vLLM issue #59086: VLLM_BATCH_INVARIANT=1 is not batch-invariant for AWQ models on sm8x in the default mode

OPEN as of 2026-10-03, five days old at time of access. Same shape as #57016 but for AWQ quantized models on sm8x (Ampere-class) in the DEFAULT mode. Together the two issues establish that batch invariance in vLLM is currently opt-in per kernel path and not yet universal across (architecture x quantization) combinations - so the correct test posture is to assert the invariant on the exact config you ship, and treat a pass on another config as no evidence.