fix: specialist-suite sweep — surface silent node divergences

README.md

@@ -6,6 +6,10 @@ A visual prompt manager + detailer + sampler suite for ComfyUI (formerly *Ribbit
 
 **~20 nodes**, one shared library, **zero ComfyUI custom-node dependencies** — every node imports only PyPI packages (torch, ultralytics, sam2, etc.) and ComfyUI core. Nodes that can leverage IPAdapter Plus (Character Anchor) lazy-import it at call time so the rest of the suite loads cleanly without it.
 
+## Specialist sweep — 2026-07-28
+
+`PromptLibraryMulti` IS_CHANGED now matches load_prompts (a deleted entry made the cache key disagree with the emitted text → stale/wrong output); conditioning-concat shape-mismatch warns instead of silently dropping the style; civitai model-override + LoRA-chain-truncation now warn (metadata was silently wrong/incomplete); `_to_latin1_safe` type fixed; +1 test.
+
 Highlights:
 - **Library / Style / Multi** — visual prompt picker with per-entry LoRA stacks; selection persists across page refreshes via three-tier mirroring (widget value → node properties → localStorage)
 - **Smart Detailer** — one node replaces the 3-node FaceDetailer chain. 6 detail targets (face / eyes / mouth / hands / feet / skin), SAM mask refinement, per-target threshold/denoise/max/steps/crop_factor overrides, color-coded detection-preview output. **No Impact Pack required** — uses `ultralytics` directly.
@@ -346,6 +350,12 @@ Going forward, new widgets land in `optional` so they don't shift required-widge
 
 The detail pass pipeline used to skip a NaN scrub between VAE decode and the alpha-blend composite, so a sample-time NaN propagated into the final image and downstream `clip(0,1).astype(uint8)` casts produced black pixels. Fixed in v0.49.0 — `nan_to_num` runs after every refined-image path. If you still see this on v0.49.0+, share the console log; there's a different bug.
 
+## Maintenance — periodic specialist sweeps
+
+Last full code-specialist sweep: **2026-07-28** (see the `sweep/specialist-suite-fixes` branch / PR).
+
+A specialist sweep is a multi-axis review — bug/logic review, silent-failure hunt, type-design, comment-accuracy, test-coverage, and blind-spot passes — with every fix adversarially re-verified before it lands. These passes catch the quiet correctness/robustness bugs (swallowed errors, unrecoverable saves, non-atomic writes, silent divergences) that accumulate between feature work and never show up in a demo. **Worth re-running every few months, or after any large feature push.** Local findings archive: `~/vault/code-sweep/`.
+
 ## Credits
 
 See [CREDITS.md](CREDITS.md) for the full list. Short version: the suite is

__init__.py

@@ -781,7 +781,7 @@ def IS_CHANGED(cls, **kwargs):
         for n in range(1, cls.PANELS + 1):
             ids = [p.strip() for p in (kwargs.get(f"prompt_id_{n}") or "").split(",") if p.strip()]
             sep = kwargs.get(f"separator_{n}", ", ")
-            sigs.append(sep.join(items.get(pid, "") for pid in ids))
+            sigs.append(sep.join(items[pid] for pid in ids if pid in items))
         return "
".join(sigs)
 
     def load_prompts(self, **kwargs):

civitai_save.py

@@ -826,6 +826,11 @@ def _walk_model_chain(prompt: dict, start_node_id: str | None):
 
         # Unknown node — give the `model` input one last try (covers most
         # third-party passthroughs we haven't catalogued).
+        if ctype and ctype not in {*_LORA_LOADER_TYPES, _RGTHREE_LORA_STACK_TYPE,
+                                    _RGTHREE_POWER_LORA_TYPE, *_MODEL_LOADER_TYPES,
+                                    *_SDXL_TUPLE_PACK_TYPES, *_MODEL_PASSTHROUGH_TYPES}:
+            print(f"[CivitaiSave] LoRA chain walk: unknown node type {ctype!r}; "
+                  f"LoRA collection may be incomplete")
         current = _link_source(inputs.get("model"))
     return None, list(reversed(loras))
 
@@ -1151,8 +1156,6 @@ def extract_workflow_metadata(prompt: dict | None) -> dict:
 
 
 def _to_latin1_safe(s: str) -> str:
-    if s is None:
-        return s
     out = s.translate({ord(k): v for k, v in _LATIN1_FALLBACKS.items()})
     try:
         out.encode("latin-1")
@@ -1413,10 +1416,10 @@ def save(self, images, filename_prefix,
         )
         model_resolved = resolve_model_path(model_label) if model_label else None
         if model_label and not model_resolved:
+            actual_model = meta.get("model_label")
             print(f"[CivitaiSave] model_override {model_label!r} doesn't resolve "
-                  f"to a known checkpoint — falling back to auto-detected "
-                  f"{meta.get('model_label')!r}")
-            model_label = meta.get("model_label")
+                  f"to a known checkpoint; metadata will use auto-detected {actual_model!r}")
+            model_label = actual_model
             model_resolved = resolve_model_path(model_label) if model_label else None
         model_name = model_resolved[0] if model_resolved else None
         model_sha = (get_cached_sha256(model_label, model_resolved[1])

style_node.py

@@ -121,9 +121,12 @@ def _concat_conditioning(base, addition):
         t1 = entry[0]
         try:
             merged = torch.cat((t1, cond_from), dim=1)
-        except Exception:
+        except Exception as e:
             # Shape mismatch (mixed model classes, etc.) — keep the base
             # entry untouched rather than crashing the workflow.
+            _log.warning("PromptLibraryStyle: conditioning concat failed (base shape %s vs "
+                        "addition shape %s); style conditioning not applied: %s",
+                        tuple(t1.shape), tuple(cond_from.shape), e)
             merged = t1
         out.append([merged, entry[1].copy()])
     return out

tests/test_prompt_library.py

@@ -1120,6 +1120,40 @@ def test_multi_node_skips_missing_ids(self):
         out = node.load_prompts(prompt_id_1="a,nope", prompt_id_2="", prompt_id_3="")
         self.assertEqual(out[0], "alpha")
 
+    def test_multi_node_cache_consistency_on_deleted_entry(self):
+        """Verify IS_CHANGED changes when an entry is deleted, matching load_prompts."""
+        # Seed with two entries
+        self._seed_library([
+            {"id": "a", "text": "alpha"},
+            {"id": "b", "text": "bravo"},
+        ])
+        node = self.mod.PromptLibraryMulti()
+        kwargs = {
+            "prompt_id_1": "a,b",
+            "prompt_id_2": "",
+            "prompt_id_3": "",
+            "separator_1": ", ",
+            "separator_2": ", ",
+            "separator_3": ", ",
+        }
+        # Get initial signature and output
+        sig_before = node.IS_CHANGED(**kwargs)
+        out_before = node.load_prompts(**kwargs)
+        self.assertEqual(out_before[0], "alpha, bravo", "both ids should be present")
+
+        # Delete "b" from library
+        self._seed_library([{"id": "a", "text": "alpha"}])
+
+        # Get new signature and output
+        sig_after = node.IS_CHANGED(**kwargs)
+        out_after = node.load_prompts(**kwargs)
+
+        # Verify both change consistently
+        self.assertNotEqual(sig_before, sig_after, "IS_CHANGED should detect deletion")
+        self.assertEqual(out_after[0], "alpha", "output should skip deleted id")
+        # The crucial part: when the output content changes, IS_CHANGED must also change
+        # so cache invalidation works correctly
+
     # ---- Comic-strip nodes ---------------------------------------------
 
     def test_scene_node_joins_non_empty_fields(self):