[phoenixray2000]

Automatic speaker counting switches from spectral clustering to UMAP/HDBSCAN at 2048 speaker embeddings. The density-based branch ignores the spectral speaker limit, so a configured maximum of 5 can still produce 8 speaker groups.

Expose a construction-time top-level AutoModel(..., max_num_spks=5) argument, defaulting to 15. Forward that value explicitly to ClusterBackend so both automatic clustering paths obey the same cap. max_num_spks inside spk_kwargs or spk_kwargs["cb_kwargs"] raises TypeError, including when a top-level value is also supplied. This leaves one AutoModel configuration entry point without alias or precedence rules. Other backend options, such as merge_thr, continue to use the existing component configuration.

Excess UMAP/HDBSCAN clusters are greedily merged by centroid cosine similarity. Noise labels are preserved and excluded from the count; pair selection excludes the diagonal and handles nonpositive similarities and zero centroids. An explicit preset_spk_num/oracle_num takes precedence on the known-count spectral and KMeans paths.

The default limit of 15 now also applies to automatic UMAP/HDBSCAN results. Callers can increase it for larger meetings. This is an upper bound, not an exact count: setting it below the actual number of speakers can merge different people. The English and Chinese SDK guides document the top-level construction argument, defaults, validation and examples.

Validation:

• The 2048-embedding limit regression fails against the original implementation for caps 1, 3 and 5 with NumPy and Torch inputs. Top-level routing and nested-location rejection regressions also fail before the API change; all pass afterwards.
• Core clustering and SDK documentation checks: 66 passed, 22 subtests passed. Coverage includes top-level caps 1/3/16, the default of 15, invalid values, both nested locations alone or alongside equal/conflicting top-level values, preservation of other backend options and caller dictionaries, known-count precedence, the 2047/2048 boundary, noise, closest-centroid selection and zero/nonpositive similarities. The existing real UMAP/HDBSCAN wrapper test also passes.
• NumPy workflow selection plus SDK, component-device and timestamp regressions: 284 passed, 42 subtests passed, 1 deselected locally. The deselected MP3 decoder case reopens a still-open NamedTemporaryFile on Windows; it remains selected in Linux CI.
• Black 24.4.0 with the repository's 100-character line length and git diff --check pass.
• tests/test_cluster_backend.py is included in the path filters and test selection of the NumPy 1.26.4 / 2.4.0 CI matrix.

The existing product-site browser environment has a separate incompatibility: its ubuntu-latest runner uses Ubuntu 26.04 with Playwright 1.55.0. The previous browser run fails at Install Chromium, before browser tests run, with Error: ERROR: Playwright does not support chromium on ubuntu26.04-x64.

Latest head 7bec62b9763bee22c513a156ddfcc35c55d91774: NumPy 1.26.4, NumPy 2.4.0, KWS output, model contracts, and legacy document links pass (NumPy, KWS, model contracts). The current product-site build/validation is still in progress; its browser stage is pending.

[LauraGPT]

Thanks @phoenixray2000. Your seven reviewed speaker-cap changes are now integrated through #3809 on current main, with contributor credit in the signed+DCO commit. The original fork/head is unchanged. This PR is being closed as superseded, not described as a merged original commit.

Validation: the previous source reproduces 41 expected failures, while the candidate passes 66 focused CPU tests with no skips. Real L20Z execution used random-initialized reduced CAMPPlus parameters, FP32, batch 32, and 2048 unique synthetic feature rows. Actual CPU UMAP/HDBSCAN and spectral clustering exercised the cap; the CUDA profiler recorded 972 kernel events in its first batch. This is not pretrained diarization accuracy, BF16 or a full-suite claim.

The old browser failure was not bypassed or rerun. The integrated head and final main each passed their new CI, including browser/build validation and both NumPy versions. Actual evidence artifacts and logs were downloaded and checked against each exact commit, including all 50 cluster tests and six existing empty-batch regressions per NumPy matrix. Completed local experiments were reused only after full source/receipt hash verification. No new package release is claimed.