fix(embedder): list(dense) returns ndarray objects not float lists -- Qdrant serialization crash on every BgeM3 ingest #286

Summary

BgeM3Embedder.embed() returns list(dense) where dense is a 2D numpy array of shape (N, 1024). list(numpy_2d_array) yields a list of 1-D numpy arrays, NOT a list of Python float lists. _validate_embed_dimension passes them silently (numpy len() works). But when Qdrant's upsert_points() serializes the vector payload, it raises TypeError: Object of type ndarray is not JSON serializable, crashing the ingest pipeline after all parse/chunk work is done.

Root cause

Fix #245 added the three-key probe (dense_vecs / dense / embeddings) but the return statement at line 242 did not convert numpy arrays to Python lists.

Fix

ingestion/embedder.py line 242: change return list(dense) to return dense.tolist() which recursively converts numpy arrays to Python floats.

Impact

Every document ingestion using BgeM3Embedder crashes at the Qdrant upsert step with a confusing serialization error instead of indexing the document.