feat(lab): add recorded realtime spatial playback

docs/22_MILESTONE_4_OBJECT_CENTRIC_RECORDED_REALTIME_CV.md

@@ -1058,9 +1058,37 @@ The accepted corrected M4.6 result is
 The common viewer keeps VIDEO/CAMERA/3D/PLAN synchronized and independently
 toggles `CURRENT INCREMENT` and `ROLLING MAP`. CAMERA requests one exact JPEG
 decoded from the bounded fMP4 GOP instead of preparing the complete video.
-Regression frames are now `138`, `274`, `1880` and `2584`. M4.7 may proceed only
-from the corrected identities above; physical-live, collision, navigation and
-actuation authority remain false.
+Regression frames are now `138`, `274`, `1880` and `2584`.
+
+### 2026-08-05 — M4.6 recorded-realtime visual timeline
+
+The 32 sealed visual samples remain immutable regression checkpoints, but they
+are no longer the playback mechanism. A product-owned recorded spatial evidence
+timeline now exposes every one of the `4,489` source frames without loading or
+copying the complete `345 MB` frame ledger into the browser:
+
+- timeline metadata contains only the exact ordered source timestamps and the
+  frozen source, rig, corridor and authority identities;
+- spatial evidence is read on demand from indexed ledger offsets in bounded
+  chunks of at most `24` frames; the GUI requests `12` and retains at most four
+  chunks;
+- each frame carries the bounded current point sample, metric/rolling obstacles,
+  camera proposals, decision counts and one exact CAMERA URL;
+- VIDEO, 3D and PLAN run from one source-time clock at `0.5×`, `1×` or `2×`;
+  CAMERA pauses that clock and decodes exactly the selected sequence;
+- the Three.js scene separates immutable rig/corridor/grid objects from dynamic
+  point and obstacle layers, so a frame update does not rebuild the scene or
+  reset the operator view;
+- playback never changes the sealed M4.6 result, its visual ledger or its
+  `replay-simulated`/no-actuation authority.
+
+Browser acceptance verified live frame advance and changing current/rolling
+geometry in both 3D and PLAN, exact CAMERA sequence binding, VIDEO clock drift
+below `0.1 s` after initial decode, `2×` pacing, fullscreen layout and an empty
+browser error log. The first cold timeline index took approximately `3.5 s`;
+after indexing, a 12-frame spatial chunk was served in approximately `33 ms` and
+was about `1.27 MB`. M4.7 may proceed only from the corrected identities above;
+physical-live, collision, navigation and actuation authority remain false.
 
 ## Implementation order