diff --git a/ariv/runner/app.py b/ariv/runner/app.py
index 9f736d3..4ba7d43 100644
--- a/ariv/runner/app.py
+++ b/ariv/runner/app.py
@@ -31,7 +31,10 @@ class EvalRequest(BaseModel):
 
 
 def _registry_path() -> Path:
-    return Path(os.getenv("ARIV_MODELS_YAML", "ariv/models/models.yaml"))
+    env_path = os.getenv("ARIV_MODELS_YAML")
+    if env_path:
+        return Path(env_path).expanduser().resolve()
+    return (Path(__file__).resolve().parents[1] / "models" / "models.yaml").resolve()
 
 
 app = FastAPI(title="ARIV Runner", version="0.2.0")
diff --git a/ariv/runner/llama_cli.py b/ariv/runner/llama_cli.py
index 442bbab..b134036 100644
--- a/ariv/runner/llama_cli.py
+++ b/ariv/runner/llama_cli.py
@@ -61,6 +61,15 @@ async def stream_chat(
         )
 
         assert process.stdout
+        stderr_chunks: List[bytes] = []
+
+        async def _collect_stderr() -> None:
+            assert process.stderr
+            async for chunk in process.stderr:
+                stderr_chunks.append(chunk)
+
+        stderr_task = asyncio.create_task(_collect_stderr())
+
         async for raw_line in process.stdout:
             line = raw_line.decode("utf-8").strip()
             if not line:
@@ -77,6 +86,17 @@ async def stream_chat(
                 yield str(token)
 
         await process.wait()
+        await stderr_task
+
+        if process.returncode != 0:
+            stderr_tail = b"".join(stderr_chunks).decode("utf-8", errors="replace").strip()
+            if len(stderr_tail) > 1200:
+                stderr_tail = stderr_tail[-1200:]
+            raise RuntimeError(
+                "llama-cli failed: "
+                f"binary={self._binary}, model={model_path}, exit_code={process.returncode}, "
+                f"stderr={stderr_tail or '<empty>'}"
+            )
 
     async def run_chat(
         self,
diff --git a/core/orchestrator.py b/core/orchestrator.py
index df3658f..3e73df0 100644
--- a/core/orchestrator.py
+++ b/core/orchestrator.py
@@ -6,9 +6,14 @@
 import os
 import time
 from typing import Dict, Optional, Any, List, Callable
-from llama_cpp import Llama
+try:
+    from llama_cpp import Llama
+    _HAS_LLAMA = True
+except ModuleNotFoundError:  # pragma: no cover - optional runtime dependency
+    Llama = Any  # type: ignore[assignment]
+    _HAS_LLAMA = False
 from .vram_manager import VRAMManager, MemoryProfiler
-from ..tools.registry import ToolRegistry
+from tools.registry import ToolRegistry
 import logging
 
 logger = logging.getLogger("JugaadOrchestrator")
@@ -87,6 +92,11 @@ def load_model(self,
         load_start = time.time()
 
         try:
+            if not _HAS_LLAMA:
+                raise ModelLoadError(
+                    "llama_cpp is not installed. Install project dependencies to load GGUF models."
+                )
+
             # Get memory optimization recommendations
             model_size_gb = os.path.getsize(model_path) / (1024**3)
             mem_opt = self.vram_manager.optimize_for_model(model_size_gb)
diff --git a/core/trv_pipeline.py b/core/trv_pipeline.py
index 9ecd9cf..2b03bdd 100644
--- a/core/trv_pipeline.py
+++ b/core/trv_pipeline.py
@@ -7,8 +7,8 @@
 import time
 from typing import Dict, List, Optional, Any
 from .orchestrator import JugaadOrchestrator
-from ..config import (INDIAN_LANGUAGES_22, PIPELINE_CONFIG, 
-                     COT_CONFIG, TOOL_CONFIG, ARC_CONFIG)
+from config import (INDIAN_LANGUAGES_22, PIPELINE_CONFIG, 
+                    COT_CONFIG, TOOL_CONFIG, ARC_CONFIG)
 import json
 
 logger = logging.getLogger("TRVPipeline")
diff --git a/core/vram_manager.py b/core/vram_manager.py
index 99c1eef..a2c57f5 100644
--- a/core/vram_manager.py
+++ b/core/vram_manager.py
@@ -5,7 +5,6 @@
 """
 
 import gc
-import torch
 import logging
 import time
 from typing import Optional, Dict, Any
@@ -14,6 +13,11 @@
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger("VRAMManager")
 
