[ice-black-tea]

Problem

On an ARM64 Docker image store, Compose can start a container using the selected platform manifest while tag inspection reports the enclosing index and container inspection reports a different image identity. Comparing those IDs makes a successful up fail and can force recreation of an untouched container.

Fix

• Let native Compose reconcile image changes. Check replica count/start state after detached up, while retaining the image-reference drift guard before deployment.
• Use observed container IDs and start timestamps for dependency actions. After a partial failure, preserve untouched original running replicas and do not record a previous successful one-shot exit as completion of this batch.
• Verify that the original immutable image reference resolves locally before attempting to recreate for recovery; do not fall back to a mutable tag.

Verification

• manage.py check linktools-cntr (focused on batch application, deployment recovery, lifecycle, runtime inspection, image snapshots, readiness and generated-input consumers): 352 passed. Ruff and Python 3.6 static compatibility passed.
• Regression cases cover different image ID layers, actual replacement, same-container restart dependencies, partial replica replacement/loss, an untouched completed job in both up and restart and an unavailable original image.
• The prior head failed eight CLI/running-state tests because their shared test fixture omitted inspection identity fields. The follow-up uses the real inspection data type; all eight reproduced before the change and pass with original assertions unchanged. Production code is unchanged by the follow-up.
• Real macOS Docker not rerun; latest natural CI pending.