fix(manifest): keep per-item metrics written to disk when a generation run saves

metrics_runner computes metrics for an existing session straight to disk; a
generation run's in-memory copy of those items had no metrics and would have
wiped them on its next save. Disk wins for an item's metrics when the memory
item has none (merge_manifest_user_changes, flush_batch_with_vae's disk sync,
the remote VAE worker's disk sync). Fresh in-memory metrics are kept.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

image_generation.py

@@ -1123,6 +1123,8 @@ def flush_batch_with_vae(pending_batch, vae, img_dir, existing_data, session_nam
                             local_item["favorited"] = disk_item["favorited"]
                         if "rejected" in disk_item:
                             local_item["rejected"] = disk_item["rejected"]
+                        if "metrics" in disk_item and "metrics" not in local_item:
+                            local_item["metrics"] = disk_item["metrics"]  # computed later for an existing session
 
             except Exception as e:
                 print(f"[GridTester] ⚠️ Error syncing with disk manifest: {e}")

manifest_utils.py

@@ -185,6 +185,10 @@ def merge_manifest_user_changes(manifest_path, existing_data):
                 if current_item.get("note"):
                     item["note"] = current_item["note"]
                     notes_preserved += 1
+                # Per-image metrics computed for an existing session (metrics_runner) live only on
+                # disk; a generation run's in-memory copy must not wipe them on its next save.
+                if "metrics" in current_item and "metrics" not in item:
+                    item["metrics"] = current_item["metrics"]
             # else: This is a new item, no merge needed
         
         # Log merge results if any user changes were preserved

remote_vae.py

@@ -262,6 +262,8 @@ class RemoteVAEDecodeWorker:
                                     local_item["favorited"] = disk_item["favorited"]
                                 if "rejected" in disk_item:
                                     local_item["rejected"] = disk_item["rejected"]
+                                if "metrics" in disk_item and "metrics" not in local_item:
+                                    local_item["metrics"] = disk_item["metrics"]  # computed later for an existing session
 
                     except Exception as e:
                         print(f"[GridTester] ⚠️ Error syncing with disk manifest: {e}")

tests/test_metrics_preserved_on_save.py

@@ -0,0 +1,55 @@
+"""A generation run's manifest save must not wipe per-image metrics that were computed
+for an existing session (metrics_runner writes them straight to disk while the run's
+in-memory copy of those items has no `metrics`). Disk wins for an item's metrics when
+the in-memory item has none; freshly computed in-memory metrics are kept.
+
+Loads the real manifest_utils.py via importlib (tests/conftest.py registers a stub
+under the package name), like test_comparisons_manifest.py does.
+"""
+import importlib.util
+import json
+import os
+
+_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+_spec = importlib.util.spec_from_file_location("manifest_utils_real_metrics", os.path.join(_ROOT, "manifest_utils.py"))
+mu = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(mu)
+
+
+def _item(i, **extra):
+    d = {"id": i, "file": f"/view?filename=img_{i}.webp", "favorited": False, "rejected": False, "note": ""}
+    d.update(extra)
+    return d
+
+
+def test_disk_metrics_survive_a_stale_in_memory_save(tmp_path):
+    path = str(tmp_path / "manifest.json")
+    disk = {"session_name": "s", "meta": {}, "items": [
+        _item(2, metrics={"sharpness": 12.5, "entropy": 7.1}),
+        _item(1, metrics={"sharpness": 3.0}),
+    ]}
+    with open(path, "w", encoding="utf-8") as f:
+        json.dump(disk, f)
+
+    # the generation run's copy predates the metrics job: same items, no metrics, plus a new image
+    stale = {"session_name": "s", "meta": {}, "items": [_item(3), _item(2), _item(1)]}
+    mu.save_manifest(path, stale)
+
+    with open(path, encoding="utf-8") as f:
+        after = json.load(f)
+    by_id = {it["id"]: it for it in after["items"]}
+    assert set(by_id) == {1, 2, 3}, "the new image is kept"
+    assert by_id[2]["metrics"] == {"sharpness": 12.5, "entropy": 7.1}
+    assert by_id[1]["metrics"] == {"sharpness": 3.0}
+    assert "metrics" not in by_id[3]
+
+
+def test_fresh_in_memory_metrics_are_not_overwritten_by_disk(tmp_path):
+    path = str(tmp_path / "manifest.json")
+    with open(path, "w", encoding="utf-8") as f:
+        json.dump({"session_name": "s", "meta": {}, "items": [_item(1, metrics={"sharpness": 1.0})]}, f)
+    fresh = {"session_name": "s", "meta": {}, "items": [_item(1, metrics={"sharpness": 99.0})]}
+    mu.save_manifest(path, fresh)
+    with open(path, encoding="utf-8") as f:
+        after = json.load(f)
+    assert after["items"][0]["metrics"] == {"sharpness": 99.0}