[tae898]

What does this PR do?

I changed the native transaction entry point arcadedb_n_tx to return the transaction state as one long, not as a malloc'd JSON text. A begin, a commit, a rollback or a mark is now one foreign call: no string built in Java, no strlen, no copy into a str, no arcadedb_free, no json.loads. It works the way arcadedb_n_cmd already returns a cursor id.

New signature: public static long tx(IsolateThread thread, long h, int op) in ArcadeNative.java (was CCharPointer), restype c_longlong in lib.py, and a new Python door lib.tx(handle, op) -> int. The integer is the state after the call: bit 0 active, bits 1 to 31 depth, bits 32 to 62 the low 31 bits of commits (callers only compare two counts for equality), and -1 for an error.

	before	after
entries for with db.transaction() (mark, begin, commit)	3	3
foreign calls for with db.transaction()	12 (per entry: arcadedb_n_tx, strlen, PyUnicode_FromStringAndSize, arcadedb_free)	3
foreign calls for one begin, commit, rollback or is_transaction_active	4	1
json.loads for the mark	1	0

The error path is unchanged for the caller. On -1 the Java side has stored the same errorJson(t) text per thread (as arcadedb_n_cmd does), and Python reads it with lib.last_error() and passes it to the same _error(prefix, reply). The exception type, the message (Failed to commit transaction: <java class>: <message>), .native and .java_chain are the ones from before. A closed database still raises Database is closed before any crossing.

How I verified it: only by reading and by the light checks. I did not build the wheel or run a test, because the native image build and the suite run in CI. uvx pre-commit run --files on the changed files and Bandit on the changed Python files pass. The CI run on the fork is the proof.

Tests are in tests/test_native_tx_crossings.py, in the file that already counted these crossings, and the first commit holds only the tests:

• with db.transaction() is exactly three foreign calls, all arcadedb_n_tx, counted through a proxy of the loaded library and the two reply-reading helpers. Begin, commit, rollback, is_transaction_active and the mark are one each. A transaction never calls lib._take.
• json.loads is not called by a transaction (the old test expected one decode for the mark).
• The layout of the integer, -1 and the error text for an unknown handle, commit without begin, rollback with nothing to roll back, a failed statement inside a block, nested with blocks, an injected failure on each entry point, and a database closed with a transaction open.

They fail on the old code, where lib.tx does not exist and the counts are higher. tests/test_native_one_row_crossings.py now counts lib.tx under the name arcadedb_n_tx, so its existing expectations hold.

Not measured yet: the speed claim waits for a timed A/B of a candidate wheel against 26.11.1.post2. This draft only claims fewer foreign calls and no reply text, which the tests count.

Things I am not sure of: a failing commit that the engine refuses at commit time (not at the statement) is covered by the injected failure, not by a real conflict, because I did not find a deterministic way to make one on one thread. The wheel ships the library and the Python files together, so a library built before this change (ARCADEDB_NATIVE_LIB) would be read wrongly by the new Python.

Related issues

Part of #315

Performance changes

None claimed yet, see above.

Docs

No page or README text describes this entry point, so none needs an update.

Checklist

• A regression test fails on the old code and passes on the new (CI on the fork)
• The test suite that the change can affect passes (CI on the fork)
• uvx pre-commit run --files <changed files> passes

🤖 Generated with Claude Code

[tae898]

This change shipped in 26.11.1.post3 (on PyPI since 2026-11-14) through the release line, which #387 merged into main with a merge commit. The release line carries the same change as its own commits, so I close this draft. The measurements are on #315.