known issues: ArcadeData/arcadedb#9377 and #9378 are fixed in the 26.11.1 snapshot (#9383); tripwires become regression tests, light_edges=True into an undeclared type is refused - #257

Fixes #256

Upstream ArcadeData#9383 (merged f5bb4ac on 2026-10-07) fixes ArcadeData#9377 and ArcadeData#9378, and the 26.11.1 snapshot image that CI builds against now carries it. Main CI is red on every platform (run 37556915393): 8 of 895 tests fail, all of them in tests/test_light_edge_and_view_known_issues.py, and nothing else.

The three ArcadeData#9377 strict xfail tripwires pass (XPASS strict), as designed. They are now plain regression tests that pass on the snapshot and fail on the 26.10.1 wheel.

The five ArcadeData#9378 tests that loaded light edges into an edge type not declared LIGHTWEIGHT failed because that load now raises: ArcadeDBError from new_edge, Failed to buffer batch edge 'E': java.lang.IllegalArgumentException: Edge type 'E' does not declare LIGHTWEIGHT, so a batch built with withLightEdges(true) cannot store its property-less edges as light edges. Declare the type with CREATE EDGE TYPE E LIGHTWEIGHT, or build the batch without withLightEdges(true). One plain test now asserts the refusal. Declared LIGHTWEIGHT types, light_edges=False, an omitted option, and edges with properties in an undeclared type all count right and have tests.

Upstream says edges that older versions already wrote are not detected, so the workaround stays documented. Those tests would have had no way to build their data on the snapshot, so the graph is now built with the engine's newLightEdge (through the Java vertex), which writes the same light edge into an undeclared type on every engine, so the tests run in CI with no skip. I measured that the snapshot still answers CONSTANT COUNT 0 for such a graph, and one strict xfail keeps that caveat honest: it fails the suite if an engine ever repairs them. WITH a, b, count(b), and count(r) over a named relationship count 2 on the 26.10.1 wheel and on the snapshot.

Docs: both Known Engine Issues entries now say fixed in 26.11.1 (ArcadeData#9383), not released, still present in the 26.10.1 wheel, with the workarounds kept for the wheel, the new refusal quoted for light_edges=True, and the caveat that data written by an older version stays wrong. The light_edges warnings in docs/api/graph_batch.md, docs/api/database.md, and the Database.graph_batch docstring state the refusal, and the docs/development/testing.md row is updated.

Verification: I put the jars of the newest fork Build ArcadeDB engine jars artifact (run 37559131707, arcadedb-lib, 26.11.1-SNAPSHOT, which carries the refusal) into a scratch copy of the 26.10.1 wheel's package, swapping only the 11 arcadedb-* jars, and ran the suite from it with Python 3.12. The 8 failures reproduced exactly as in the CI log before the change. After it: the new file gives 19 passed and 1 xfailed on the snapshot, and 15 passed, 4 failed (the refusal test and the three ArcadeData#9377 tests), 1 xfailed on the 26.10.1 wheel. The full suite on the snapshot gives 895 passed, 1 xfailed, 0 skipped, 0 failed, and scripts/check_test_skips.py accepts the JUnit file. mkdocs build --strict passes and both new links resolve to the entry's anchor. pre-commit and bandit at low/low are clean.

Not verified: the exact jar set CI assembles (I used the wheel's own selection of jars with the snapshot's versions), so the CI run on this PR is the real check.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ubz3m6QXHESerCkehC1qgi