clarify_query intrinsic enters repetition loop on embedded ALORA switch model

[yairallouche]

Summary

The rag.clarify_query intrinsic enters a degenerate repetition loop when run against
ibm-granite/granite-switch-4.1-3b-preview (ALORA embedded adapter, via vLLM/OpenAI backend).
The model produces a valid JSON opening but then repeats program names indefinitely until hitting max tokens, never closing the JSON.

Controlled experiments

Backend	Model	Adapter mechanism	Result
vLLM	granite-switch-4.1-3b-preview	ALORA embedded in switch checkpoint	FAIL
vLLM (enforce_eager)	granite-switch-4.1-3b-preview	ALORA embedded in switch checkpoint	FAIL
vLLM	GrizleeBer/gs-test-2	LORA embedded in switch checkpoint	PASS
HF	granite-4.1-3b	ALORA loaded via PEFT from ibm-granite/granitelib-rag-r1.0	PASS
HF	granite-4.1-3b	LORA loaded via PEFT from ibm-granite/granitelib-rag-r1.0	PASS

The same ALORA adapter weights work correctly when loaded via PEFT on the HF backend. The failure is specific to the embedded ALORA in the switch checkpoint.

[lastras]

The problem appears to be intermittent, caused by variability at inference time.