[tae898]

Closes #219.

Native NativeResultSet.to_json_list now walks the rows with codec.json_native only when a reply of the result restored a tagged value, and json_native returns JSON-safe values for an RID and for map keys. Draft: the timed A/B (one-library overlay, ab.py) is prepared and not run, so the speed claim is the untimed probe in #219 until it is.

What changes

• codec.loads_tagged(text, floats) is loads plus a flag, true when a tagged value was restored. NativeResultSet keeps _tagged across every reply it decodes: the first batch in _exec, every arcadedb_n_next and arcadedb_n_next_bin batch in _fill, and a lazy command result starts false. The flag is sticky, so a result with an untagged first batch and a later tagged one (or the reverse) walks. to_json_list returns the decoded rows as they are when the flag is false. A caller that builds a result set without saying (NativeResultSet(db, reply)) gets true, the old behaviour.
• A reply whose only tags are ["B", offset, count] (the float tail of vectors) decodes to plain lists of Python floats, so it does not set the flag and is not walked. This is tested, and a reply with vectors and a date does walk.
• json_native returns the text of an RID ("#10:5"), also in a list and in a map. A map key that is not a string becomes a string: an RID its text, True "true", None "null", a number its str.
• to_list is unchanged: it still returns RID objects and the Python key types (tested).
• Tagged results give the same output as before for DATE and DATETIME (epoch milliseconds), DECIMAL (float) and sets (lists).

The value for a link, and the one deliberate difference. The JVM wheel's to_json_list returns the RID text for a link property, a list of links, SELECT @rid and a map value (read on the 26.11.1.dev0 JVM wheel, 2026-11-04), and string keys for a map stored with integer keys ({'1': 'a'}; JSON objects have string keys). The native wheel now returns the same, and tests/test_to_json_list_links.py states those values and runs on both backends (it passes on the JVM wheel too). The one change to an output that was JSON-safe: a tagged map with integer keys used to come back with int keys ({1: 'a'}) and now has str keys, as the JVM wheel has. json.dumps of either gives the same text.

Tests (all new, none skipped; tests/test_native_json_list.py is native only through conftest.py, tests/test_to_json_list_links.py runs on both)

I ran the new files against the unmodified CI native wheel (tree of arcadedb-26.11.1.dev0-py3-none-manylinux_2_28_x86_64.whl, commit 413f54f, whose codec.py and database.py equal origin/main): 25 failed, 5 passed in 0.45s. The tests that fail on the old code are the ones that state the change (an untagged result is not walked, a float-tail-only reply is not walked, an RID and a non-string key become strings, the loads_tagged flag, the links on both builds). The five that pass on the old code are guards for what must not change (an empty result, a lazy command result that must still walk, to_list keeping the Python types, to_list keeping RID objects and integer keys, and the other json_native conversions). The tests of a tagged or mixed result also fail on the old code, but only through their premise check (the old code has no flag to read); their walk-count asserts hold on the old code by construction, and the mutants below are what shows they guard the new code. I also checked the guards with two mutants of the new code: a non-sticky flag fails the tagged-first-then-untagged test, and counting a "B" tag as restored fails the float-tail tests.

How I ran the tests. The CI native wheel installed as in test-native.yml (uv run --no-project --python 3.12 --with <wheel> --with pytest --with numpy --with requests --with pyarrow --with jpype1 --with setuptools --with pandas python -m pytest), with the extracted wheel tree first on PYTHONPATH and the two changed modules copied over it (libarcadedb.so is the wheel's), taskset -c 8-15. 519 passed and 23 skipped (all NA-JVM or NOTYET as the native parity run reports them) over test_native_json_list.py, test_to_json_list_links.py, test_native_codec_loads.py, test_native_result_transport.py, test_native_point_read.py, test_resultset.py, test_nonfinite_floats.py, test_results_after_close.py, test_resultset_read_errors.py, test_columnar_readers.py, test_float_vector_bulk.py, test_bulk_json_paths.py and test_core.py. Pre-commit (black 26.10.1, isort, the name check), compile, pyflakes (only the existing # noqa imports) and Bandit 1.9.4 over src tests are clean.

Not in this PR. The timed A/B. The overlay study only changes _native/codec.py and _native/database.py over the main CI wheel, so libarcadedb.so is identical in both arms.

🤖 Generated with Claude Code

Docs updated in 7c17592 (docs and docstrings only).

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

[tae898]

Timed A/B (native build; arm A = main 413f54f, arm B = A's tree with only _native/database.py and _native/codec.py from this PR (PR 223 88aafa5), A/A control; laptop timed lane, ab.py --noise, ABBA pairs; the rule written before the run). Each figure is the statement plus to_json_list() on a result shaped like the benchmark's graph and sparse reads, time ratio B/A:

figure	median B/A	95% CI	clean pairs
one-row point read (jl_point_us)	0.958	[0.942, 0.964]	12
top 10, three columns (jl_top10_3col_us)	0.987	[0.970, 0.999]	12
100-row histogram (jl_hist100_us)	0.976	[0.971, 0.980]	12
100-row scan, four columns (jl_scan4_r100_us)	0.953	[0.946, 0.959]	12
sparse search top 10 (jl_sparse10_us)	0.984	[0.979, 0.987]	12
control: to_list() on the 100-row scan	0.993	[0.990, 1.001]	12

Result by the rule: SHIP (every gating interval below 1.00, the to_list() control not moved, the A/A control clean). The gain is the second walk that untagged results no longer take; it is a few percent of the whole call because the engine's statement time dominates these small results. The same PR makes RID values (links, @rid, map keys) JSON-safe in to_json_list(), as on the JVM build.

Marking this ready. It merges after the other pending timed A/B that pins src/arcadedb/_native has run, so its pins stay valid.