[tae898]

Closes #222

NativeGraphBatch._op sent every op argument as JSON text, so a float32 ndarray in create_vertices rows or in new_edge and new_edges props went through .tolist() into an F-tagged number list. This adds arcadedb_n_gb_bin(handle, op, json, floats, n_floats), arcadedb_n_gb plus the float32 side buffer that arcadedb_n_exec_bin already reads. The JSON carries a B tag with an offset and a count, and Json.decodeTag returns the float[] the F tag gave, so the stored values are bit-identical.

What changes:

• native/.../NativeExtras.java: arcadedb_n_gb and arcadedb_n_gb_bin share gbImpl, which passes the buffer to the parser. Json.object (the streaming parser of create_vertices) takes the buffer too; its old signature stays.
• _native/codec.py: encode_gb_args writes one-dimensional, non-empty float32 ndarrays to the side buffer and everything else as before. Lists, float64 arrays and rows without a vector keep their text.
• _native/lib.py: GB_BIN (the loader clears it when the library has no arcadedb_n_gb_bin; ARCADEDB_NATIVE_GB_BIN=0 turns it off for an A/B in one image), and the ctypes signature, next to arcadedb_n_exec_bin.
• _native/extras.py: _op picks the entry. A library built before the entry keeps the text path.
• No change to the JVM wheel.

Tests (run in the native job, against the wheel this branch builds):

• tests/test_native_graph_batch_binary_vectors.py (native only): a spy on lib.call sees arcadedb_n_gb_bin for float32 rows of 64, 96 and 128 dimensions and the stored vectors are bit-equal to the inputs (-0.0, the float32 maximum, the smallest normal and the smallest subnormal included). It also covers the binary path against the text path in the same image, the fallback by the probe flag, None, list, float64 and empty vectors staying on the text entry, mixed rows (every offset), the chunks of a 2,500-row load (one buffer each), edge props, and a tag outside the buffer being refused by the library.
• tests/test_float_vector_bulk.py::TestCreateVertices::test_float32_rows_are_stored_bit_equal: the same bit-equality at the three dimensions on both packages.

What I could and could not run: the laptop has no native build, so the Java was compiled with javac (Corretto 25, Maven artifacts of 26.10.2) and the Python tests were run against the CI wheel of main with this branch's three Python files laid over it, which is a library without the new entry. There the fallback tests pass, the 7 tests that need the entry fail as they must, and the shared bit-equality test passes. Against main's own Python code, 12 of the 13 new native tests fail (they need lib.GB_BIN or the new call name) and the one that pins the unchanged text path passes. The binary entry itself is exercised only by the native job of this branch, which is dispatched.

Draft until that job is green. A wheel-versus-wheel A/B (timed) follows from the CI wheels of main and of this branch; the numbers in the issue are an untimed probe.

🤖 Generated with Claude Code

Docs updated in 8e04007 (docs and docstrings only).

Edited 2026-11-04: added the docs line above.

[tae898]

Timed A/B (native build; arm A = main 7e64c5e, arm B = PR 224 9ac30ec, both the CI native wheels of their commits; A/A control; laptop timed lane, ab.py --noise, ABBA pairs; the rule written before the run). GraphBatch.create_vertices loads in the shapes of the benchmark's dense and hybrid lanes, time ratio B/A:

figure	median B/A	95% CI	clean pairs
128-d vectors, batch 10,000 (sift_load_s, gating)	0.517	[0.515, 0.520]	12
96-d vectors, batch 10,000 (deep_load_s, gating)	0.218	[0.217, 0.219]	12
64-d vectors, batch 5,000 (e2_load_s, gating)	0.271	[0.270, 0.272]	12
graph_batch load without vectors (ctl_load_s, control)	1.009	[1.003, 1.012]	12
peak RSS	1.004	[0.999, 1.007]	12

Result by the rule: SHIP (every gating interval below 1.00, the control inside [0.97, 1.03], the A/A control clean). Sending float32 vectors in the binary side buffer instead of as F-tagged JSON text removes most of the load time for vector-carrying vertices; loads without vectors stay inside the control band (the control interval is just above 1.00).

Marking this ready. It merges after the other pending timed A/Bs that pin src/arcadedb/_native have run, so their pins stay valid.

Edit: the sentence on loads without vectors said "unchanged"; the control interval lies just above 1.00, inside the rule's band, so it now says that.

[tae898]

Edited the comment above: it said loads without vectors are "unchanged"; the control interval lies just above 1.00 (inside the rule's band), and the sentence now says so. The verdict is the same.