fix(amd): Qwen3.8-Flash-Next accepts image input

[MicroHEROX]

Corrects the AMD Token Factory entry for Qwen3.8-Flash-Next: the merged #5731 file declared this host text-only, but live testing on 2026-09-03 shows the AMD deployment accepts image input.

• GET /v1/models on the Token Factory API reports input_modalities: ["text", "image"] for this model.
• Image understanding verified live: a programmatically rendered test image with a random digit was read back correctly (HTTP 200 + correct answer). Control group: a text-only model on the same host (MiniCPM5-1B) rejects image input with 400, proving the gateway enforces per-model modality rather than swallowing image parts.
• Drop the attachment = false and text-only modality override; the multimodal alibaba/qwen3.8-flash-next base entry is now narrowed to text + image input (no video on this host). attachment inherits true from the base, matching the resolved DeepSeek-V4-Flash-Vision-Exp entry.
• Pricing ($0.15 / $0.47 / $0.016 per MTok) and reasoning controls (low|medium|xhigh, xhigh default, no none, no toggle, no reasoning_content) unchanged from feat: add Qwen3.8-Flash-Next to AMD provider #5731, re-verified live on 2026-09-03.

bun validate passes against current dev (resolved entry: attachment = true, modalities.input = ["text", "image"]).

[github-actions]

No actionable findings.