[LauraGPT]

Summary

Release the already-merged standalone Torch runtime fixes as 0.1.7: mixed-length Paraformer timestamps (#3871), long/empty Contextual and Seaco hotwords (#3882), terminal abbreviations (#3869), and rich-transcription word boundaries. Keep root FunASR, ONNX and Python WebSocket server releases separate.

Add installed wheel and source-distribution regressions to the existing Python 3.10/3.12 packaging workflow, preserving read-only permissions and unrelated-engine absence checks. Numeric hotword weights are not added; long terms retain the first ten characters supported by the exported embedding.

Verification

• 118 related source/package/native-wrapper/postprocessing/CMVN CPU tests pass, with zero failures or skips.
• Wheel and sdist installed outside the source package each pass five regressions. The official 0.1.6 wheel reproduces all four behavior failures and the expected old-version mismatch.
• The first private negative harness had two setup mistakes (a non-exported alias and missing synthetic device field). Their failures are retained, not counted as product regressions; only the two affected cases were rerun after correction, and both then failed for the intended runtime defect.
• All ten runtime Python modules are byte-identical to the previously accepted real Shanghai L20Z fixture source. Existing GPU evidence is hash-bound, not rerun: no new pretrained weights or model-quality/full-suite claims.
• New-head and merged-main hosted logs, checkout trees, JUnit and actual archives require separate acceptance. Formal GitHub Release and PyPI publication follow only after those gates.

This is a standalone runtime package release, not a production model-service upgrade.