fix(review): harden write-back, confirmation gate, and edge cases

- write-back: call upsert_tag inside _apply_tag's try so a tag-create failure
  is caught per-asset and apply_all no longer aborts mid-batch (was P1)
- write-back: surface _pipeline/processed write failures in the result instead
  of discarding them (was reported as success)
- ui: add title to the high confidence badge so approve-high-confidence's
  pre-action count is non-zero; confirm() before single-cluster apply
- core: Store context manager; close DB connection even if a route raises;
  guard split_cluster against a first-member boundary (empty cluster); add
  writeback_log lookup index
- ingest: split thumbnail download/write error handling and clean up the
  .tmp file on a write failure
- tests: upsert/processed write-failure regression tests + split-guard test

apps/trip-cluster/app/ingest.py

@@ -45,14 +45,23 @@ def run_ingest(client, store, thumbs_dir, *, date_from=None, date_to=None,
         if need_thumb:
             try:
                 data = client.download_thumbnail(a["id"])
-                # Write to a temp path and os.replace into place so a failure
-                # never leaves a 0-byte <id>.jpg behind.
-                tmp_path = f"{thumb_path}.tmp"
+            except Exception:
+                log.exception("Thumbnail download failed for asset %s", a["id"])
+                continue
+            # Write to a temp path and os.replace into place so a failure
+            # never leaves a 0-byte <id>.jpg behind; clean up the temp file
+            # if the write itself fails (e.g. disk full).
+            tmp_path = f"{thumb_path}.tmp"
+            try:
                 with open(tmp_path, "wb") as f:
                     f.write(data)
                 os.replace(tmp_path, thumb_path)
             except Exception:
-                log.exception("Thumbnail download failed for asset %s", a["id"])
+                log.exception("Thumbnail write failed for asset %s", a["id"])
+                try:
+                    os.unlink(tmp_path)
+                except OSError:
+                    pass
                 continue
 
         store.upsert_asset(Asset(

apps/trip-cluster/app/routes/review.py

@@ -113,9 +113,8 @@ def apply_one(cid):
     # Lazy import: app.writeback is owned by a later task and may be absent
     # at app-startup; importing here keeps the blueprint importable regardless.
     from app.writeback import apply_cluster
-    s = _store()
-    res = apply_cluster(_client(), s, cid)
-    s.close()
+    with _store() as s:               # close even if the Immich write-back raises
+        res = apply_cluster(_client(), s, cid)
     return jsonify(res)
 
 
@@ -123,7 +122,6 @@ def apply_one(cid):
 def apply_everything():
     # Lazy import: see apply_one above.
     from app.writeback import apply_all
-    s = _store()
-    results = apply_all(_client(), s)
-    s.close()
+    with _store() as s:               # close even if the Immich write-back raises
+        results = apply_all(_client(), s)
     return jsonify({"results": results})

apps/trip-cluster/app/static/app.js

@@ -136,6 +136,8 @@ function clusterReview() {
       location.reload();
     },
     async apply() {
+      // Write-back needs explicit confirmation (mirrors applyAll/merge).
+      if (!confirm('Apply this cluster’s decision to Immich?')) return;
       const r = await this.post(`/cluster/${this.selected}/apply`, {});
       alert(`Applied ${r.succeeded.length}, failed ${r.failed.length}.`);
     },

apps/trip-cluster/app/writeback.py

@@ -3,11 +3,15 @@ from photoflow.immich import pipeline
 APPLYABLE = ("approved", "non_trip", "skipped")
 
 
-def _apply_tag(client, store, asset_ids, action, tag, tag_id):
+def _apply_tag(client, store, asset_ids, action, tag):
     todo = [a for a in asset_ids if not store.already_applied(a, action, tag)]
     if not todo:
         return [], []
     try:
+        # upsert_tag is part of the write: a failure here must be caught and
+        # recorded like a tag_assets failure, not propagate out of apply_cluster
+        # (which would abort the whole apply_all batch and skip later clusters).
+        tag_id = client.upsert_tag(tag)
         client.tag_assets(tag_id, todo)
     except Exception as e:                       # noqa: BLE001 — recorded, surfaced
         for a in todo:
@@ -29,12 +33,11 @@ def apply_cluster(client, store, cluster_id) -> dict:
 
     if c.status == "approved":
         tag = c.decided_name or c.suggested_name
-        ok, fail = _apply_tag(client, store, included, "trip", tag, client.upsert_tag(tag))
+        ok, fail = _apply_tag(client, store, included, "trip", tag)
         succeeded += ok
         failed += fail
     elif c.status == "non_trip":
-        ok, fail = _apply_tag(client, store, included, "non-trip", pipeline.NON_TRIP,
-                              client.upsert_tag(pipeline.NON_TRIP))
+        ok, fail = _apply_tag(client, store, included, "non-trip", pipeline.NON_TRIP)
         succeeded += ok
         failed += fail
     # 'skipped': no content/non-trip tag, only processed below.
@@ -42,8 +45,10 @@ def apply_cluster(client, store, cluster_id) -> dict:
     failed_ids = {i for i, _ in failed}
     proc_targets = [a for a in included if a not in failed_ids]
     if proc_targets:
-        _apply_tag(client, store, proc_targets, "processed", pipeline.PROCESSED,
-                   client.upsert_tag(pipeline.PROCESSED))
+        # Surface processed-marker failures too — a discarded return here makes a
+        # failed _pipeline/processed write look like success in the CLI/UI summary.
+        _, proc_fail = _apply_tag(client, store, proc_targets, "processed", pipeline.PROCESSED)
+        failed += proc_fail
         for a in proc_targets:
             if store.already_applied(a, "processed", pipeline.PROCESSED):
                 store.mark_processed(a)