fix(sampler): H3 audit — context windows and ALG guide blur off and hidden

Both are structurally inert on H3; context windowing also mis-reported the
cause as an old ComfyUI. alg_anchor deliberately kept: a continuation scene
carries real latent frames for it to act on.

CHANGELOG.md

@@ -99,6 +99,12 @@
   looked exactly like one where the feature was switched off.
 
 ### Fixed
+- **Context windowing and ALG's guide blur are switched off and hidden on MiniMax H3.**
+  Core's windowing unpacks the LTXAV stream specifically and measures its window on LTX's 8x
+  latent ratio; ALG's guide blur acts on guide frames appended to the latent, and an H3 guide
+  is a condition row, so that tail is always empty. Context windowing also reported the wrong
+  reason on H3 — it blamed the ComfyUI version for what is a model difference.
+
 - **JoyAI-Echo is switched off and hidden on MiniMax H3, where it was doing damage.** The
   memory bank places frame *i* at sequence position *i*, but H3's packed layout pins only the
   first or last frame — so every frame past the first was refused, and the one that landed

movie_editor/frontend/pipeline_caps.js

@@ -128,6 +128,14 @@
     bounded_attention_enabled:
       "Bounded Attention masks text cross-attention; H3 puts text in the same self-attention "
       + "stream as the video, where the mask would be far too large to hold.",
+    context_windows:
+      "Core's context windowing unpacks the LTXAV stream specifically — mapping each video "
+      + "window onto its audio window and re-slicing the guide entries. H3 packs its sequence "
+      + "by a different layout, and the window length is measured on LTX's 8x latent ratio.",
+    alg_blur_guides:
+      "This blurs the trailing GUIDE frames appended to the latent. On H3 a guide is a "
+      + "condition row rather than an appended frame, so that tail is always empty. "
+      + "(alg_anchor is NOT hidden — a continuation scene does carry real latent frames.)",
     joyai_memory:
       "JoyAI-Echo places memory frame i at sequence position i. H3's packed layout pins only "
       + "the FIRST or LAST frame, so every frame past the first is refused — and the one that "
@@ -135,6 +143,13 @@
       + "one is not merely inert on H3, it is harmful, so the sampler forces it off.",
   };
 
+  // Sub-settings of the two groups above: they configure machinery that cannot run here.
+  const CONTEXT_SUB_INPUTS = [
+    "context_window_length", "context_window_overlap", "context_window_schedule",
+    "context_window_fuse", "context_window_freenoise", "context_window_retain_first",
+  ];
+  const ALG_GUIDE_SUB_INPUTS = ["alg_guide_blur_strength", "alg_guide_blur_sigma_threshold"];
+
   // JoyAI's own sub-settings: they configure a bank that cannot be built here.
   const JOYAI_SUB_INPUTS = [
     "joyai_memory_size", "joyai_fix_frames", "joyai_frame_select", "joyai_memory_strength",
@@ -182,7 +197,8 @@
     if (!usesChainSampler(st)) return new Set();
     if (!isH3(st)) return new Set(Object.keys(H3_ONLY_SAMPLER_INPUTS));
     return new Set([...Object.keys(H3_DEAD_SAMPLER_INPUTS), ...IDENTITY_SUB_INPUTS,
-                    ...JOYAI_SUB_INPUTS, ...H3_DEAD_VALUE_INPUTS]);
+                    ...JOYAI_SUB_INPUTS, ...CONTEXT_SUB_INPUTS, ...ALG_GUIDE_SUB_INPUTS,
+                    ...H3_DEAD_VALUE_INPUTS]);
   }
 
   // Returns [{short, detail}] for settings that are ON but cannot do anything on H3.

