FAISS backends ignore the requested distance metric #89

The faiss-cpu and faiss-gpu backends accept metric but always build IndexFlatL2, so e.g. knn_dist_metric="cosine" silently returns Euclidean neighbors.

• cellmapper/src/cellmapper/model/_knn_backend.py

  Lines 53 to 77 in f68b451

  	class _FaissCpuBackend(_KNNBackend):
  	def __init__(
  	self,
  	n_neighbors: int,
  	metric: str,
  	random_state: int = 0,
  	**kwargs: Any,
  	):
  	check_deps("faiss-cpu")
  	import faiss
  	self.faiss = faiss
  	self._index = None
  	def fit(self, data: np.ndarray) -> None:
  	dims = data.shape[1]
  	index = self.faiss.IndexFlatL2(dims)
  	# Ensure data is float32 and C-contiguous
  	data_f32 = np.ascontiguousarray(data.astype(np.float32))
  	index.add(data_f32)
  	self._index = index
  	def query(self, points: np.ndarray, k: int) -> tuple[np.ndarray, np.ndarray]:
  	# Ensure points are float32 and C-contiguous
  	points_f32 = np.ascontiguousarray(points.astype(np.float32))

• cellmapper/src/cellmapper/model/_knn_backend.py

  Lines 80 to 101 in f68b451

  	class _FaissGpuBackend(_KNNBackend):
  	def __init__(
  	self,
  	n_neighbors: int,
  	metric: str,
  	random_state: int = 0,
  	**kwargs: Any,
  	):
  	check_deps("faiss-gpu")
  	import faiss
  	self.faiss = faiss
  	self.res = faiss.StandardGpuResources()
  	self._index = None
  	def fit(self, data: np.ndarray) -> None:
  	dims = data.shape[1]
  	flat = self.faiss.IndexFlatL2(dims)
  	gpu_index = self.faiss.index_cpu_to_gpu(self.res, 0, flat)
  	# Ensure data is float32 and C-contiguous

Expected: honour the metric (e.g. IndexFlatIP on L2-normalised data for cosine) or raise for unsupported metrics. The only FAISS test checks importability (tests/model/test_kernel.py), so a test comparing neighbors against the sklearn backend would catch this.

Related: squared distances from FAISS (separate issue).