docs(issues): #464 reproduced from the script — camera in front, zoomed out through the open face into the hall; stairwell draws without walls

The eye is legitimate (retail's sweep passes the same opening); the defect
is the draw of the stairwell cells seen back through the building's
exterior portals. Matrix row invariant re-worded; the one retail capture
to request is the oh-capture walk + alphadepth at exactly this eye.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/ISSUES.md

@@ -101,6 +101,8 @@ then `/teleloc F4180113 39.502899 19.589222 181.119247 1 0 0 0`, zoom out, tilt
 
 **Investigation report (2026-09-03 21:47, `docs/research/2026-09-01-overhaul/464-camera-investigation.md`):** the read-only decomp + offline-replay investigation reaches the same conclusion independently: eleven enumerated sweep/root differences are each unreachable at this pose or make acdream MORE constrained than retail; the real-DAT replay seats the pivot in `0x114`, stops a straight-back boom on `0x114`'s east pier at y=16.448 for every pitch to 55 degrees and under the vault at 70 degrees, and the production walk from root `0x114` floods exactly `[0x114, 0x113, 0x112]` with seven exit views. **Reclassified: #464 is an interior-root DRAW defect with seven exit views (landscape flush / Z clear / exit seals under `outside_view.view_count > 0`, `PView::DrawCells` @0x005a4840), S4's depth/alpha domain, not the camera; the section 10 matrix row's invariant is re-worded from 'the eye stays inside' to 'root 0x114 with 0x113/0x112 flooded draws the stairwell solid'.** Two REAL openings let the eye out in both clients (a 4 m archway in `0x114` at y=16.15, x 33.98-38.02, reached by a boom yawed left; the open y=24 face behind a player facing up the stairs); retail's sweep passes them too (difference 7), so a frame from beyond them is again a draw-side question. The owner reproduction should also set `ACDREAM_PROBE_FACILITY_STAIRS=1` (the `[cathedral-stair]` line: viewer/player/root/eye/rootFlood per frame) next to `ACDREAM_PROBE_FLAP=1`. Owed when the fix lands: promote the two replays (the sweep stops on the pier under flat traversal; the walk from root `0x114` floods 114/113/112) into InstalledDat pins so the fix cannot regress the camera side.
 
+**REPRODUCED from the script (lead, 2026-09-03 21:51, `logs/selfgate-20260903-215041-464-arch/artifacts/screenshots/464-arch-05-yawright500-out.png`, route `route-464-arch.txt`, code `c3b989305`, both probes on):** at the top pose, CameraRotateRight held 500 ms swings the chase camera to the FRONT of the player (the orbit covers about 180 degrees in that time); zooming out from there drives the boom 11.03 m forward with `pulledIn=0.00` and `collNormValid=False`: the eye passes through the stairwell's open y=24 face and comes to rest at `(36.07, 29.58, 185.79)` in cell `0xF4180109`, the hall (`[flap-sweep] ok=True viewerCell=0xF4180109`, `[flap-cam] root=0xF4180109 eyeInRoot=Y res=None`, `[cathedral-stair] rootFlood=[107=1,112=0] visited=[107=1,112=1]`). The frame is the owner's picture: the player on stair slabs that float with no side walls, the hall's arches, green terrain and sky through the arch, one pier. The eye is legitimate (the investigation's difference 7: retail's viewer sphere passes that open face too, the tower shell collides only from its origin land cell), so the defect is the DRAW of the stairwell cells from an outside root: `0x112` is visited but not root-flooded (reached through the building's exterior portals), and its walls, `0x113`'s and `0x114`'s are not on screen. This is the S3/S4 building-seen-from-outside path (the outside-view block, exit views, building portal stab lists), not the camera. A walk-transcript run at this exact eye follows; the retail walk/alphadepth capture at this eye does not exist in the oh-capture kit and is the ONE capture to ask the owner for (retail: `/teleloc F4180113 39.502899 19.589222 181.119247 1 0 0 0`, swing the camera to the front of the character, zoom fully out, run the `oh-capture` walk + alphadepth templates).
+
 ## #463 — Soak `turn` samples can report a 0.1 ms GPU p50 (the turn window catches non-world frames)
 
 **Status:** OPEN — found 2026-09-03 by the Campaign OVERHAUL v2 S3 review's perf re-measure.