" + r.ArticleLabel + "|" + text
+		if idx, ok := pos[key]; ok {
+			if r.Score > out[idx].Score {
+				out[idx] = r
+			}
+			continue
+		}
+		pos[key] = len(out)
+		out = append(out, r)
+	}
+	return out
+}