9.2 Atlas consumer findings

This log records integration findings at the boundary between ml-eng-lab and its pinned Atlas submodule. Do not patch Atlas implementation files from this repository; report upstream defects there and record the consumer impact here.

9.2.1 Native Ollama is mandatory

Status: enforced in scripts/atlas-up.sh and covered by lifecycle tests.

The local workstation's containerized Ollama path consumes excessive memory and is too slow for this consumer. The committed contract therefore requires ollama-localhost, clears ambient source overrides, verifies the loopback native daemon before startup, and rejects container/automatic ComfyUI overrides. Keep this finding in place unless a reviewed Atlas change gives equivalent native behavior with a different explicit source.

9.2.2 Atlas Jupyter runtime is distinct from local CI

Status: documented with live probe evidence on 2026-07-30.

Atlas JupyterHub is a distinct runtime and is not Issue #62 acceptance evidence. Issue #62 qualifies the repository Torch 2.11 CPU stack; Issue #65 completed the retained Atlas runtime review, and Issue #66 separately requires full quantization execution in both environments. See dependency-contracts.md and atlas-pin-bump-runbook.md before changing either side of that boundary.

Issue #62 did not upgrade Atlas. The host-native Ollama boundary is unchanged, and no containerized Ollama service is added. There is no containerized Ollama source in the Issue #66 quantization workflow.