Function responses_parity 

pub async fn responses_parity(
    handle: &PluginHandle,
    smaller: Value,
    larger: Value,
    expect_delta: usize,
    want: &(dyn Fn(&str) -> bool + Sync),
    timeout: Duration,
) -> Cell

2026-09-26: The Responses API sees the image, and sizes it the same.

/v1/responses has its own content vocabulary (input_image, input_text) and its own lowering into the IR (openai/responses_lowering.rs).

The assertion is a difference inside this one surface, not a comparison with chat completions: the chat-completions legs turn thinking off, this request asks for reasoning.effort “low”, so their prompts differ in envelope as well as image. Two images whose vision-token counts differ by a known amount both go through Responses; the envelope is the same twice, so it cancels:

  (tokens_b - tokens_a)  ==  (vision_b - vision_a)

Both must also answer correctly, which shows the pixels arrived rather than merely being counted. A 404 on the first request is a skip.