AbstractVision v0.3.6

• Model registry/downloads: expand vision_model_capabilities.json and curated preset coverage for current Hugging Face text-to-image and image-to-image models, including FLUX.1/2, Qwen Image, ERNIE Image, GLM Image, Z-Image, SDXL, SD Turbo, and SD3.5. The registry now remains the packaged source of truth for both downloadable artifacts and task capability metadata.
• Cache/catalog: keep curated downloads in the Hugging Face cache by default, migrate older ~/models/<preset> trees into that cache on first use, and align Diffusers/MFLUX cache discovery across CLI, REPL, playground, and AbstractCore so cached local models surface consistently.
• API-level normalization: move model-specific parameter normalization into shared backend hooks used by VisionManager and the playground server. Model constraints such as MFLUX distilled guidance_scale=1.0, unsupported negative prompts, and GLM steps/guidance_scale/32-multiple dimensions now apply consistently through the CLI/REPL, playground, and AbstractCore integration.
• Playground/API: expose structured per-model task_specs from /v1/vision/models, only enable the edit surface for models that advertise image_to_image, switch models by auto-loading the new selection, and truncate b64_json in logs instead of dumping the full base64 payload.
• AbstractCore/provider catalogs: keep remote provider/model listing robust when internet or API credentials are unavailable, and keep local provider/model catalogs aligned with the same cache-backed discovery used by the interactive surfaces.
• CLI/docs: keep abstractvision cli as the canonical interactive command (with repl as a legacy alias), expand model catalog/docs coverage, refresh backend/configuration/AbstractCore/playground references, and regenerate llms.txt / llms-full.txt for release.