[LauraGPT]

Summary

Address #3801: Chinese VAD chunk boundaries must stay separated in the timestamp-aligned token stream. Punctuation input and top-level punctuated text retain their existing compact Chinese surface.

• Derive fallback timestamp text from the already space-joined raw ASR text, removing rich tags without fusing adjacent chunk tokens.
• Keep explicit words metadata, subword/surface alignment, no-punctuation and English behavior unchanged.
• Exercise the real AutoModel VAD merge, duration sorting/restoration, timestamp sentence helpers and speaker assignment with deterministic model-response fixtures. Check exact sentence text, every timestamp, speaker IDs, raw-text output and mixed/English controls.

Validation

• Before the fix: 3 Chinese regressions failed with the reported shifted text; 2 language controls passed. An initial mixed-language expectation was corrected to preserve the existing leading-space behavior; that initial negative receipt is retained.
• Focused CPU regression suites: 207 tests plus 171 related KWS/VAD/speaker/MOSS tests passed, zero skips/failures/errors. Dependency/deprecation warnings remain in the logs. A missing pytorch-wpe dependency was installed only in the maintenance virtual environment; both failed environment runs were retained. This is not the full repository suite, which includes unapproved model downloads.
• Real NVIDIA L20Z: 6 FP32/BF16 CUDA CIF timestamp pipelines matched CPU and explicit timestamp oracles; 8 downstream VAD merge/speaker/raw-text cases passed.
• The GPU inputs and model responses are synthetic fixtures. No pretrained ASR, original recording, transcription quality, or public benchmark validation is claimed.
• New exact-head hosted CI is required before normal merge. No branch protection change or administrator bypass.

This is a post-v1.4.17 fix, not a republish. Keep #3801 open for verification on the reporter's original recording.