docs(settings): record the head sharing finding in the sparse hints

Per-head selection is what breaks head-consistency sensitive models, measured on Anima across fourteen arms: the only clean sparse selections are the ones every head shares. The head sharing hint names itself as the first resort for banded output, and the exclusion hint carries why Cosmos stays listed.

ui/locale/locale_en.json

@@ -1476,8 +1476,8 @@
     {"id":"","label":"Sparse Attention minimum sequence","localized":"","hint":"Shortest sequence that is sparsified. Below it attention stays dense, because choosing the tiles costs more than skipping them saves.<br>Sequence length is not resolution: a diffusion transformer sees roughly (width/16) x (height/16) tokens for an image, so 1024x1024 is about 4k tokens and 2048x2048 about 16k, and video multiplies that by the frame count.<br><br><b>0</b> sparsifies every sequence that reaches the stage.<br>Default 8192, around 1450x1450 for an image.","ui":"settings_cuda"},
     {"id":"","label":"Sparse Attention dense steps","localized":"","hint":"Number of steps at the start and end of sampling that receive a larger budget, where composition and fine detail are set. Counted at each end, and capped at half the run.<br><br>Only takes effect when <b><i>Sparse Attention dense step bonus</i></b> is above 0.<br><b>0</b> applies one budget to every step.<br>Default 0.","ui":"settings_cuda"},
     {"id":"","label":"Sparse Attention dense step bonus","localized":"","hint":"Percentage points added to the budget on the dense steps, capped at 100.<br><br>Only takes effect when <b><i>Sparse Attention dense steps</i></b> is above 0.<br>Default 30, so a budget of 30 rises to 60 on those steps.","ui":"settings_cuda"},
-    {"id":"","label":"Sparse Attention share selection across heads","localized":"","hint":"Computes one selection for all attention heads rather than one per head, by averaging the heads before scoring. Cheaper to select and coarser in what it keeps, since heads that attend to different regions are served by a single compromise.<br>Worth trying when selection is itself a visible share of the cost, which happens on models with many heads.<br><br>Disabled by default.","ui":"settings_cuda"},
-    {"id":"","label":"Sparse Attention excluded models","localized":"","hint":"Models that stay dense no matter how the rest of this section is set. Comma separated, matched case insensitively against the architecture, the pipeline class and the denoiser class, so whichever of those names is to hand works as an entry.<br>The class names appear in the model load log; the architecture is the short name used elsewhere in the settings, such as <b>f1</b> or <b>anima</b>.<br><br>Listed by default is <b>CosmosTransformer3DModel</b>, the transformer Anima runs, which returns banded noise at every budget tested against a sound dense baseline. The class is listed rather than the architecture because the other models built on it have not been checked.<br>A model whose output breaks up rather than merely softening belongs here.<br><br>Default <b>CosmosTransformer3DModel</b>.","ui":"settings_cuda"}
+    {"id":"","label":"Sparse Attention share selection across heads","localized":"","hint":"Computes one selection for all attention heads rather than one per head, by averaging the heads before scoring. Cheaper to select and coarser in what it keeps, since heads that attend to different regions are served by a single compromise.<br>Worth trying when selection is itself a visible share of the cost, which happens on models with many heads.<br>Also the first thing to try when a model breaks up into bands under sparse attention: some architectures need every head to see one consistent context, and per-head selection is what breaks them.<br><br>Disabled by default.","ui":"settings_cuda"},
+    {"id":"","label":"Sparse Attention excluded models","localized":"","hint":"Models that stay dense no matter how the rest of this section is set. Comma separated, matched case insensitively against the architecture, the pipeline class and the denoiser class, so whichever of those names is to hand works as an entry.<br>The class names appear in the model load log; the architecture is the short name used elsewhere in the settings, such as <b>f1</b> or <b>anima</b>.<br><br>Listed by default is <b>CosmosTransformer3DModel</b>, the transformer Anima runs, which collapses into banded noise when each head selects its own tiles and stays degraded even with <b><i>Sparse Attention share selection across heads</i></b> enabled. The class is listed rather than the architecture because the other models built on it have not been checked.<br>A model whose output breaks up rather than merely softening belongs here.<br><br>Default <b>CosmosTransformer3DModel</b>.","ui":"settings_cuda"}
   ],
   "t": [
     {"id":"txt2img_nav","label":"T2I","localized":"","hint":"Create image from text<br>Legacy interface that mimics original text-to-image interface and behavior"},