Function cache_leak 

pub async fn cache_leak(
    handle: &PluginHandle,
    first: Value,
    second: Value,
    second_want: &(dyn Fn(&str) -> bool + Sync),
    first_marker: &(dyn Fn(&str) -> bool + Sync),
    timeout: Duration,
) -> Cell

2026-09-26: The same prompt text with a different image, back to back.

Vision prompts must not be served from the prefix cache. The model-engine guard is tokens_have_vision_pad, which must recognise both the image and the video pad; if it missed one, the second request would match the first’s cached prefix and answer about the previous image. first_marker tells that case apart from a merely wrong reply.