Same image fails with no logs despite succeeding on a prior identical query

[aidanbruneel]

Observed Behavior:
While testing Qwen3.6 and Gemma 4 VLMs for #1167, a query for the cathedral test image succeeded on the first attempt for both models, but failed on subsequent, identical queries.

Both the orchestrator logs and the object-detection-llm preprocessor logs show no trace of the failed
request, suggesting that the request never reached (or was never processed by) the pipeline on that attempt.

Expected Behavior:
The same image, submitted the same way, should produce a consistent result across repeated queries (either always succeeding or failing with a traceable error). If a request is rejected or dropped for any reason, this should be visible in the orchestrator and preprocessor logs.

Reproduction:
Use web extension to query an image (e.g., cathedral test image) and see both orchestrator and preprocessor logs.

Note: Not reliably reproducible. I have tried to reproduce the same issue with the cathedral test image and several other images that gave me this issue but have been successful only 2-3 times.

Runtime:

• Tested on unicorn
• Models: unsloth/qwen3.6-27B-NVFP4 and unsloth/gemma-4-31B-it-qat-w4a16
• Image confirmed under the 10MB payload limit (succeeded on first attempt). Ruling out the cause described in #1197.