[Timothysit]

Prepares the Python pipeline for spike-sorted data with hundreds of units. On the 306-unit sorted recording (steps 1–4, full figures, one recording):

	Before	After
Step 3 (STTC + 200 shuffles × 3 lags)	1,494 s	121 s
Step 4	1,749 s	332 s
Total	3,258 s	463 s

Adjacency matrices and the network CSVs are identical to the old run.

Changes

• STTC (sttc.py): numba prange across rows, paired from both ends to balance load: 2.2 s → 0.22 s per matrix. If numba's threading layer is missing (e.g. a frozen app), it warns and runs serially with the same results. The build spec now collects numba.np.ufunc, and app_smoke checks it.
• NMF (nmf.py): 933 s → 26 s.
  • Fits run on threads. Seeds are drawn in the original order and BLAS is pinned to 1 thread, so results don't depend on thread count.
  • The rank sweep stops at the 95% crossing, or when the curve levels off under the new nmf_plateau_gain (default 0.005; 0 = MATLAB's full sweep). Skipped ranks are NaN.
  • The setting has a widget under Advanced in the Network metrics box on the Connectivity tab and is read by both pipelines.
• Normalised participation coefficient (null_models.py, network_metrics.py): 379 s → 33 s. All random draws for the 100 null models depend only on the original edge counts, so they are planned up front in order and the rewiring then runs on threads with identical results.
• randmio/latmio v2 and randmio_und_signed: compiled with numba and consume the same RNG stream as before.

Caveats

• NMF arrays differ from old runs by about 1e-14, because of the single-thread BLAS pin. Old runs already varied by this much between machines.
• On sorted data the variance curve rises roughly linearly and never plateaus, so in practice the sweep stops at 95% (rank 228/306 here).
• Figures are now most of the remaining time: about 250 s of step 4, of which plot_network takes about 190 s.
• The frozen-app threading layer still needs to be confirmed on a real build.
• This PR bumps to 1.15.0 / CAT-NAP 1.11.0, the same bump as the spatiotemporal-autocorrelation branch. Whichever merges second must re-bump and update the section count in test_gui_advanced.

Tests

• test_pipeline_fast_paths: 91/91. It compares each rewritten piece against the replaced code (including RNG state) at 1 and many threads.
• MATLAB parity tests for steps 3 and 4 pass (step 4: 60/60).
• test_step4_optional_metrics 21/21; CAT-NAP 112/112; catnap_subnetwork 45/45; sorted_import, null_models, nmf and small_worldness pass.
• Full suite: the only failures are test_pipeline_streaming_pool, test_viewer_server and test_stats_report, which also fail on main.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LBjNAqNrYVFQSwnaKgi25j