movie_editor/frontend/pipeline_caps.test.js

@@ -93,3 +93,44 @@ test("JoyAI stays available on LTX", () => {
     assert.ok(!inert.has(k), `${k} must remain reachable on LTX`);
   }
 });
+
+// ── the rest of the H3 audit ──────────────────────────────────────────────────
+
+test("context windowing and its settings disappear on H3", () => {
+  const inert = PC.familyInertInputs(h3);
+  for (const k of ["context_windows", "context_window_length", "context_window_overlap",
+                   "context_window_schedule", "context_window_fuse",
+                   "context_window_freenoise", "context_window_retain_first"]) {
+    assert.ok(inert.has(k), `${k} should be hidden on H3`);
+  }
+});
+
+test("ALG's GUIDE blur goes on H3 but its ANCHOR blur stays", () => {
+  // The guide blur acts on frames appended to the latent, and H3 guides are condition rows
+  // (tail always 0). The anchor blur acts on latent frame 0, which a continuation scene
+  // really does carry — hiding that one would neuter a knob that works.
+  const inert = PC.familyInertInputs(h3);
+  assert.ok(inert.has("alg_blur_guides"));
+  assert.ok(inert.has("alg_guide_blur_strength"));
+  assert.ok(inert.has("alg_guide_blur_sigma_threshold"));
+  assert.ok(!inert.has("alg_anchor"));
+  assert.ok(!inert.has("alg_anchor_strength"));
+});
+
+test("knobs that depend on the INSTALL, not the family, stay visible", () => {
+  // decode noise (which VAE), second pass / detailing (which upsampler) are reported at run
+  // time against the actual object. A family rule here would be a guess about a models folder.
+  const inert = PC.familyInertInputs(h3);
+  for (const k of ["decode_noise_scale", "decode_timestep", "second_pass_op",
+                   "segmented_detailing", "plateau_cache"]) {
+    assert.ok(!inert.has(k), `${k} is install-dependent and must not be hidden by family`);
+  }
+});
+
+test("LTX keeps everything H3 cannot run", () => {
+  const inert = PC.familyInertInputs(ltx);
+  for (const k of ["context_windows", "alg_blur_guides", "joyai_memory",
+                   "bounded_attention_enabled", "identity_transfer_enabled"]) {
+    assert.ok(!inert.has(k), `${k} must remain reachable on LTX`);
+  }
+});

samplers.py

@@ -6062,6 +6062,16 @@ class FunPackLTXAVSceneChainSampler:
         window the packed tensor blindly and quietly wreck audio and guides — the same class
         of packed-vs-unpacked mistake that killed DynamicConditioning. Refuse instead.
         """
+        if self._is_h3:
+            # Checked BEFORE the capability gate below, which would otherwise report a
+            # missing LTXAV core feature and blame the ComfyUI version for what is really a
+            # model difference. Core's windowing unpacks LTXAV's packed AV stream and
+            # re-slices its guide entries; H3's sequence is packed by a different layout
+            # entirely, and the latent window length here is computed on LTX's 8x temporal
+            # ratio besides.
+            return None, None, ("MiniMax H3 has no context-window support — core's windowing "
+                                "unpacks the LTXAV stream specifically, and the window length "
+                                "is measured on LTX's 8x latent ratio")
         try:
             import comfy.context_windows as _cw
             import comfy.patcher_extension as _pe
@@ -6680,6 +6690,15 @@ class FunPackLTXAVSceneChainSampler:
                              "against LTX's 4096-wide cross-attention context, and the overlap "
                              "tokens need LTX's patchifier; H3's native ref2va reference blocks "
                              "are the equivalent and are not wired yet)")
+            if context_windows:
+                _dead.append("context_windows (core's windowing unpacks the LTXAV stream and "
+                             "re-slices its guide entries; H3 packs its sequence differently, "
+                             "and the window length is measured on LTX's 8x latent ratio)")
+            if alg_blur_guides:
+                _dead.append("alg_blur_guides (it blurs the trailing GUIDE frames appended to "
+                             "the latent; on H3 a guide is a condition row, so the appended "
+                             "tail is always 0 and there is nothing to blur. alg_anchor still "
+                             "works — a continuation scene does carry real latent frames)")
             if joyai_memory:
                 _dead.append("joyai_memory / JoyAI-Echo (the memory bank places frame i at "
                              "sequence position i, and H3's packed layout pins only the FIRST "
@@ -6709,6 +6728,10 @@ class FunPackLTXAVSceneChainSampler:
             # an LTX project must not keep doing that here.
             joyai_memory = False
             joyai_audio_memory = False
+            # These two only ever no-op on H3 — off here so the run report does not list a
+            # mechanism that did nothing.
+            context_windows = False
+            alg_blur_guides = False
             v2a_grad_scale = 1.0
             # second_pass_op is NOT switched off here. It only needs a latent upsampler that
             # matches this model's latent width; which file that is depends on what is