Vision admission release: multi-image requests past 16,384 rows (fixes #10)

[data-angel]

Pins Atlas sparkqwen/atlas-20261017-vision (92dd5435, tree 34e71b7b): the drafter release plus 20 vision commits. Image tag 3b2deea3d187. Fixes #10.

• The Qwen encoder's output window holds one prefill chunk's images, so a request is no longer capped at 16,384 vision rows (one 2048x2048 image's worth). The issue's 16,680-row request and 5 or 8 x 2048x2048 now return 200 with every image in order, cold and warm.
• Fixed a second bug found along the way: warm multi-chunk image prompts could splice a cached image's rows in place of the next image's.
• Admission is now decided from image headers, before any decode or GPU work. A request over the limit gets 400 vision_input_too_large. A request that would overflow the in-flight budget gets 503 vision_capacity_busy with Retry-After: 2. Defaults are 32,768 rows per request and 65,536 in flight, set by --vision-max-request-rows and --vision-max-inflight-rows.
• Decode memory is bounded at 640 MiB per image, with at most 2 large decodes at once.
• Images are read with their EXIF orientation applied.
• Pixels are released at the end of prefill.

Text is unchanged against the drafter release:

• Greedy output is 8/8 identical at c1, c4 and c8, and with the prefix cache at c8.
• The prompt-logprob hash is 798fc6cc2641.
• Decode speed is level.

The recipe image built from a fresh clone passed its build-time tests and a smoke run on the shipped 8x262k profile.

Profiles are unchanged. The performance tables stay as the drafter release's. Receipts are in results/2026-10-17-vision/.

🤖 Generated with Claude Code