[Bug]: GPU vision warm-up stays at 4096 tokens with an 8192 image cap #1945

What happened

With strata-vision --gpu --max-tokens 8192, startup still encodes a fixed 2048 × 2048 dummy image. For the Qwen3.8-Flash-Next BF16 projector this warms only 4096 image tokens, even though real images can reach the larger configured cap.

A same-day build from main (61b3fb5dd3f1e8ec09cf7e4e05208bc6d3c46406) logged:

strata-vision: warmed up at 4096 image tokens in 8.2 s
READY 2560

The encoder's Windows dedicated-GPU-memory counter was 1,699,991,552 bytes at READY. Encoding a synthetic 3840 × 2160 image then produced 8160 image tokens and raised that counter to 2,305,748,992 bytes: 605,757,440 bytes allocated after READY.

The server starts the language engine after the encoder reaches READY so it can budget its expert cache against the remaining GPU memory. A late allocation defeats that ordering. This test observed the allocation growth; it did not reproduce an OOM or paging failure.

Expected: an explicitly higher image cap should also increase the GPU prewarm size before READY.

Reproduction (substitute matching local GGUF/projector paths and a sufficiently large 3840 × 2160 PNG):

strata-vision --mmproj <mmproj-Qwen3.8-Flash-Next-BF16.gguf> --model <first-text-GGUF-shard> --gpu --threads 16 --min-tokens 1024 --max-tokens 8192
ENC <large-image.png> <output.sve>
QUIT

Strata version, GPU, OS

Strata 0.1.42 source at 61b3fb5; AMD Ryzen AI Max+ 395 / Radeon 8060S (gfx1151), 128 GB unified RAM with 96 GB assigned to the GPU; Windows 11 Home build 26200; AMD driver 32.0.31032.1003. Standalone image encoder built with Vulkan SDK 1.4.350.0 and LLVM MinGW 23.1.3 (20261006), native CPU kernels. The deployed language engine uses HIP.

Engine log

The relevant encoder log is included above. A narrow fix and paired CPU/Vulkan verification are prepared. Related Windows AMD image support is discussed in #1155 and PRs #1287 / #1789. Those threads cover encoder builds and installer discovery; this report concerns the warm-up size in the existing encoder.