Automate visual failure comparisons in /review tests

.github/aw/actions-lock.json

@@ -1,5 +1,10 @@
 {
   "entries": {
+    "actions/checkout@v4": {
+      "repo": "actions/checkout",
+      "version": "v4",
+      "sha": "11d5960a326750d5838078e36cf38b85af677262"
+    },
     "actions/checkout@v7.0.1": {
       "repo": "actions/checkout",
       "version": "v7.0.1",
@@ -25,11 +30,6 @@
       "version": "v7.0.1",
       "sha": "043fb46d1a93c77aae656e7c1c64a875d1fc6a0a"
     },
-    "github/gh-aw-actions/setup-cli@v0.82.14": {
-      "repo": "github/gh-aw-actions/setup-cli",
-      "version": "v0.82.14",
-      "sha": "b6d1443e05b8716267fa19425b99aa4f12006b4a"
-    },
     "github/gh-aw-actions/setup@v0.82.14": {
       "repo": "github/gh-aw-actions/setup",
       "version": "v0.82.14",

.github/docs/maui-ci-facts.md

@@ -268,6 +268,21 @@ Messages like `Baseline snapshot not yet created`, missing snapshot paths, or sn
 environment-version mismatches are strong **unrelated** evidence — unless the PR adds or
 modifies that visual test or the affected snapshot/platform.
 
+The automated `/review tests` lane gathers failed UI result IDs from the public
+`vstmr.dev.azure.com/.../testresults/resultsbybuild` endpoint, then reads the public
+result-detail and attachment APIs. It publishes validated baseline/actual/diff PNGs to
+the repository's `review-tests-assets` branch. A trusted post-step inserts as many
+complete expandable comparison panels as fit inside the single test-failure analysis
+comment while enforcing gh-aw's URL, mention, and character limits; excess panels are
+reported as omitted rather than creating another comment. Visual publishing is
+supplementary evidence only: missing images never raise or lower the deterministic
+verdict ceiling. Each panel also shows a conservative relationship label derived from
+the exact test-and-platform `deterministicAttribution` plus exact changed snapshot/test
+scope: `regressed-vs-base` or directly changed visual coverage is Likely PR-caused,
+`pre-existing-on-base` or `known-issue` is Likely unrelated, and indeterminate or
+unmatched evidence remains Needs human investigation. A same-named snapshot on another
+platform and platform or area mismatch alone never change the label.
+
 ## Platform mismatch
 
 Platform mismatch is **supporting** evidence, not proof. An iOS-only test failing on a