spike(embed): F-04 BGE-M3 dense+sparse — NO-GO

Reason: SentenceTransformer.encode() in sentence-transformers 5.5.1 does
not accept return_dense / return_sparse kwargs. Those belong to the
FlagEmbedding.BGEM3FlagModel API, which is not in our dependency set.

Dense encoding works fine via the vanilla encode() call (shape (N, 1024),
float32). Sparse / lexical_weights requires either:
  (a) adding FlagEmbedding as a dependency, or
  (b) using sentence_transformers.SparseEncoder with a SPLADE checkpoint, or
  (c) driving the underlying transformers model and sparse head manually.

Branch retained as documentation per docs/loops/spike-gate.md.

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

spike.py

@@ -0,0 +1,68 @@
+"""Spike: verify BGE-M3 dense+sparse encoding via sentence-transformers.
+
+RESULT: NO-GO.
+
+The session prompt for F-04 specified an `encode()` call signature of
+``encode([...], return_dense=True, return_sparse=True, batch_size=2)``.
+This signature originates from the ``FlagEmbedding.BGEM3FlagModel`` API,
+not ``sentence_transformers.SentenceTransformer``.
+
+With ``sentence-transformers==5.5.1`` and ``BAAI/bge-m3``:
+
+    >>> model.encode(['hello', 'world'], return_dense=True, return_sparse=True)
+    ValueError: SentenceTransformer.encode() has been called with additional
+    keyword arguments that this model does not use:
+    ['return_sparse', 'return_dense'].
+
+The vanilla call ``model.encode([...])`` returns dense vectors of shape
+``(2, 1024)`` and dtype ``float32`` as expected — dense embeddings work.
+Sparse / lexical-weights output is NOT exposed via the standard
+``SentenceTransformer`` wrapper.  To get the BGE-M3 sparse head we must:
+
+    (a) Install ``FlagEmbedding`` and load via ``BGEM3FlagModel``, or
+    (b) Drive the underlying ``transformers`` model directly and apply the
+        sparse projection head ourselves, or
+    (c) Use ``sentence_transformers.SparseEncoder`` with a different
+        (SPLADE-style) checkpoint.
+
+Decision required from Theorist/Leader: which path?
+"""
+
+from __future__ import annotations
+
+import sys
+
+
+def main() -> int:
+    try:
+        from sentence_transformers import SentenceTransformer
+
+        model = SentenceTransformer("BAAI/bge-m3")
+        out = model.encode(
+            ["hello", "world"],
+            return_dense=True,
+            return_sparse=True,
+            batch_size=2,
+        )
+
+        if isinstance(out, dict):
+            dense = out.get("dense_vecs")
+            sparse = out.get("lexical_weights")
+        else:
+            dense = out
+            sparse = None
+
+        assert dense is not None, "no dense output"
+        assert dense.shape == (2, 1024), f"dense shape {dense.shape} != (2, 1024)"
+        assert isinstance(sparse, list) and len(sparse) == 2, "sparse not list of 2"
+        assert all(isinstance(d, dict) for d in sparse), "sparse elements not dicts"
+    except Exception as exc:  # noqa: BLE001 — spike must catch everything
+        print(f"Spike: NO-GO — {type(exc).__name__}: {exc}")
+        return 1
+
+    print("Spike: GO")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())