[LauraGPT]

Summary

Preserve the original contributor commit from #3802 and complete its timestamp alignment contract on a maintainer-owned branch. The fork is left unchanged.

• Reindex timestamps by original token indices after filtering special tokens in all three implementations.
• Raise an explicit ValueError when a retained token has no timestamp, instead of silently returning shifted values. A trailing filtered EOS without a timestamp remains valid.
• Keep untimed transcription and filtered-only token inputs working, with exact timestamp assertions and explicit invalid-input regressions.

Validation

• Test-first check on fix(postprocess): keep word timestamps aligned after filtered tokens #3802: 2 failed, 6 passed; both failures are the missing explicit rejection.
• Updated source: 68 focused CPU tests passed across six postprocessing/timestamp suites.
• NVIDIA L20Z real CUDA CIF + timestamp helper: 16 FP32/BF16 pipeline/offset cases; all three postprocessors passed 48 exact-output checks and 48 malformed-input rejection checks, with CPU/CUDA agreement.
• These GPU inputs are synthetic. No pretrained ASR, ONNX engine, libtorch model inference, audio-quality evaluation, or full repository suite is claimed. The broad suite includes unapproved model downloads and has not been run.
• New exact-head hosted CI must finish before merge; previous fix(postprocess): keep word timestamps aligned after filtered tokens #3802 CI does not stand in for it.

This is separate from #3801's VAD chunk concatenation issue. It does not change or republish v1.4.17.