fix(batch): set has_errors on per-object/reference errors in stream recv

[vaibzde]

Closes #2107.

What's broken

_BatchBaseSync.__recv (and the identical code in _BatchBaseAsync) builds a fresh BatchObjectReturn for each failed object coming back on the batch stream:

result_objs += BatchObjectReturn(
    _all_responses=[err],
    errors={cached.index: err},
)

has_errors isn't passed, so it keeps its dataclass default of False. BatchObjectReturn.__add__ only ORs it forward (self.has_errors = self.has_errors or other.has_errors), so it never flips to True even though errors has an entry. Same thing for the reference-error branch (BatchReferenceReturn).

This is the code both data.ingest() and collection.batch.stream() use, so result.has_errors stays False on both, and the idiomatic if result.has_errors: check never fires. failed_objects/failed_references are still correct because those lists are populated separately in the same block, from the raw ErrorObject/ErrorReference, not through the buggy has_errors accumulation — that's why #2107's own repro saw failed_objects come back right on batch.stream() while only checking has_errors on ingest().

Fix

Pass has_errors=True at the four call sites that construct these objects from a server-reported error (2 in sync.py, 2 in async_.py).

How to verify

Added test_ingest_has_errors_on_failed_object to mock_tests/test_batch.py — a mock gRPC BatchStream service that returns one object error, driven through collection.data.ingest(), asserting result.has_errors is True. It fails on main (has_errors comes back False even though errors has one entry) and passes with this fix.

Ran locally (uv venv --python 3.12 .venv, uv pip install -e . -r requirements-test.txt -r requirements-devel.txt):

• pytest mock_tests/ — 52 passed (including the new test)
• pytest test/ — 386 passed, 1 skipped
• ruff format --check / ruff check — clean
• flake8 --config=.flake8 — clean
• pyright (v1.1.400, matching .pre-commit-config.yaml) — 0 errors

I didn't run the integration suite (integration/) since it needs a live Weaviate server — deferring to CI there.

What I didn't touch

Left the actual server-side error paths, retry/backoff logic, and failed_objects/failed_references tracking alone — this is scoped to the one missing kwarg per call site.

[orca-security-eu]

Orca Security Scan Summary

Status	Check	Issues by priority
Passed	Infrastructure as Code	0   0   0   0	View in Orca
Passed	SAST	0   0   0   0	View in Orca
Passed	Secrets	0   0   0   0	View in Orca
Passed	Vulnerabilities	0   0   0   0	View in Orca

[weaviate-git-bot]

To avoid any confusion in the future about your contribution to Weaviate, we work with a Contributor License Agreement. If you agree, you can simply add a comment to this PR that you agree with the CLA so that we can merge.

beep boop - the Weaviate bot 👋🤖

PS:
Are you already a member of the Weaviate Forum?

[vaibzde]

I agree with the CLA.

[g-despot]

Thanks for this, and sorry for the collision: we had a fix for #2107 in flight and I opened #2120 without checking that you already had one up. That was my mistake. Yours came first, so I've closed mine and pushed our work onto this branch instead.

What I pushed on top of your commit:

• Moved the fix from the four call sites into __post_init__ on BatchObjectReturn and BatchReferenceReturn, deriving has_errors from errors. Your diagnosis was exactly right; the reason for the different seam is that setting the flag per call site keeps it as a second source of truth, and that is how it drifted in the first place — the stream receive loop was added without the kwarg. Deriving it means a future construction site cannot reintroduce the bug, and it covers BatchReferenceReturn in the same move.
• Dropped the four has_errors=True kwargs your commit added, since they are redundant once the flag is derived.
• Kept your test_ingest_has_errors_on_failed_object as is. With the kwargs gone it now guards the seam directly: revert __post_init__ and your test is one of the ones that fails.
• Extended your mock servicer to reject every other object rather than all of them, so it also covers a mixed result, and added cases for the async ingest path, batch.stream(), and the two BatchReferenceReturn paths.