[LauraGPT]

Summary

Integrate the original #3882 hotword fix on the current accepted 1.4.26 main, retaining the original author's commit as the parent. Relative to current main this changes only the original three files: the ONNX and Torch hotword preprocessing and its native-runtime tests. The original fork branch is untouched. Refs #3881.

The exported embedding input is ten characters wide. The fix keeps the supported prefix of longer words, computes positions after truncation, and drops empty entries produced by whitespace.

Verification

• Four author regression cases fail on baseline; the current-main CLI/native CPU set passes all 63 cases without skips.
• Real L20Z FP32 validation passed 24 combinations (ONNX/Torch, Contextual/Seaco, batches 1/2, long/empty/whitespace hotwords), producing 72 recorded outputs from three distinct synthetic audio files. Torch recorded 856 CUDA events, and eight ONNX session profiles confirmed CUDA node execution.
• The four baseline long-input errors occur in CPU preprocessing before GPU graph execution; they are not counted as GPU inferences.
• These are fixed synthetic graph fixtures, not pretrained ASR-quality, full-suite or Nano results. Torch fixtures are numerically identical GPU-serialized copies, not a fix or claim for automatic device placement.
• Completed tests are reused through exact source hashes. Only eight release version/docs/test files differ from the originally tested tree; all runtime and fixture sources are byte-identical. No unchanged tests were rerun to inflate coverage.
• The original head's three workflows were accepted from actual artifacts/logs; its ONNX checkout used an older PR merge. Those records are historical evidence, not acceptance of this new head. New-head hosted CI and current protected-branch gates remain required before merging.

This does not publish standalone ONNX or Torch distributions, and the already published root 1.4.26 does not contain this later runtime change. No production service or model-weight download was performed.