test(3d): populate baseline-webgl (47 browser goldens) + clean stale diff artifacts per run

- baseline-webgl/: first population from the passing local Chromium renders
  (47/47 under the parity floor) — activates the webgl-self regression gate
  (floor 0.005). CI may re-promote its own renders if the Linux/swiftshader
  rasterization differs beyond the floor (the README's CI-promote flow).
- compare-dirs.ts: recreate the output dir on every run — triptychs/heatmaps
  are only written for CHANGED pairs, so artifacts from earlier (redder) runs
  lingered and misreported scenarios that had since gone green (a stale
  red-state gen-cylinder triptych survived three green runs).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/tools/screenshots/compare-dirs.ts

@@ -113,6 +113,10 @@ function main(): void {
     const levelName = (args.level as string) || null;
     const floorFor = makeFloorResolver(args);
 
+    // Recreate the output dir from scratch: triptychs/heatmaps are only written
+    // for CHANGED pairs, so artifacts from earlier (redder) runs would otherwise
+    // linger and misreport scenarios that have since gone green.
+    fs.rmSync(outDir, { recursive: true, force: true });
     fs.mkdirSync(outDir, { recursive: true });
     const oldNames = listPngs(oldDir);
     const newNames = new Set(listPngs(newDir));