fix header sub-text/links contrast in light mode

cap video duration at 3.5s

set social thumbnail in card metadata

add social thumbnail image

enable Gradio MCP server (mcp_server=True) + tool docstrings/type hints

perf: optional First-Block-Cache "Faster" toggle (~1.4x, A/B-verified near-clean); default off

perf: guidance-interval CFG (skip uncond forward on low-noise tail, stop=0.7) — ~14% faster, quality-neutral (A/B verified)

UI: strip gradio link padding (fixes fake space inside (Qwen3.6-27B))

UI: zero header wrapper padding, drop bottom border

UI: borderless header flush with container, sub-text full width (2 lines)

UI: left-align header links (clear HF widget), I2V examples as labeled chips (fix empty prompt column)

UI: compact sleek header, clean single-column examples (natural-language prompts + labels), styled scrollbars

revert batched-CFG attempt: FA3 has no sm_120 build; drop kernels dep (broke diffusers import)

Delete files fa3_shim.py with huggingface_hub

fix: lazy fa3_shim instead of top-level flash_attn_interface module

Delete files flash_attn_interface.py with huggingface_hub

batched CFG via FA3 kernel (kernels-community) with sequential fallback

calibrate GPU duration estimator from measurements, cap duration at 4s, README polish

fix Textbox kwarg for gradio 6, remove diagnostics endpoint

add official two-step prompt rewriter (Qwen3.6-27B via Inference Providers), TF32

diagnostics: add weight integrity check

diagnostics v2: component isolation

add backend diagnostics endpoint

init: LingBot-Video MoE 30B-A3B ZeroGPU demo

initial commit