fetch_objects(limit=0) returns a default page of objects instead of none

[nazsats]

What happened

limit=0 on the gRPC query path is dropped before it reaches the server, so
the server applies QUERY_DEFAULTS_LIMIT instead of returning nothing.

With 30 objects in a collection and the server started with
QUERY_DEFAULTS_LIMIT=25:

call	returned
collection.query.fetch_objects(limit=0)	25 objects
collection.query.fetch_objects(limit=1)	1
collection.query.fetch_objects(limit=2)	2
collection.query.near_vector(..., limit=0)	25
collection.query.hybrid(..., limit=0)	25
GET /v1/objects?class=X&limit=0	0 objects
GraphQL Difftest(limit: 0)	error: invalid pagination params: invalid default limit: 0

bm25 is unaffected, it fails earlier on a different validation path.

Versions

• weaviate-client 4.23.0
• weaviate server 1.38.13 (semitechnologies/weaviate:latest)

Reproduction

docker run -d -p 8080:8080 -p 50051:50051 \
  -e AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED=true \
  -e PERSISTENCE_DATA_PATH=/var/lib/weaviate \
  -e DEFAULT_VECTORIZER_MODULE=none \
  -e QUERY_DEFAULTS_LIMIT=25 \
  semitechnologies/weaviate:latest

import weaviate
from weaviate.classes.config import Configure, DataType, Property

client = weaviate.connect_to_local()
coll = client.collections.create(
    name="LimitZero",
    properties=[Property(name="n", data_type=DataType.INT)],
    vectorizer_config=Configure.Vectorizer.none(),
)
with coll.batch.fixed_size(batch_size=30) as batch:
    for i in range(30):
        batch.add_object(properties={"n": i}, vector=[float(i), 0.0, 0.0])

print(len(coll.query.fetch_objects(limit=0).objects))  # 25, expected 0
client.close()

Why it happens

SearchRequest.limit is a proto3 uint32 with no field presence, so 0 is the
type default and is not serialised:

from weaviate.proto.v1 import search_get_pb2
search_get_pb2.SearchRequest(collection="X", limit=0).SerializeToString()
# b'\n\x01X'
search_get_pb2.SearchRequest(collection="X", limit=2).SerializeToString()
# b'\n\x01X\xf0\x01\x02'

limit=0 is byte-identical to sending no limit at all.

I want to be clear that I think this part is intended, not a wire bug.
grpc/proto/v1/search_get.proto in weaviate/weaviate says so:

// affects order and length of results. 0/empty (default value) means disabled
uint32 limit = 30;
uint32 offset = 31;
uint32 autocut = 32;

So 0 meaning "disabled" is a deliberate convention shared with offset and
autocut, and I am not proposing changing the proto.

What I think the actual problem is

The Python signature is limit: Optional[int] = None. None already means
"no limit", so there is nothing in the Python API to suggest 0 is a second
spelling of the same thing. The convention is documented in a .proto file
that a Python user has no reason to open, and it is not in the docstring or
the type.

A caller passing a computed limit — min(page_size, remaining),
wanted - already_fetched, a quota that reaches zero — asks for no objects and
receives a full default page, with no error and no warning.

The client already rejects other degenerate query inputs. contains_any([])
raises WeaviateInvalidInputError: Filter contains_any must have at least one value. limit=0 gets no such treatment.

There is also existing precedent for guarding this exact class of silent
coercion. weaviate/validator.py carries this from #2088:

# bool is a subclass of int, so isinstance(True, int) is True. Reject a
# bool where a plain int is expected so it is not silently coerced to 0/1.

What would you prefer

Two options, and I do not want to pick one for you:

1. Reject limit=0 with WeaviateInvalidInputError inside the existing
   if self._validate_arguments: block in collections/grpc/query.py, the way
   contains_any([]) is handled. This is a behaviour change for anyone
   currently relying on 0 meaning "disabled" from Python.
2. Document it — state in the docstring that 0 means no limit, same as None.

I have option 1 written with a test in test_bad_query_inputs, and the unit
suite passes. Happy to open a PR for either once you say which you want, or to
drop it if you would rather leave the behaviour alone.