[LauraGPT]

Summary

Integrate #3880 on current main, preserving its original author commit and hotword normalization. Fixes #3879. Weighted JSON object strings now reach the Python WebSocket model as words; AutoModel does not accept numeric hotword weights, so values remain intentionally omitted. Plain strings and other non-object values are unchanged. The already-merged offline model-selection change is retained.

Extend the existing Python 3.10/3.11/3.12 WebSocket workflow to execute hotword tests with actual scipy/websockets/CPU torch dependencies. No skipped-test acceptance.

Validation

• Current integrated source: 40 selected CPU tests passed, zero skips/errors. The prior original-patch negative control (3 failures/2 passes) is preserved, not rerun.
• Real Shanghai L20Z FP32: ws_serve coroutine -> real source AutoModel/local state loading -> native Seaco and Contextual hotword tokenization -> CUDA kernels. Fifteen combinations of offline/online/2pass and five hotword formats, with two distinct synthetic PCM inputs; baseline nine failures/six passes, candidate fifteen passes. Raw outputs and CUDA traces retained.
• The first private GPU harness failed before any cases because its local config lacked tokenizer_conf. That failure is retained; the corrected isolated fixture uses the same existing synthetic weights, inputs and literal expected values.
• This validates protocol/model input and CUDA execution contracts, not pretrained ASR recognition quality, real network transport, arbitrary model compatibility, NPU, Nano or the full repository suite. No new pretrained weights downloaded.
• Actual new-head/main CI checkout and artifact bytes must pass before completion. No self-approval, protection changes or admin bypass.

This server-script fix is not part of the separately versioned funasr-torch package.