My biggest hesitation to enabling canisters to transfer ICP would be in de facto also enabling neuron transfer, if this is in fact an unavoidable consequence, per the exchange above between you and @wang. This hesitation is rooted in the security risks of neuron transfer suggested in this thread here, for example.
Neuron immobility is currently at the heart of the model of secure good governance underpinning the IC as a whole. (Other risks of enabling canisters to transfer ICP seem more or less limited to particular potentially insecure or malicious canisters, and mitigations seem possible, based on the discussion so far.)