+try:
+    import torch
+except ModuleNotFoundError:  # pragma: no cover - optional runtime dependency
+    torch = None
+
 @dataclass
 class MemoryStats:
     """Memory statistics data class"""
@@ -60,7 +64,7 @@ def flush(self, aggressive: bool = True) -> Dict[str, Any]:
             gc.collect()
             
         # Step 2: Clear PyTorch CUDA cache multiple times
-        if torch.cuda.is_available():
+        if torch is not None and torch.cuda.is_available():
             for i in range(2 if aggressive else 1):
                 torch.cuda.empty_cache()
                 
@@ -92,7 +96,7 @@ def flush(self, aggressive: bool = True) -> Dict[str, Any]:
 
     def get_memory_stats(self) -> MemoryStats:
         """Get current GPU memory statistics with fragmentation analysis"""
-        if not torch.cuda.is_available():
+        if torch is None or not torch.cuda.is_available():
             return MemoryStats(
                 allocated_gb=0,
                 reserved_gb=0,
diff --git a/tests/test_llama_cli.py b/tests/test_llama_cli.py
new file mode 100644
index 0000000..a2986ae
--- /dev/null
+++ b/tests/test_llama_cli.py
@@ -0,0 +1,57 @@
+import asyncio
+
+from ariv.runner.llama_cli import LlamaCLI
+
+
+def test_stream_chat_raises_on_nonzero_exit(monkeypatch, tmp_path) -> None:
+    monkeypatch.delenv("ARIV_MOCK_LLAMA", raising=False)
+    model = tmp_path / "dummy.gguf"
+    model.write_text("x", encoding="utf-8")
+
+    class FakeStream:
+        def __init__(self, chunks):
+            self._chunks = chunks
+
+        def __aiter__(self):
+            self._iter = iter(self._chunks)
+            return self
+
+        async def __anext__(self):
+            try:
+                return next(self._iter)
+            except StopIteration as exc:
+                raise StopAsyncIteration from exc
+
+    class FakeProcess:
+        def __init__(self):
+            self.stdout = FakeStream([b'{"token":"ok"}\n'])
+            self.stderr = FakeStream([b"fatal error\n"])
+            self.returncode = 17
+
+        async def wait(self):
+            return self.returncode
+
+    async def fake_exec(*args, **kwargs):
+        return FakeProcess()
+
+    monkeypatch.setattr("asyncio.create_subprocess_exec", fake_exec)
+
+    cli = LlamaCLI(binary="llama-cli")
+
+    async def _consume() -> None:
+        async for _ in cli.stream_chat(
+            model_path=str(model),
+            prompt="hello",
+            num_gpu_layers=0,
+            max_tokens=2,
+        ):
+            pass
+
+    try:
+        asyncio.run(_consume())
+        raise AssertionError("Expected RuntimeError")
+    except RuntimeError as exc:
+        message = str(exc)
+        assert "exit_code=17" in message
+        assert "fatal error" in message
+        assert "binary=llama-cli" in message
diff --git a/tests/test_pipeline.py b/tests/test_pipeline.py
index 65f1c94..f1c0552 100644
--- a/tests/test_pipeline.py
+++ b/tests/test_pipeline.py
@@ -9,6 +9,7 @@
 def test_pipeline_phases():
     """Test that all 4 phases execute"""
     mock_orch = Mock()
+    mock_orch.models_config = {}
     mock_orch.generate.return_value = "test output"
 
     pipeline = TRVPipeline(mock_orch, {})
@@ -19,7 +20,7 @@ def test_pipeline_phases():
     pipeline._phase3_critic = Mock(return_value="PASS")
     pipeline._phase4_synthesis = Mock(return_value="final")
 
