Summarize successful and failed runs

CHANGELOG.md

@@ -2,7 +2,7 @@
 
 ## 0.3.0 — reproducible evidence scaffold
 
-Verified locally with the checked-in `run_checks.ps1` gate (31 tests):
+Verified locally with the checked-in `run_checks.ps1` gate (32 tests):
 
 - V2 depth baseline with optional DA3 camera-aware adapter and early contract
   validation.
@@ -21,6 +21,8 @@ Verified locally with the checked-in `run_checks.ps1` gate (31 tests):
   whose `DepthAnything3` API symbol is unavailable.
 - Failure manifests now have a validated schema, so failed model requests are
   reproducible diagnostics rather than unstructured log lines.
+- Run summarization now includes failed manifests, their stage, and structured
+  error details alongside successful backend comparisons.
 - Pinned Torch/TorchVision baseline and commit-pinned optional DA3 research
   dependency.