docs(issues): #464 camera-cell probe rounds — the eye never leaves the cell; re-read as the exit-portal seam-band family

Four probe-on self-gate rounds (run/zoom/tilt/mid) with the camera cell,
root and eye logged every frame: sweep ok, eye in root, no fallback, in
every frame. The DAT shows 0xF4180113/0xF4180114 are one stairwell split
horizontally; the zoomed-out eye stops 0.31 m in front of 0x114's
nine-vertex EXIT portal. Owner asked for one probe-on reproduction.

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

docs/ISSUES.md

@@ -90,6 +90,15 @@ remain closed. See
 
 **Moving reproduction (lead, 2026-09-03 21:19, Release `590fbb66c`, `logs/selfgate-20260903-211930-464-run/`, route `tools/overhaul-selfgate/route-464-run.txt`):** from the mid pose `F4180113 [39.36 21.29 179.77]` at rest the frame is the enclosed stairwell (`464-run-21.png`); 0.75 s after MovementForward is pressed, `464-run-22.png` shows the chase camera several metres above and behind the stairwell, outside the cell: a dark exterior slab fills the lower half of the frame, the balcony wall and the vaulted hall are drawn beyond it, and only the character's head shows above the slab; 0.78 s later (`464-run-23.png`) the camera is back inside and the frame is clean. So the owner's "running down the stairs" symptom is this defect: the chase camera transiently occupies a position outside the stairwell when it starts to follow, and the frame roots from there. From the top pose the moving frames (`464-run-01..03`) stay consistent. The frames are 0.77 s apart (the screenshot readback costs about 0.6 s), so the transient's true duration is unknown (at most about 1.5 s).
 
+**Camera-cell probe rounds (lead, 2026-09-03 21:28–21:37, `ACDREAM_PROBE_FLAP=1`, runs `logs/selfgate-20260903-212850-464-run-flap`, `-213145-464-zoom`, `-213343-464-tilt`, `-213626-464-mid`; routes `route-464-run/-zoom/-tilt/-mid.txt`):** in every frame of all four runs the sweep succeeded (`ok=True`), the eye was inside its resolved cell (`eyeInRoot=Y`), and the root fell back never (`res=None`) — the camera does NOT leave the cell. The DAT explains the two cells: `0xF4180113` (local AABB z 20.00–21.89, world z 180.0–181.9) and `0xF4180114` (z 181.89–190.7, y 16.14–24.0) are the SAME stairwell split horizontally at z 181.89 by an ExactMatch/PortalSide portal covering the whole ceiling; a player standing at the top (feet 181.12) already has their head in `0x114`, so the default eye 2.24 m up sits in `0x114` legitimately and the sweep's `pulledIn=0.00` is right. Zooming out (held CameraZoomOut, or the wheel — both call `RetailChaseCamera.AdjustDistance`) drives the desired eye to 10.31 m but the sweep stops it at 3.28 m, eye `(39.50,16.45,183.56)`, `collNormValid=True`: that contact is 0.31 m (the 0.3 m viewer sphere) in front of `0x114`'s nine-vertex EXIT portal (poly 29, flags 7, plane y=16.14, the whole back wall of the upper stairwell). Raising the camera stops it at the ceiling (`(39.50,19.59,188.27)`); look-up at the zoomed distance did not move the eye (desired stayed 3.44 m); look-up at default zoom lowered it to 1.0 m back / 0.8 m up; yaw ±1000 ms at the middle landing swung the eye to x 37.1–40.5 at z 180.8–181.0 inside `0x113`. None of those frames shows the owner's hall-through-the-walls view. **Revised reading:** the artifact is not the camera leaving the cell; the zoomed-out eye is pinned within the viewer-sphere radius of a large EXIT portal directly behind it, which is the exit-portal seam-band family (#456/#458: a portal the eye practically touches, the exit punch/seal and the `portalsDrawnCount`-gated Z clear of the outside-view block) — S4's depth/alpha territory, which is why the first-run `464-run-22.png` transient (the camera mid-descent, again next to an exit) shows the exterior slab. **Owner morning ask:** reproduce the zoom-out artifact once with the probe on so the exact eye, cell and `[flap]` portal lines are captured:
+
+```powershell
+$env:ACDREAM_PROBE_FLAP = "1"; $env:ACDREAM_SELFGATE_PASS = "testpassword"; & .\src\AcDream.Appin\Release
+et10.0\AcDream.App.exe --session-config tools\overhaul-selfgate\session.json 2>&1 | Tee-Object -FilePath launch-464.log
+```
+
+then `/teleloc F4180113 39.502899 19.589222 181.119247 1 0 0 0`, zoom out, tilt until the hall shows, take a screenshot, and note the time; the matching `[flap-cam]`/`[flap-sweep]`/`[flap]` lines in `launch-464.log` pin the frame.
+
 ## #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.