v1.8.1

AXQuant 1.8.1

Added

• Capability-gated multimodal certification policy (Certification Spec §8): unsupported
  vision/audio modalities are not-applicable; supported modalities require
  present-not-certified, smoke-certified, or quality-certified with bound evidence.
  Helpers in axquant.modality_certification; optional modalities block on public Tier 1
  certificates (all in-tree certs backfilled).
• Multimodal recertification evidence for the 1.8.x policy pass:
  docs/certifications/evidence/modality-recert-1.8.1/REPORT.json.

Changed

• Public Tier 1 records now state vision/audio claims explicitly instead of free-text only.
• Hybrid Qwen 3.6/3.8 packs with vision.safetensors remain present-not-certified: stock
  MLX-VLM smoke fails on model.visual.* layout (expects vision_tower.*); text dual-suite
  Tier 1 and BF16 protection are unchanged. Qwen3-VL keeps prior smoke-certified status.

Operator action

• Install with python -m pip install 'axquant[mlx]==1.8.1'.
• New certificates should set modalities via build_modalities_block(...).
• Do not claim VLM quality from text Tier 1 or from weight presence alone.