Correct the pc_features axis order and indexing example

docs/changelog.md

@@ -111,6 +111,10 @@ behavior they verify rather than listed separately.
 - Added a task-oriented documentation navigation, synchronized API, shortcut,
   and plugin references, and strict documentation checks in continuous
   integration.
+- Corrected the `pc_features.npy` description. The file stores
+  `(n_spikes, n_pcs, n_channels_loc)`, phy swaps the last two axes on load,
+  array indices are zero-based, and `pc_feature_ind` is indexed by the
+  spike's template.
 
 ## [2.1.0] — 2026-07-17
 

docs/dataset.md

@@ -101,7 +101,7 @@ Most real sorting exports should also contain the following files:
 | `similar_templates.npy` | `(n_templates, n_templates)` similarity matrix used by the Template GUI's default similarity ranking. Larger values rank first; the score's scale is defined by the exporter. |
 | `whitening_mat.npy` | `(n_channels, n_channels)` whitening matrix. |
 | `whitening_mat_inv.npy` | `(n_channels, n_channels)` inverse whitening matrix used to display unwhitened templates. If absent, phy computes and writes it. |
-| `pc_features.npy` | `(n_spikes_subset, n_channels_loc, n_pcs)` floating-point principal-component features used by the Feature View. |
+| `pc_features.npy` | `(n_spikes_subset, n_pcs, n_channels_loc)` floating-point principal-component features used by the Feature View. phy swaps the last two axes on load, so `model.get_features()` returns `(n_spikes, n_channels, n_pcs)`. |
 | `pc_feature_ind.npy` | `(n_templates, n_channels_loc)` channel indices corresponding to PC features. |
 | `template_features.npy` | `(n_spikes_subset, n_template_features)` projections onto nearby templates. |
 | `template_feature_ind.npy` | `(n_templates, n_template_features)` template indices corresponding to template features. |

docs/sorting_user_guide.md

@@ -63,7 +63,7 @@ _______________
     * `amplitudes.npy` - `[nSpikes, ] double` vector with the amplitude scaling factor that was applied to the template when extracting that spike
     * `channel_map.npy` - `[nChannels, ] int32` vector with the channel map, i.e. which row of the data file to look in for the channel in question
     * `channel_positions.npy` - `[nChannels, 2] double` matrix with each row giving the x and y coordinates of that channel. Together with the channel map, this determines how waveforms will be plotted in WaveformView (see below).
-    * `pc_features.npy` - `[nSpikes, nFeaturesPerChannel, nPCFeatures] single` matrix giving the PC values for each spike. The channels that those features came from are specified in pc_features_ind.npy. E.g. the value at `pc_features[123, 1, 5]` is the projection of the 123rd spike onto the 1st PC on the channel given by `pc_feature_ind[5]`.
+    * `pc_features.npy` - `[nSpikes, nFeaturesPerChannel, nPCFeatures] single` matrix giving the PC values for each spike. The channels that those features came from are specified in `pc_feature_ind.npy`. All indices are zero-based, so the value at `pc_features[123, 1, 5]` is the projection of the spike with index 123 onto the second PC, on the channel given by `pc_feature_ind[t, 5]`, where `t` is that spike's template, `spike_templates[123]`. phy swaps the last two axes when it loads the file, so `model.get_features()` returns `[nSpikes, nChannels, nFeaturesPerChannel]`.
     * `pc_feature_ind.npy` - `[nTemplates, nPCFeatures] uint32` matrix specifying which pcFeatures are included in the pc_features matrix.
     * `similar_templates.npy` - `[nTemplates, nTemplates] single` matrix giving the similarity score (larger is more similar) between each pair of templates
     * `spike_templates.npy` - `[nSpikes, ] uint32` vector specifying the identity of the template that was used to extract each spike