[tae898]

The native wheel's point reads cost 49.6 us per lookup_by_key and 33.7 us per SELECT ... WHERE id = ? plus .first() in the untimed profile (1M-document database, 100k calls each), against 20.1 and 17.5 us on the JVM wheel. The image's own timers account for about 55% of that wall; this change takes the Python side that was measurable and leaves every signature, result and error as it was. encode_params writes a lone int or str parameter directly (byte for byte the text json.dumps gives; bool, subclasses and numpy scalars keep the general path), lib.exec_query is one function for arcadedb_n_exec (no getattr by name, no argument tuple, cached bytes for language and statement), NativeResultSet.first() takes a buffered first row without the iterator protocol, lookup_by_key remembers its statement text per (type, keys), and the record wrapper no longer imports a module and makes a stand-in object per row. It does not touch call_bin, _fill, loads or the Java side, so it does not overlap #520 or #521.

What was measured, untimed (one run each on the dev1 image, perf_counter_ns around each step on cores 8-15, the timer read subtracted; a decomposition, not a speed claim), microseconds per call before and after, from repros/native-point-call-ab-laptop-20261025/decomp.out and decomp-after.out in the notes repository:

component (us per call)	SELECT .. WHERE id = ? .first() before	after	lookup_by_key before	after
build_sql	0.00	0.00	0.97	0.46
encode_params	2.39	0.58	2.81	0.71
lib_b_x3	0.34	0.27	0.40	0.35
json_loads	4.86	4.49	5.68	5.14
rs_init	0.72	0.69	0.91	0.82
first	0.74	0.55	0.96	0.67
rs_dealloc	0.19	0.18	0.22	0.22
get_element	0.00	0.00	2.41	0.91
whole call incl. two get() calls (real API, no timers inside)	34.36	29.60	53.37	43.21

The ctypes crossing itself is not the cost: the call minus the image's own timers is 0.6 us. The largest piece left on the Python side is the reply decode (json_loads, 4.9 and 5.7 us), which belongs to the result-transport work, and I left it alone.

What was not measured: any timed speed. The timed A/B is queued as native-point-call-ab-laptop-20261025 (arm A the published dev1 wheel, arm B dev1 plus these four Python files, arm Z the A/A control; ab.py with noise log, ABBA, cores 6,7); its rule is written before any timed run: ship only if the upper end of the 95% CI of B/A is below 0.95 for both point-read figures, no figure is slower (lower CI end above 1.03), and the A/A control is clean. Until it reports, nothing here is a speed claim.

Tests first: bindings/python/tests/test_native_point_call.py (38 tests, 32 behaviour tests that pass before and after on both backends where they apply, and 6 helper tests that fail on native-preview and pass here). With the dev1 image the new file plus test_resultset, test_results_after_close, test_graph_api, test_graph and test_vector give 116 passed, 24 skipped; a wider run of test_core, test_parameter_binding, test_percall_calls, test_typed_rows, test_bulk_insert and test_cypher before the last comment-only edits gave 255 passed, 43 skipped, 1 failed (a test of mine that was wrong, fixed).

Parity fix found by this PR's CI (commit b2d6481). The JVM leg failed test_lookup_by_key_returns_the_record_for_int_str_and_composite_keys with ArcadeDBError: Failed to lookup by key in 'P': java.lang.IllegalArgumentException: No index has been created on type 'P' properties [id, name]: the test looked up the composite key ["id", "name"] while _docs() created only single-property indexes. The test was wrong, and the JVM wheel was right. lookup_by_key is documented as an index lookup (docs/api/database.md) and the engine refuses keys no index covers. The native wheel ran SELECT FROM type WHERE k = ? AND ... LIMIT 1, which works without an index and silently scans the whole type, so the same test passed there for the wrong reason. _docs() now creates CREATE INDEX ON P (id, name) UNIQUE, and native lookup_by_key makes the engine's check: on the first lookup of a (type, keys) pair on a database handle it reads schema:indexes for the type once and raises ArcadeDBError("Failed to lookup by key in 'P': java.lang.IllegalArgumentException: No index has been created on type 'P' properties [id, name]") unless an LSM_TREE or HASH index covers exactly those properties in that order, on the type or a supertype (a FULL_TEXT index on the keys gets the engine's No key index has been created ... text; a type that does not exist keeps its schema error). A pair that passed is remembered on the handle next to the statement text, so every later lookup is still the one crossing this PR measured; the remembered set is per handle (not in the module-wide _LOOKUP_SQL text cache, because whether an index exists is a fact about one database) and DROP or ALTER through command(), and any sqlscript, clears it. Limit: an index dropped by another connection to the same database is not noticed until the handle's next DDL; that falls back to the old unchecked statement, never to a false error, and only refusals are never cached.

Tests for the parity fix: test_native_point_call.py is now 43 tests. Four new ones run on both wheels (keys no index covers: a property with no index, the composite in the wrong order, a prefix of the composite, a superset; three calls each; a missing type; another database with a type of the same name; DROP INDEX then a re-created index; an index on a supertype and a FULL_TEXT one) and one is native-only (exactly one schema:indexes read per pair, and 2 x 5 lookups plus 2 reads in total). On the unfixed native source the four parity tests fail (DID NOT RAISE ArcadeDBError) and the cache test fails; on the fixed source all pass. Run with the nnext image: this file plus test_graph_api, test_conversion_fixes, test_percall_calls, test_vector, test_graph, test_core, test_resultset and test_results_after_close give 198 passed, 46 skipped. On the JVM main wheel (78e97b1, Python 3.12): this file gives 10 passed, 33 skipped (the skips are the native-only helper tests), and the same wider set gives 208 passed, 35 skipped, 1 failed: test_vector.py::TestLSMVectorIndex::test_lsm_vector_search_uses_database_rid_conversion, a file this PR does not touch and a code path (_native) the JVM wheel does not load.

🤖 Generated with Claude Code

[tae898]

Closing here: the bindings moved to ArcadeData/arcadedb-python. The branch stays for the queued A/B native-point-call-ab-laptop-20261025; if it passes, the change is ported there.