[LauraGPT]

Summary

Integrates @monsterlag's #3788 onto current main without changing the contributor's fork. Preserve the original optional speaker-import exception and raise an actionable chained ImportError only when speaker construction is requested. ASR-only use remains available. The original ten regression cases are unchanged; both NumPy CI matrices now select them and trigger on their path.

This is an error-reporting fix related to #2546, not evidence that its unidentified Windows dependency failure is resolved. That issue remains open. The original PR's old potential-merge tree did not include current main; it is not used as acceptance evidence.

Validation

• Current main reproduced eight expected NameError failures, with two healthy-dependency controls passing.
• 374 focused CPU tests passed with no failures or skips: all ten fresh-process import/configuration regressions, all 368 cases selected by the existing NumPy workflow, and six adjacent CPU-thread-setting cases. One extra performance diagnostic is explicitly deselected: test_auto_model_ncpu.py::test_the_cap_actually_reduces_blas_work segfaults in SciPy eigh after setting 192 threads on both untouched current main and this candidate. The initial combined 375-case process returned -11 and produced no completed XML; that failure is retained, not counted as a pass. This pre-existing native performance problem is not fixed here and is not part of the existing CI selection.
• Five fresh-process real L20Z FP32 cases exercised the actual AutoModel constructor and reduced native SANM/CIF inference on two synthetic audio inputs. Four injected optional-import failures preserve their exact cause while ASR still runs on CUDA. The healthy case builds the actual CAM++ model and backend and extracts CUDA speaker embeddings from audio using previously saved synthetic parameters. CUDA profiler evidence, exact inputs/weights/RNG, raw outputs, dependencies, code hashes and exit receipts are retained on CPFS. No pretrained weights were downloaded.
• Synthetic execution does not establish pretrained ASR/diarization quality, BF16, Windows compatibility, the complete repository suite or Nano leaderboard results. The broad suite's existing unavailable model/hub integration limitations remain; those completed failures were not rerun.
• New-head and merged-main CI require their own actual logs and both NumPy artifacts containing all ten regressions. No release/PyPI upload, review fabrication, administrator bypass or protection change is included.

Original contribution is credited in the signed+DCO commit. #3788 will be closed as superseded only after final integration acceptance, not described as a merged original commit.