feat(agb): wire validated routed AGB engine into live check path

Consolidate the AGB C-lean engine (71% FP -> ~0, validated vs 7-company
Opus GT) onto the canonical checker library and into the live check path.

- AGBAgent.evaluate now runs routed C-lean: keyword (L1/L2) -> business-
  model gate -> per-item decision_method routing (embedding/reference/llm
  via services/checkers/) -> severity re-tiering (LOW -> recommendation),
  honoring context.skip_llm.
- New agb/_pipeline.py orchestrates the routing; agent.py stays thin.
- Remove the 3 AGB-local checker duplicates (_reference_check,
  _embedding_rescue, _llm_judge); services/checkers/ is now canonical.
- Wire "agb" into _agent_outputs._TOPIC_AGENTS so the live check emits a
  validated AGB tab (was snapshot-only).
- Run topic agents concurrently (asyncio.gather) + emit each tab via SSE
  as it finishes -> progressive results, no wait on the slowest agent.
- Tests: checker units (mocked), routed agent (gate/rescue/re-tier),
  topic wiring; existing AGB tests made offline-safe.

dev-only, no deploy.

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

backend-compliance/compliance/api/agent_check/_agent_outputs.py

@@ -13,6 +13,7 @@ the map). Once the tabs are the source of truth, B18's v1 path retires.
 
 from __future__ import annotations
 
+import asyncio
 import logging
 
 from compliance.services.specialist_agents import REGISTRY, AgentInput
@@ -27,6 +28,7 @@ logger = logging.getLogger(__name__)
 # topic key (matches state["doc_texts"]) -> registered agent_id
 _TOPIC_AGENTS: dict[str, str] = {
     "impressum": "impressum",
+    "agb": "agb",  # v2: AGBAgent mit decision_method-Routing (71% FP -> ~0)
 }
 
 _MIN_TEXT = 100
@@ -112,14 +114,17 @@ async def run_agent_outputs(state: dict) -> None:
     )
 
     outputs: dict[str, dict] = state.get("agent_outputs") or {}
-    for topic, agent_id in _TOPIC_AGENTS.items():
+
+    async def _run_one(topic: str, agent_id: str):
+        """Einen Topic-Agent laufen lassen + sein Tab-Event sofort emittieren
+        (Zwischenbefund). Fängt eigene Fehler → ein Agent reißt den Run nicht ab."""
         text = (doc_texts.get(topic) or "").strip()
         if len(text) < _MIN_TEXT:
-            continue
+            return None
         agent = REGISTRY.get(agent_id)
         if agent is None:
             logger.warning("agent_outputs: agent '%s' not registered", agent_id)
-            continue
+            return None
         try:
             out = await agent.evaluate(AgentInput(
                 doc_type=topic,
@@ -128,15 +133,25 @@ async def run_agent_outputs(state: dict) -> None:
                 company_name=company_name,
                 origin_domain=origin_domain,
             ))
-            outputs[topic] = out.model_dump(mode="json")
-            emit(check_id, {"type": "topic", "topic": topic,
-                            "output": outputs[topic]})
+            dump = out.model_dump(mode="json")
+            emit(check_id, {"type": "topic", "topic": topic, "output": dump})
             logger.info(
                 "agent_outputs[%s]: %d findings, confidence %.2f",
                 topic, len(out.findings), out.confidence,
             )
+            return topic, dump
         except Exception as e:  # noqa: BLE001 — best-effort, never break the run
             logger.warning("agent_outputs[%s] failed: %s", topic, e)
+            return None
+
+    # Topic-Agenten laufen NEBENLÄUFIG (ihre Embedding-/LLM-Waits überlappen) und
+    # füllen ihren Tab via SSE, sobald sie fertig sind — kein Warten aufs Schlusslicht.
+    results = await asyncio.gather(
+        *(_run_one(topic, agent_id) for topic, agent_id in _TOPIC_AGENTS.items())
+    )
+    for r in results:
+        if r:
+            outputs[r[0]] = r[1]
 
     if outputs:
         state["agent_outputs"] = outputs