[tae898]

Closes four NotOnNativeYet gaps of the native build and rewrites the module docstring of src/arcadedb/_native/__init__.py to what is true today. Lands after #67: this branch is restacked on native-callbacks and the base of this PR is native-callbacks, so the diff shows only the changes of this PR. Retarget to main once #67 is merged.

What works now on native

• VectorIndex.get_stats: the counters of the primary bucket index, from its schema:index:<name> row (a statement, no glue). The row's own columns (name, fileId, ...) are left out.
• VectorIndex.get_metadata and get_quantization of an index this session did not create (a reopened database). The engine's configuration comes from one new glue entry point, arcadedb_n_vindex (op metadata), because no statement reports it. The dict has the same 20 keys as on the JVM build, engine defaults included; before, native returned only what the session had passed.
• VectorIndex.find_nearest_approximate over PRODUCT quantization: findNeighborsFromVectorApproximate through the same entry point (op approximate). No SQL option runs the PQ search (vectorNeighbors takes efSearch, filter, where, groupBy, groupSize, maxDistance only).
• warm_up on native was a no-op; it now runs the first search that loads a persisted graph (this test had been hidden behind the get_stats skip).
• AsyncExecutor.transaction: a new op of arcadedb_n_async. The engine begins the transaction on a writer thread and Python runs the block in the upcall. A block that raises is rolled back and reaches the per-call and the global error callback, as on the JVM build.
• get_java_database and get_java_document stay JVM-only by design. which-package.adoc and installation.adoc now say so, and no longer list the four calls above as missing.

Callback registration: AsyncExecutor.transaction registers up to three callbacks per call. They are kept in the executor's _cb_ids (declared in __init__, unused until now) and forgotten (_forget, from #67) when wait_completion has returned or the executor is closed. A retry calls the block again, so they cannot be released earlier. The other async operations still register callbacks that are never released; that leak is known and unchanged here.

Dropped when restacking on #67, and why: this PR used to carry two fixes that #67 solves in its own way, and #67 wins in both places.

• The reply buffer of an upcall (a dict keyed by thread id in extras._dispatch, added so that an {"error": ...} reply from an async writer thread outlives that thread's Python state). Native: Python functions in SQL and openCypher, record events, Ctrl-C, and the async writer crash #67 removes every Python-owned reply buffer: the callback is void and writes library memory (arcadedb_n_alloc) into an out slot that Java zeroes before the call, and Java copies and frees it. That memory does not depend on the Python thread state, so the failure path of transaction needs no code of its own: the block's exception comes back as {"error": ...} through the out slot, the Java transaction case throws it, and the writer rolls back and reaches the error callback (test_async_transaction).
• lib._thread_handle() taking its isolate thread from graal_get_current_thread. Native: Python functions in SQL and openCypher, record events, Ctrl-C, and the async writer crash #67 has the same change, so nothing is left here.

Tests (written first, shared by both builds): tests/test_vector_index_introspection.py (14) and tests/test_async_transaction.py (4). They use no Java object, so they run on both backends unchanged, and neither file is collect-ignored.

• Native, one image built from this restacked tree (XMX=12g, glue includes Native: Python functions in SQL and openCypher, record events, Ctrl-C, and the async writer crash #67 and this PR) and the tests run in the foreground: test_async_transaction 4, test_vector_index_introspection 14, test_async_executor 21, test_python_callbacks 40, test_python_function_interrupt 3, test_native_udf 8, test_percall_calls 30, test_concurrency 5, test_free_threading 4, test_cypher 30, test_core 39 and the other six test_vector*.py files (test_vector 41, test_vector_delta_visibility 2, test_vector_params_verification 2, test_vector_sql 39, test_vector_second_pass 1, test_vector_product_quantization 12): 295 passed, 0 failed, 40 skipped (the existing NOTYET and NA-JVM skips only).
• Also run on the same image: test_sigint, test_interrupt_hygiene, test_transaction_context_interrupt, test_results_after_close, test_resultset, test_native_startup_runtime: 51 passed, 0 failed, 11 skipped.
• JVM build, in the sibling repo with the same test files (humemai/arcadedb-embedded-python): 18 passed, three runs. Not re-run after the restack; the two test files are unchanged.
• Before the change the two new files gave 16 NOTYET skips and 1 failure on native.
• pre-commit, scripts/check_no_embedded_name.py and Bandit (--severity-level low --confidence-level low, 0 issues) pass.

Glue change: native/src/main/java/org/humemai/nativelib/NativeExtras.java only (the transaction case in async() and the new arcadedb_n_vindex entry point), so it should merge with the other glue PRs except for the import lines. Not run in CI yet: the native build and parity jobs of native.yml. test_free_threading ran on Python 3.12 (GIL enabled), not on a free-threaded build.

🤖 Generated with Claude Code

[tae898]

Rebased onto native-two-dist 56aab39 (the one conflict was get_metadata in schema_objects.py: the base added the optional keys rerank_oversample and fused_pq, which the JVM build reports when the engine has them; native reports them once the glue does, with the move to 26.11.1, and the shared test treats them as optional). The glue is unchanged by the rebase, so the image built from this branch still applies; after the rebase the same suites give 171 passed, 0 failed, 37 skipped (existing NOTYET and NA-JVM skips). The diff in NativeExtras.java and _native/lib.py stays minimal: one import block and the transaction case plus the vindex entry point in the glue, and one block in _thread_handle in lib.py. The Python-only fix in extras.py means this PR needs no second image build.

🤖 Generated with Claude Code