[tae898]

Closes #225

Python only, native build. NativeDatabase.begin(), commit(), rollback() and is_transaction_active() no longer decode the arcadedb_n_tx reply with json.loads. Success, and the active bool, are read from the fixed prefix that ArcadeNative.txState writes ({"ok":true,"active":true,...). The JSON is still decoded for an error reply (so the ArcadeDBError text, .native and .java_chain are the same), for a reply with any other shape, and for the callers that read depth or commits (_transaction_stack_mark, _begin_state, _rollback_state, the unwind in run_in_transaction). begin() and rollback() also skip two wrapper frames (_begin_state and _tx/_call): five frames before, three now.

What is not changed: the crossings (the existing counts in tests/test_native_tx_crossings.py are untouched and lib.call is still the one door), the Java side, the JVM build.

New tests in tests/test_native_tx_crossings.py (native-only file, runs in the native parity run) count json.loads per operation. test_begin_commit_rollback_decode_nothing, test_is_transaction_active_decodes_nothing_and_reports_the_state and test_with_transaction_decodes_only_the_mark are the ones that prove the cut: on main's code they count 4, 5 and 3 decodes and fail (to be shown by the CI run on main, not run locally). test_a_state_reply_still_carries_depth_and_commits, test_an_error_reply_is_decoded_into_the_same_error and test_a_differently_spaced_ok_reply_is_still_read pin what must stay the same and pass on both.

Verification: not run locally (a timed study holds the laptop); the CI runs below are the evidence. Static checks done: compile, pyflakes (only the existing availability-probe imports), bandit==1.9.4 over src and tests files touched, pre-commit (black, isort).

Speed: the numbers in #225 are an untimed discovery pass, directions only. A timed one-library A/B (begin+commit, empty with db.transaction(), a 10-row insert transaction, control: a non-transactional query) is pre-registered and runs on the laptop timed lane before merge, so this stays a draft until its verdict. A further cut (fusing mark and begin, or a packed return value from a new entry) would need a Java change and is not part of this PR.

🤖 Generated with Claude Code

[tae898]

Timed A/B (native wheel; arm A = main's CI native wheel, arm B = the same tree with only this PR's files, A/A control; laptop timed lane, tools/ab.py --noise, ABBA, round 0 dropped). Per-operation time, B/A median with 95% CI (below 1 = this PR is faster):

operation	role	B/A	A/A
empty transaction (begin, commit)	gating	0.738 [0.735, 0.739]	0.999 [0.994, 1.008]
begin and commit around one read	gating	0.576 [0.570, 0.579]	1.001 [0.990, 1.010]
transaction inserting 10 records	gating (no-harm bar 1.03)	0.986 [0.983, 0.990]	1.000 [0.997, 1.006]
query outside a transaction	control	1.000 [0.995, 1.007]	1.006 [0.993, 1.015]
rollback	info	0.554 [0.548, 0.561]	1.003 [0.996, 1.011]
is_transaction_active	info	0.569 [0.566, 0.578]	1.007 [0.997, 1.011]
transaction inserting 1 record	info	0.929 [0.926, 0.933]	1.001 [0.999, 1.002]

Clean pairs per figure: 12; every run passed its arm check, and the error-path differential (same values, states, error classes and texts on both arms) passed before the run. By the rule written before the run this is a SHIP: both gating transaction figures are below 0.90, the 10-record transaction is within its no-harm bar, and the control is flat.

The merge waits for the end of the current batch of timed studies that pin main (merging now would move their pins); the PR leaves draft at that point.