docs(i18n): rework lora hosting and stack mode hints

Restructure the seven hints to the settings pattern: lead definition,
tradeoff, scope, special values and stated default, with value and
cross-setting markup. Mode bullets and special values match loader
behavior: sub-8-bit hosting gate, rank 0 fallback, ramp 0 freezing
the balance.

ui/locale/locale_en.json

@@ -865,13 +865,13 @@
     {"id":"","label":"LoRA native fuse with model","localized":"","hint":"Merge LoRA into the model for lower memory usage.<br><br><b style=\"color: #ef4444\">Warning:</b> After removing or switching a LoRA, you may still see its style in generated images. To get a clean model, reload it from the model selector.","ui":"settings_lora"},
     {"id":"","label":"LoRA diffusers fuse with model","localized":"","hint":"Merge LoRA into the model for lower memory usage and torch.compile compatibility.<br><br><b style=\"color: #ef4444\">Warning:</b> After removing or switching a LoRA, you may still see its style in generated images. To get a clean model, reload it from the model selector.","ui":"settings_lora"},
     {"id":"","label":"LoRA quantized apply method","localized":"","hint":"How networks are applied to SDNQ-quantized model weights:<br>- <b>exact</b>: adapters are carried alongside the quantized weights at full precision; apply and removal are exact and the quantized weights are never modified. The carried factors take additional VRAM, growing with adapter rank, size and count<br>- <b>requantize</b>: adapters are merged into the quantized weights, matching the behavior of earlier releases. Uses no additional VRAM (a weight backup for network removal is held in system RAM); on models quantized below 8 bits rounding typically loses much of the adapter effect, with strong adapters retaining more<br><br>With <b>requantize</b> selected, the host rank, calibration and cache options below have no effect.<br><br>Default is <b>exact</b>.","ui":"settings_lora"},
-    {"id":"","label":"LoRA quantized host rank","localized":"","hint":"Maximum rank used to carry non-factorable adapter types (LoKR, LoHA, OFT, DoRA) on the side-channel of SDNQ models quantized below 8 bits, where merging would erase most of the adapter. Higher values keep more of the adapter at proportionally higher memory cost. Set to 0 to disable and merge into the quantized weights instead.","ui":"settings_lora"},
-    {"id":"","label":"LoRA quantized host calibration","localized":"","hint":"Collection of per-channel activation statistics during normal generations on models quantized below 8 bits, cached per checkpoint. When available, the statistics weight side-channel hosting of non-factorable adapter types toward the channels carrying the most activation energy, improving delivered fidelity at the same host rank.","ui":"settings_lora"},
-    {"id":"","label":"LoRA quantized host cache","localized":"","hint":"Disk budget in GB for caching the side-channel factors computed when hosting non-factorable adapter types on quantized models. A cached configuration skips the truncation math on reapply; least recently used entries are evicted past the budget. Set to 0 to disable.","ui":"settings_lora"},
-    {"id":"","label":"LoRA stack mode","localized":"","hint":"Combination rule applied when multiple LoRA networks target the same layer. sum adds all contributions. ties merges only elements whose signs agree after keeping the strongest ones; dare variants randomly drop elements and rescale the survivors; magnitude_prune keeps only the strongest elements of each network. klora and estlora select one of exactly two networks per layer, with the first network in the prompt treated as subject and the second as style, and the balance shifting from subject toward style over the sampling steps. Applied by the native loader only; other load methods behave as sum.","ui":"settings_lora"},
-    {"id":"","label":"LoRA stack density","localized":"","hint":"Fraction of elements kept by the ties, dare and magnitude_prune stack modes. Lower values keep only the strongest contributions of each network.","ui":"settings_lora"},
-    {"id":"","label":"LoRA stack ramp","localized":"","hint":"Steepness of the subject-to-style shift over the sampling steps for the klora and estlora stack modes. Higher values strengthen the late-step shift toward the style network.","ui":"settings_lora"},
-    {"id":"","label":"LoRA stack discrepancy","localized":"","hint":"Stand-in for the estlora mode's data-derived style separation estimate. Higher values shift the overall balance toward the subject network.","ui":"settings_lora"},
