Naming injection_ml in Guard(scanners=) blocks every input #199

Naming injection_ml in Guard(scanners=[...]) gives you a Guard that blocks everything.

Guard(scanners=["injection_ml"], model="tiny").scan("What is the capital of France?").risk_score
# 1.0

Guard(model="tiny").scan("What is the capital of France?").risk_score
# 0.0

Two injection_ml scanners get built. One is backed by TransformersSpanProvider and works. The other is backed by NullModelProvider, whose predict() returns None, so injection_ml.py:124 does list(prediction.spans) on None and raises AttributeError. base.py:48 catches it and emits a scanner_error finding with score=1.0, which becomes the risk for every input.

Nothing in the output says the scanner failed unless you are reading logs. The caller sees a confident block on plain text.

injection_ml.py:95 is the part that lets it through:

if self._model.loaded is False:
    self._model.load()

A provider that cannot produce a prediction gets asked for one anyway, and the None that comes back is trusted. Same result with scanners=["injection", "injection_ml"].

Two things worth deciding separately. Why a second scanner is constructed with a null provider when a real one is available, and whether NullModelProvider.predict should return an empty SpanPrediction or raise rather than returning None.