fix(agents): accept Primary-hot warmup KV path

scripts/agent_gan_inference_demo.py

@@ -18,7 +18,10 @@
 
 def _agent_cache_gate(warm_delta: dict, actual_delta: dict) -> bool:
     return (
-        warm_delta["remote_hits"] >= 1
+        (
+            warm_delta.get("remote_hits", 0) >= 1
+            or warm_delta.get("local_hits", 0) >= 1
+        )
         and warm_delta["tokens_reused"] >= 1
         and warm_delta["tokens_computed"] == 0
         and warm_delta["fallbacks"] == 0

scripts/agent_gan_repl.py

@@ -157,6 +157,23 @@ def _stage(
     return stage
 
 
+def _gate_failure(name: str, warm: dict, actual: dict) -> RuntimeError:
+    keys = (
+        "local_hits",
+        "remote_hits",
+        "remote_jobs",
+        "tokens_reused",
+        "tokens_computed",
+        "fallbacks",
+        "remote_job_failures",
+    )
+    compact = lambda delta: {key: delta.get(key, 0) for key in keys}
+    return RuntimeError(
+        f"{name} KV gate failed: "
+        f"warm={compact(warm['delta'])} actual={compact(actual['delta'])}",
+    )
+
+
 def main() -> int:
     install_signal_protection()
     parser = argparse.ArgumentParser()
@@ -274,7 +291,11 @@ def get_stats():
                     generator_text,
                 )
                 if not generator_stage["ok"] and not telemetry_state["degraded"]:
-                    raise RuntimeError("Generator KV gate failed")
+                    raise _gate_failure(
+                        "Generator",
+                        generator_warm,
+                        generator_actual,
+                    )
                 if remote_run:
                     _telemetry_request(
                         f"{args.dashboard}/v1/network/benchmarks/{run_id}",
@@ -337,7 +358,7 @@ def get_stats():
                     extra_metrics=context_metrics,
                 )
                 if not critic_stage["ok"] and not telemetry_state["degraded"]:
-                    raise RuntimeError("Critic KV gate failed")
+                    raise _gate_failure("Critic", critic_warm, critic_actual)
                 completed = None
                 if remote_run:
                     completed = _telemetry_request(

tests/inference_engine/bridge/test_agent_gan_demo.py

@@ -23,6 +23,10 @@ def test_agent_gate_accepts_remote_compute_or_exact_remote_cache_hit():
     }
     assert _agent_cache_gate(warm, actual)
     assert _agent_cache_gate({**warm, "remote_jobs": 0}, actual)
+    assert _agent_cache_gate(
+        {**warm, "remote_hits": 0, "local_hits": 1, "remote_jobs": 0},
+        actual,
+    )
     assert not _agent_cache_gate({**warm, "remote_hits": 0}, actual)
     assert not _agent_cache_gate({**warm, "tokens_reused": 0}, actual)
     assert not _agent_cache_gate({**warm, "fallbacks": 1}, actual)

tests/inference_engine/bridge/test_agent_gan_repl.py

@@ -5,6 +5,7 @@
 from scripts.agent_gan_repl import (
     PrefillHeartbeat,
     TokenPrinter,
+    _gate_failure,
     _stage,
     _telemetry_request,
     build_critic_messages,
@@ -160,6 +161,18 @@ def timeout(*_args, **_kwargs):
     assert "inference will continue" in output
 
 
+def test_gate_failure_exposes_reuse_counters():
+    error = _gate_failure(
+        "Generator",
+        {"delta": {"local_hits": 0, "remote_hits": 0}},
+        {"delta": {"local_hits": 0, "fallbacks": 1}},
+    )
+    message = str(error)
+    assert "Generator KV gate failed" in message
+    assert "'remote_hits': 0" in message
+    assert "'fallbacks': 1" in message
+
+
 def test_interactive_prompts_are_deterministic_for_kv_reuse():
     generator_a = build_generator_messages("prove RH")
     generator_b = build_generator_messages("prove RH")