[tae898]

Port of humemai/arcadedb-embedded-python#522 (branch perf/native-tx-crossings, closed there when the bindings moved, never merged) onto current main. Refs #19. Opened as a draft: a timed A/B decides it before merge.

NativeDatabase.run_in_transaction made 4 library crossings on the happy path (_transaction_stack_mark, begin, _transaction_commit_mark, commit), each one a ctypes call, a malloc'd JSON reply, a decode and a json.loads, although the begin reply already carries {"active","depth","commits"} and begin() discarded it. The commit-count witness now comes from the begin reply (new internal _begin_state(); public begin() still returns None). The unwind after a failed block asked is_transaction_active() before every rollback and once more at the end, and the retry check asked for the commit count again; every rollback reply carries the same state, so the new _unwind_opened_since(mark) returns the state of its last crossing and the retry check reads the commit count from it. A crossing that failed leaves the state unknown and the next step reads it again, as before. _end_transactions_opened_since(mark) keeps its signature and behaviour. Python only, no Java change, no image rebuild.

Counts (lib.call invocations, tests/test_native_tx_crossings.py, native only and not collected on JPype): run_in_transaction happy path 4 to 3 (mark, begin, commit); a block that raises 6 to 4; a refused attempt that is retried and then commits 11 to 7; with db.transaction() stays at 3. Two crossings is not reachable from Python alone: the mark has to be taken before begin (a mark taken after it counts the level begin pushed as the caller's), so it is a crossing of its own. Semantics are unchanged by design and pinned by tests that pass before and after (a nested block that fails leaves the caller's transaction open, a refused attempt is rolled back before the retry, a duplicated key is retried once, an error without a retry kind is raised after one attempt). The P-03 guard (the witness check raises outside its try) is already on main and is kept.

tests/test_interrupt_hygiene.py patched Database.begin and Database.rollback, which the native run_in_transaction no longer calls, so on native the patches would never apply. The tests now patch the seam the active backend's run_in_transaction really calls (begin and rollback on the JPype Database, _begin_state and _rollback_state on the native one). The one test with no native equivalent (an interrupt between begin and a separate commit-mark read, which the begin reply made one crossing) is removed on native instead of skipped, because scripts/check_test_skips.py fails a native run on any skip. JVM behaviour is untouched.

No speed claim is made. Whether 1 of 4 crossings shows in a small-transaction rate is decided by a timed A/B (tools/ab.py, laptop, main's native wheel against this PR's wheel) before this leaves draft. Draft PRs skip the test-native.yml legs, so the PR's wheel and the parity run come from gh workflow run test-native.yml --ref perf/native-tx-crossings. Checks run locally on the changed files: compile, pyflakes (only the findings main already has), black. The tests themselves have not been run on this port; the parity run is the first run.

🤖 Generated with Claude Code

[tae898]

Timed A/B of this PR against main (laptop timed lane, ab.py --noise, ABBA pairs, an A/A control; the rule was written in the study before the first run). Both arms use main's native wheel; arm B swaps in this PR's database.py and records.py, so libarcadedb.so is the same file in both.

figure	role	clean pairs	median PR / main	95% CI
rit_empty_us	headline	29	0.773	[0.770, 0.777]
rit_insert1_us	headline	29	0.988	[0.982, 0.993]
rit_raise_us	headline	29	0.715	[0.711, 0.717]
rit_insert10_us	target	29	0.993	[0.990, 0.997]
ctl_with_tx_us	guard	29	1.003	[0.994, 1.004]
ctl_first_us	guard	29	0.993	[0.989, 1.001]

Result: ok (all 6 B/A figures have >= 10 clean pairs and CI width <= 0.08, the A/A control is clean). By the rule the result was GRAY at round 12 (the one-insert transaction gains less than the 3 % the rule asks of every headline figure), so it was extended to round 30 as the rule says; still GRAY there, the median decides: every headline and target median favours the PR and the guards did not move, so it ships. In short: the fixed cost of run_in_transaction drops by about a quarter (empty and raising blocks), which is about 1 % of a transaction that inserts a record.