[LauraGPT]

Summary

Fix the default WavFrontendOnline(lfr_m=1, lfr_n=1) path, which references an uninitialized lfr_splice_frame_idx after skipping LFR. Initialize the consumed-frame cursor from the actual per-item feature length, including zero-length items; the LFR branch continues to replace it with its existing splice cursor.

Real GPU integration exposed a second issue: final offline audio whose length is an exact multiple of the VAD chunk size gets an unnecessary empty chunk. With no LFR right-context frames, that empty call cannot finalize the last score and can discard a pending dynamic-VAD segment. Mark the last nonempty offline chunk final by adding a remainder chunk only when one exists.

Add 42 regressions covering valid lengths, padding, CMVN, packetization, empty frontend flushes, short inputs, and offline finalization with no-LFR/LFR controls. Run them in the existing KWS/frontend workflow and include the VAD model in its path filters.

Validation

• Unchanged frontend: 24 UnboundLocalError failures and 6 passing short-input controls.
• The first waveform reference omitted the frontend's existing 32768 PCM scaling. That test-fixture error and its 16 failures were preserved; after correcting the reference without changing tolerances, all 16 waveform cases still failed on unchanged product code and all 30 frontend cases passed on the repair.
• The new offline-finalization regression reproduced 5 failures with 7 passing controls before the chunk-count fix.
• Final candidate: all 214 selected CPU cases passed, including KWS, frontend, VAD, realtime websocket and example contracts.
• Real NVIDIA L20Z: 16 direct FP32/BF16 CUDA feature cases passed. Initial VAD integration retained 6 fixed-threshold passes and 6 dynamic-invariance failures. After the finalization fix, all 12 VAD integration cases passed, with 26,200 CUDA profiler events; the unchanged 16 feature cases were source-hash bound and reused, not rerun.
• The VAD integration uses real feature extraction and VAD state handling with an analytic CUDA posterior encoder. No pretrained weights/downloads, ASR quality claim, or complete-suite claim.
• Exact-head hosted CI is required before normal merge; merged-main CI is accepted separately.

Scope

No model weights, dependencies, release versions, repository protections, or production configuration change. The generic explicit streaming-empty-input early-return path is unchanged; this PR does not claim to solve every empty-EOS or short-tail case. Root 1.4.22 is already published and does not contain this candidate. The ONNX offline VAD work in #3846 is separate.

Type Of Change

• Bug fix
• Regression tests
• Model/training change