A fixed network input shape does not establish that image preprocessing is warm

A fixed network input shape does not establish that image preprocessing is warm. When latency varies between new and repeated images, time decoding, preprocessing, model forward and postprocessing separately with appropriate accelerator synchronization. Compare a first image size, repeats of that size, a different size, and different content returning to the first size. A first-use spike confined to preprocessing is evidence for source-shape-sensitive initialization, even if every network tensor has the same dimensions. Do not attribute it to model forward, OCR fallback, or JIT compilation without corresponding evidence.

Applicability and limits: GPU image pipelines with resize/color/layout stages, including restored workers. First-use behavior may exist in both old and new backends; historical differences in concurrency and cache state do not isolate a backend regression. Inspect stage routing separately so identical final decisions do not hide fallback changes. Kernel-cache inspection that finds no files is not proof that compilation did not occur; its path or mechanism may be different.