feat(onnx): load-time warning when the model keys on integration-context fields

[omob]

Follow-up to finding F3 from the efficacy validation (#99) — the detection half; the retrain that actually fixes the model is a separate piece of work.

What

A contextSensitivityProbe runs next to the existing calibration probe on every model load and registry hot-swap. It scores the same plausible transaction twice — once with the optional integration-context fields populated (is_authenticated, device_is_trusted, channel_code, currency_code, session_to_txn_seconds), once with them absent, exactly what an integrator sending only the six required fields produces. If the score gap exceeds 0.5 it logs a loud warning with remediation guidance; the measured gap is exposed via getModelInfo().

Warning only — never a readiness failure. The model still functions; the operator needs to know its scores track payload richness rather than behaviour.

Why

#99 measured the shipped model as a trust-context detector: 687/687 bare-payload transactions scored >0.65 while 770/770 full-context transactions scored ≤0.65 — mule networks, trusted-device velocity bursts, and rings all invisible. Nothing in the stack could see this failure mode. Now every load self-checks.

Verification

Run against the shipped fraud_model.onnx:

contextSensitivityProbe: Warn. Model score swings by 0.9998 on optional
integration-context fields alone (same transaction, context fields present
vs absent). bareScore=0.9998 fullScore=0.0000 gap=0.9998

Lint clean, existing onnx suite green.

🤖 Generated with Claude Code