-    result = pipeline.execute("query", "hindi", enable_critic=True)
+    result = pipeline.execute("query", "hindi", enable_critic=True, enable_deep_cot=False)
 
     # Verify all phases called
     pipeline._phase1_ingestion.assert_called_once()
@@ -32,6 +33,7 @@ def test_pipeline_phases():
 def test_critic_loop():
     """Test critic iteration logic"""
     mock_orch = Mock()
+    mock_orch.models_config = {}
     pipeline = TRVPipeline(mock_orch, {})
 
     # First critic fails, second passes
@@ -40,7 +42,7 @@ def test_critic_loop():
     pipeline._phase1_ingestion = Mock(return_value="english")
     pipeline._phase4_synthesis = Mock(return_value="final")
 
-    result = pipeline.execute("query", "hindi", enable_critic=True)
+    result = pipeline.execute("query", "hindi", enable_critic=True, enable_deep_cot=False)
 
     # Should have 2 critic calls (fail then pass)
     assert pipeline._phase3_critic.call_count == 2
diff --git a/tests/test_runner_app.py b/tests/test_runner_app.py
index 6cc40ff..4af8f46 100644
--- a/tests/test_runner_app.py
+++ b/tests/test_runner_app.py
@@ -1,3 +1,4 @@
+import importlib
 import os
 
 from fastapi.testclient import TestClient
@@ -15,3 +16,11 @@ def test_chat_streaming_response() -> None:
     assert response.status_code == 200
     text = response.text
     assert "metadata" in text
+
+
+def test_registry_path_independent_of_cwd(monkeypatch, tmp_path) -> None:
+    monkeypatch.delenv("ARIV_MODELS_YAML", raising=False)
+    monkeypatch.chdir(tmp_path)
+    runner_module = importlib.import_module("ariv.runner.app")
+    importlib.reload(runner_module)
+    assert runner_module.registry.get("mock-0.1b-q4_0").name == "mock-0.1b-q4_0"
diff --git a/tests/test_vram.py b/tests/test_vram.py
index c372f9c..5badf3e 100644
--- a/tests/test_vram.py
+++ b/tests/test_vram.py
@@ -2,23 +2,22 @@
 """Tests for VRAM Manager"""
 
 import pytest
-import torch
-from core.vram_manager import VRAMManager
+from core.vram_manager import VRAMManager, torch
 
 def test_vram_stats():
     """Test VRAM stats collection"""
-    if not torch.cuda.is_available():
+    if torch is None or not torch.cuda.is_available():
         pytest.skip("CUDA not available")
 
-    stats = VRAMManager.get_memory_stats()
-    assert stats['available'] == True
-    assert 'allocated_gb' in stats
-    assert 'total_gb' in stats
-    assert stats['total_gb'] > 0
+    manager = VRAMManager()
+    stats = manager.get_memory_stats()
+    assert stats.total_gb > 0
+    assert stats.available_gb >= 0
+    assert stats.allocated_gb >= 0
 
 def test_flush_protocol():
     """Test VRAM flush doesn't crash"""
-    if not torch.cuda.is_available():
+    if torch is None or not torch.cuda.is_available():
         pytest.skip("CUDA not available")
 
     # Allocate some memory
@@ -26,8 +25,9 @@ def test_flush_protocol():
     del x
 
     # Should not raise
-    VRAMManager.flush()
+    manager = VRAMManager()
+    manager.flush()
 
     # Memory should be reduced (or at least not crash)
-    stats = VRAMManager.get_memory_stats()
-    assert stats['available'] == True
+    stats = manager.get_memory_stats()
+    assert stats.available_gb >= 0