the moondream local rung has never done real inference — needs someone with CUDA or Apple Silicon

[LetsGetToWorkBro]

What happened

#650 fixed the moondream adapter to construct vl(local=True) only (the bare call was a cloud client that transmitted frames before auth failed — read the PR), and #651 moved to moondream 2.0.1. The local path is kestrel's Photon runtime, and it hard-requires an accelerator:

# moondream/photon_vl.py (2.0.1)
raise RuntimeError(
    "Photon local inference needs a supported accelerator, but neither "
    "CUDA nor Apple Silicon MPS is available in this Python environment.")

Every CI runner here is CPU-only, so test_moondream_stays_local.py proves the adapter's construction through a fake factory — but nobody in this repo has ever watched MoondreamClassifier return a real answer from the real runtime. Per CLAUDE.md #1, that's a gap worth being honest about: the guard tests are green and the actual capability is unverified.

Ask

You have an NVIDIA GPU or an Apple Silicon Mac:

cd host-python
pip install -e ".[vision]"
python - <<'PY'
from PIL import Image
from dreamlayer.object_lens.classify_backends import MoondreamClassifier
clf = MoondreamClassifier()
print(clf(Image.open("any-photo-of-an-object.jpg")))
PY

Report back: hardware, what it printed, wall time for first call (weights download) and second call (warm). If it fails, the traceback is exactly what this issue wants — the except path in _ensure will have swallowed it, so run with logging at DEBUG.

Stretch goal: a real_model-marked test for it. Note the tension before you write one — real-models.yml fails on any skip by design, and its runner is CPU-only, so the test cannot live there until a GPU runner exists. A manually-invoked script under tools/ (or a mark that only a self-hosted runner selects) is the honest shape.

Where

host-python/src/dreamlayer/object_lens/classify_backends.py (MoondreamClassifier), host-python/src/dreamlayer/tests/test_moondream_stays_local.py