[LauraGPT]

Summary

Publish the already merged PR #3831 offline FSMN-VAD multiple-input fix as standalone funasr-onnx 0.4.6. Each waveform is scored independently and receives its own result; requested batch sizes remain accepted while the effective model batch remains 1.

Only five established version/docs/release-contract paths change. All 13 runtime modules are unchanged from the reviewed and tested merge. Existing 0.4.5 and 0.4.4 history remains intact.

Validation

• Contract RED: 3 expected failures and 5 controls before the version/docs/workflow change; GREEN: 8 passed.
• PEP 517 sdist then wheel, Twine, isolated installation, installed version/origin, metadata, license, dependencies and all 13 runtime byte bindings passed.
• Reused by exact hash: PR3831's 74 scoped CPU cases and 42 real L20Z FP32 synthetic-contract cases (36 baseline failures, 0 candidate failures, 396 CUDA/0 CPU ORT nodes). No GPU rerun or new weight/data download.
• Exact-head and merged-main hosted CI, actual artifacts/logs, tag/release bytes and PyPI official bytes remain required before publication is complete.

This is not pretrained VAD/ASR quality, a full repository suite, Windows/NPU/BF16, or Nano acceptance. No protection change, self-approval or admin bypass.