+    {"id":"","label":"LoRA quantized host rank","localized":"","hint":"Maximum rank used to carry adapter types that are not natively low-rank (LoKR, LoHA, OFT, DoRA) alongside the quantized weights instead of merging them in.<br>Higher values retain more of the adapter at proportionally more memory. Plain LoRA files are carried exactly at their own rank.<br><br>Applies only to SDNQ models quantized below 8 bits, where merging erases most of the adapter; at 8 bits and above merging retains it and hosting is skipped.<br><br><b>0</b> disables hosting and merges every adapter into the quantized weights.<br><br>Default is <b>256</b>.","ui":"settings_lora"},
+    {"id":"","label":"LoRA quantized host calibration","localized":"","hint":"Collects per-channel activation statistics from the model's own generations and uses them to focus hosted-adapter truncation on the channels with the strongest activations.<br>Statistics accumulate in the background on models quantized below 8 bits, persist per checkpoint, and raise delivered adapter fidelity at the same <b><i>LoRA quantized host rank</i></b>, most at low ranks.<br><br>Capture is skipped while the model is compiled; previously cached statistics still apply.<br><br>Enabled by default.","ui":"settings_lora"},
+    {"id":"","label":"LoRA quantized host cache","localized":"","hint":"Disk space in GB for caching computed hosting factors.<br>A cached set skips the truncation math on the next load; least recently used entries are evicted once the budget is exceeded.<br><br><b>0</b> disables the cache.<br><br>Default is <b>10</b>.","ui":"settings_lora"},
+    {"id":"","label":"LoRA stack mode","localized":"","hint":"How multiple networks targeting the same layer are combined:<br>- <b>sum</b>: adds all contributions<br>- <b>ties</b>: keeps each network's strongest elements and merges only where signs agree<br>- <b>dare_ties</b>: randomly drops elements, rescales the survivors, then merges where signs agree<br>- <b>dare_linear</b>: randomly drops elements, rescales the survivors and sums<br>- <b>magnitude_prune</b>: keeps each network's strongest elements and sums<br>- <b>klora</b> / <b>estlora</b>: assign each layer to one of exactly two networks, the first in the prompt as subject and the second as style, shifting from subject toward style over the sampling steps<br><br>Kept fractions are set by <b><i>LoRA stack density</i></b>; the subject-to-style shift by <b><i>LoRA stack ramp</i></b> and <b><i>LoRA stack discrepancy</i></b>.<br><br>Applies to the native load path; other load methods and text encoder networks always combine as <b>sum</b>. Selection modes fall back to <b>sum</b> unless exactly two networks are loaded, or when model compile is active.<br><br>Default is <b>sum</b>.","ui":"settings_lora"},
+    {"id":"","label":"LoRA stack density","localized":"","hint":"Fraction of elements each network keeps under the <b>ties</b>, <b>dare_ties</b>, <b>dare_linear</b> and <b>magnitude_prune</b> stack modes.<br>Lower values keep only the strongest contributions and reduce interference between networks at the cost of per-network detail. The dare variants drop at random and rescale the survivors to preserve expected strength.<br><br>Default is <b>0.5</b>.","ui":"settings_lora"},
+    {"id":"","label":"LoRA stack ramp","localized":"","hint":"Slope of the subject-to-style shift across the sampling steps in the <b>klora</b> and <b>estlora</b> stack modes.<br>Higher values shift layers to the style network earlier and more broadly; lower values keep the subject network dominant for longer.<br><br><b>0</b> keeps the balance fixed for the whole generation.<br><br>Default is <b>1.5</b>.","ui":"settings_lora"},
+    {"id":"","label":"LoRA stack discrepancy","localized":"","hint":"Stand-in for the measured style separation the <b>estlora</b> stack mode would otherwise derive from data.<br>Higher values keep layers with the subject network longer; lower values let the style network take layers earlier.<br><br>Applies only when <b><i>LoRA stack mode</i></b> is <b>estlora</b>.<br><br>Default is <b>0.5</b>.","ui":"settings_lora"},
     {"id":"","label":"LoRA auto-apply tags","localized":"","hint":"Automatically add trigger words/tags from LoRA metadata to your prompt.<br>Set to the number of tags to auto-apply, e.g., 3 = add top 3 trigger tags.<br>Set to 0 to disable, -1 to add all available tags.","ui":"settings_lora"},
     {"id":"","label":"LoRA memory cache","localized":"","hint":"How many LoRAs to keep in network for future use before requiring reloading from storage","ui":"settings_lora"},
     {"id":"","label":"LoRA add hash info to metadata","localized":"","hint":"Include LoRA file hashes in generated image metadata.<br>Useful for reproducibility and tracking which exact LoRA versions were used.","ui":"settings_lora"},