test(M00): record Goldshire pose evidence

2026-07-12 01:59:42 +04:00
parent 657a1d888a
commit 03891edeb2
4 changed files with 116 additions and 13 deletions
@@ -14,10 +14,11 @@ This offline M00 diagnostic recovers reproducible checkpoint framing when an ori
  -Checkpoint elwynn_adt_boundary `
  -YawOffsets -15,-10,-5,0,5,10,15 `
  -PitchOffsets -10,-5,0,5,10 `
+  -ViewportWidth 1280 -ViewportHeight 960 `
  -WaitSeconds 2
 ```

-Each candidate receives its own output directory and comparison report. `ranking.json` sorts candidates by mean perceptual error and then changed-pixel ratio. Use `-PlanOnly` to validate the Cartesian grid and output paths without rendering. Run a coarse grid first, then a finer grid around the best candidate.
+Each candidate receives its own output directory and comparison report. The viewport must exactly match the reference image dimensions; the runner fails on `size_mismatch` instead of inventing a score. `ranking.json` sorts candidates by mean perceptual error and then changed-pixel ratio. Use `-PlanOnly` to validate the Cartesian grid and output paths without rendering. Run a coarse grid first, then a finer grid around the best candidate.

 ```mermaid
 flowchart LR
@@ -37,3 +38,11 @@ The smallest error is a candidate, not automatic fidelity approval. Geometry, ma
 ## Verification

 The headless capture dry-run must report requested yaw and pitch offsets. Comparator `--self-test` covers paired metrics. Sweep `-PlanOnly` covers deterministic grid expansion and output naming without requiring a display. A real ranked sweep requires a display and a private build 12340 reference image.
+
+## Goldshire Inn evidence — 2026-07-12
+
+The private `2560x1440` build 12340 `goldshire_inn_large_wmo` reference was evaluated at FOV 62 degrees. A coarse 5x5 grid over yaw/pitch `[-20, -10, 0, 10, 20]` reduced mean error from `0.099632` at `(0, 0)` to `0.077575` at `(-10, -20)`. Extending pitch produced `0.070901` at `(0, -40)` and `0.063574` at `(0, -60)` with changed-pixel ratio `0.549556`.
+
+This is not a camera-pose solution. Human inspection shows the original-client reference centered on the Goldshire Inn, while the zero-offset OpenWC candidate contains forest/roads and no inn; negative pitch increasingly fills the frame with grass. The monotonically improving full-frame metric rewards similar green color coverage rather than landmark alignment. Therefore the manifest pose is unchanged and this checkpoint remains spatial/placement-composition evidence. A future registration metric must use approved landmark or region masks before it can select camera pose automatically.
+
+During this run the sweep orchestration was hardened in two ways: expected comparator exit code `1` is now collected through an explicit child process, and viewport dimensions are explicit. A reference/candidate `size_mismatch` now fails before ranking instead of producing an empty metric.