From 5a90fa8a04b4d499a211e65fa35a2aa3ff883b09 Mon Sep 17 00:00:00 2001
From: silveroxides <ishimarukaito@gmail.com>
Date: Mon, 24 Aug 2026 17:45:22 +0200
Subject: [PATCH] feat(cwb): save merge metadata

---
 docs/consensus_mergers.md |  6 ++++
 nodes/consensus_merger.py | 68 ++++++++++++++++++++++++++++++++++++---
 2 files changed, 70 insertions(+), 4 deletions(-)

diff --git a/docs/consensus_mergers.md b/docs/consensus_mergers.md
index 9b8f674..88ab589 100644
--- a/docs/consensus_mergers.md
+++ b/docs/consensus_mergers.md
@@ -241,6 +241,12 @@ alignment, weighting, or merge. Genuine factor ranks remain variable during
 CWB. The first maximum-rank input supplies the output component space, so the
 output rank remains the largest genuine input rank. Merged outputs never
 contain or invent alpha tensors.
+
+Every CWB merge node writes a `cwb.merge` safetensors metadata record. It
+contains selected source model names, fully resolved CWB settings, and merge
+options. Presets are stored as effective settings rather than by preset name.
+Non-empty exclude/discard filters and their glob mode are included. CWB reports
+and diagnostics are not saved in metadata.
 Similarity alignment operates on paired latent-rank components. A rows and B
 columns share one mapping; fixed input/output feature axes are never reordered.
 For a rank-R reference and rank-r source it evaluates an R-by-r similarity
diff --git a/nodes/consensus_merger.py b/nodes/consensus_merger.py
index 9e2c678..9ffd841 100644
--- a/nodes/consensus_merger.py
+++ b/nodes/consensus_merger.py
@@ -2,10 +2,11 @@
 
 from __future__ import annotations
 
+import json
 import logging
 import os
 from contextlib import closing
-from dataclasses import dataclass, field
+from dataclasses import asdict, dataclass, field
 from typing import Iterable
 
 import comfy.utils
@@ -796,6 +797,51 @@ def build_custom_cwb_settings(**values) -> CWBSettings:
     return settings
 
 
+def _cwb_merge_metadata(
+    source_models: list[str],
+    params: dict,
+    settings: CWBSettings,
+    *,
+    operation: str,
+) -> str:
+    metadata = {
+        "schema_version": 1,
+        "operation": operation,
+        "source_models": source_models,
+        "effective_cwb_settings": asdict(settings),
+        "merge_options": {},
+    }
+    for key in (
+        "mismatch_mode",
+        "save_dtype",
+        "process_device",
+        "lazy_load",
+        "force_clear_cache",
+        "override_dtype",
+        "include_1d_diffs",
+        "counterfactual_weight_sweep",
+        "embedding_coalesce",
+        "target_vector_count",
+        "coalesce_similarity_threshold",
+        "coalesce_position_window",
+        "vision_boundary_embeddings",
+        "legacy_boundary_search",
+        "boundary_reference_embedding",
+        "boundary_similarity_threshold",
+    ):
+        if key in params:
+            metadata["merge_options"][key] = params[key]
+    exclude_patterns = params["exclude_patterns"].split()
+    discard_patterns = params["discard_patterns"].split()
+    if exclude_patterns or discard_patterns:
+        metadata["filters"] = {
+            "exclude_patterns": exclude_patterns,
+            "discard_patterns": discard_patterns,
+            "glob_patterns": params["glob_patterns"],
+        }
+    return json.dumps(metadata, sort_keys=True, separators=(",", ":"))
+
+
 def _common_prefix_length(tensors: list[torch.Tensor]) -> int:
     if not tensors:
         return 0
@@ -1622,7 +1668,7 @@ class ConsensusMergerLogic:
             )
             if lora_mode:
                 result = cls._merge_loras(
-                    handlers, low_bit_sets, model_type, params, settings
+                    handlers, low_bit_sets, model_type, params, settings, model_names
                 )
             else:
                 result = cls._merge_generic(
@@ -1633,6 +1679,7 @@ class ConsensusMergerLogic:
                     settings,
                     embedding_union=embedding_union,
                     boundary_handler=boundary_handler,
+                    source_models=model_names,
                 )
             if params.get("_return_cwb_report"):
                 report = diagnostics.render(
@@ -1672,6 +1719,7 @@ class ConsensusMergerLogic:
         *,
         embedding_union,
         boundary_handler=None,
+        source_models,
     ):
         primary = handlers[0]
         keys = set()
@@ -1699,7 +1747,14 @@ class ConsensusMergerLogic:
             }
             work_units.append((key, entries))
 
-        with atomic_uel_writer(output_path, primary.metadata()) as writer, torch.no_grad(), closing(
+        output_metadata = (primary.metadata() or {}).copy()
+        output_metadata["cwb.merge"] = _cwb_merge_metadata(
+            source_models,
+            params,
+            settings,
+            operation="cwb_embedding_merge" if embedding_union else "cwb_tensor_merge",
+        )
+        with atomic_uel_writer(output_path, output_metadata) as writer, torch.no_grad(), closing(
             stream_work_units(
                 handler_map, work_units,
                 pin_memory=str(params["process_device"]).startswith("cuda"),
@@ -1917,7 +1972,9 @@ class ConsensusMergerLogic:
         return cls._output_name(model_type, output_path)
 
     @classmethod
-    def _merge_loras(cls, handlers, low_bit_sets, model_type, params, settings):
+    def _merge_loras(
+        cls, handlers, low_bit_sets, model_type, params, settings, source_loras
+    ):
         parsed = [parse_lora_layers(handler.keys()) for handler in handlers]
         logical_maps = []
         logical_cores = []
@@ -2077,6 +2134,9 @@ class ConsensusMergerLogic:
             work_units.append((("passthrough", key), entries))
 
         output_metadata = (handlers[0].metadata() or {}).copy()
+        output_metadata["cwb.merge"] = _cwb_merge_metadata(
+            source_loras, params, settings, operation="cwb_lora_merge"
+        )
         output_metadata["alpha_normalized"] = "true"
         output_metadata["alpha_normalization"] = (
             "lora_up := lora_up * (alpha / rank); alpha tensors removed"