[LauraGPT]

Summary

Integrates #3854 while preserving its original commit and authorship. Empty CIF samples have no output frame to reset; filtering their reset offsets avoids out-of-bounds indexing when a mixed batch ends in zero-fire samples. Both eager and export functions are covered. Existing rounding and thresholds are unchanged.

Adds read-only exact-head CPU CI with required ONNX dependencies and a zero-skip assertion. The signed companion commit synchronizes retained main content without rewriting contributor history.

Maintainer validation

• Same submitted suite on unchanged main: 45 errors out of 98 subcases across 6 methods. Candidate: 98/98 subcases, 6/6 methods, zero skips, including 14 dynamic ONNX Runtime CPU batch patterns.
• Actual NVIDIA L20Z: 34 FP32 analytic cases, 2551 CUDA kernels, independent sequential frame oracle, finite gradients, and real CifPredictorV2/export callers. Exact outputs, checkpoint hash, trace and source hashes are retained.
• The intended pytest CI command additionally passed all 6 methods with zero skips. Workflow structure and shell syntax validated; new exact-head hosted CI remains required before merge.
• Private HOME/cache and kernel network isolation throughout. No new pretrained weights, training, production changes, BF16 claim, full Paraformer export, recognition-quality benchmark or full-suite claim.

This is not a new release. Nano remains deferred.