acdream/docs/research/2026-05-28-a8-cellar-flap-root-cause.md

# A8 cellar-flap — structured debugging root cause (2026-05-28 PM)

## Method

Systematic-debugging Phases 1-3, all evidence gathered **offline** via the
`tools/A8CellAudit` tool (extended with a `portals` mode) — no live launches
needed. Deterministic, instant, reproducible.

## Phase 1 — evidence

Scenario (from `launch-a8-probe-normal-20260528-194536.out.log`):
- Camera in cell `0xA9B40171`, `inside=True really=True`.
- `camBldgs=[0xA]`, `visN=7 [0x16F,0x170,0x171,0x172,0x173,0x174,0x175]`.
- Portal stencil mask = 12 verts (not the old over-punch case).
- Bisection (prior session): writer is **Step 4 content**; disabling Step-2
  punch does **not** fix it.

Offline audit findings:

**Building grouping** (`A8CellAudit buildings 0xA9B40000`):
```
buildingOrdinal=10 registryId=0xA model=0x01002232 portalCells=[0xA9B4016F,0xA9B40170]
```
Building `0xA`'s LandBlockInfo seed = `{0x16F, 0x170}`. `BuildingLoader` then
BFS-expands through interior portals → all 7 cells (incl. the cellar). The
BFS matches WB's `PortalService` (same algorithm), so the grouping is not the
divergence.

**Exit-portal ownership** (`A8CellAudit portals ...`):

| Cell | exit portals (0xFFFF) | interior | role |
|------|----|----|------|
| `0x16F` | **1** | 1 | ground floor (window/door) |
| `0x170` | **1** | 1 | ground floor (window/door) |
| `0x171` (camera) | **0** | 3 | cellar |
| `0x172`–`0x175` | **0** | 1–2 | cellar rooms |

So the 12-vert mask = `0x16F` exit (6v) + `0x170` exit (6v). **The cellar
camera (zero exit portals) is marking the two ground-floor windows.**

**Topology**:
```
0x171.portal[0] -> 0x170  (stairwell/hatch, polyId 54)
0x170.portal[1] -> 0x171  (polyId 5)
0x170.portal[0] -> EXIT   (window/door to outside, polyId 4)
```
Cellar connects directly up to ground floor `0x170`; `0x16F` is one further hop.

**Occluder geometry** (`A8CellAudit 0xA9B40170` / `0xA9B40171`):
- `0x170` floor poly `0x0002` (n.Z=+1) **emits** — the cellar's ceiling/occluder exists.
- `0x171` has a ceiling `0x0003` (n.Z=-1, emits) AND three `NoPos` polys
  `0x0036/0x0037/0x0038` (surface `0x080000DF`) that do **not** emit —
  `0x0038` is a ceiling-plane poly = the **stairwell hole** up to the ground floor.

## Phase 2 — pattern vs WB

WB `RenderInsideOut` marks the building's exit portals (flat — same as us) and
relies on **Step-3 cell depth** to occlude them: terrain only survives where the
punched/cleared far-depth isn't overwritten by rendered cell geometry.

Our code matches that structure. The difference that produces the visible flap:
WB's outside view through a portal is the world geometrically behind that
portal; from a cellar, the only un-occluded opening is the **stairwell hole**
(`0x0038`, not rendered). Through that hole, stencil=1 (ground-floor window
marked) and depth=far → **Step 4 draws the entire outdoor world (terrain +
buildings) through the hole**, not a window-sized sliver. The two ground-floor
windows are 1–2 BFS hops above the camera and should contribute essentially
nothing from the cellar, but their full silhouettes are marked.

"Disable Step-2 punch doesn't fix it" is explained: the leak pixels are the
stairwell hole, which has **cleared (far) depth** regardless of the punch
because no cell geometry covers it — terrain passes `DepthFunc.Less` either way.

## Phase 3 — single hypothesis (root cause)

**The inside-out exit-portal stencil mask is built by flat-marking the exit
portals of every visibility-BFS-reached cell. From the cellar, the BFS reaches
the ground-floor cells, whose windows get full-silhouette-marked. Where the
cellar's stairwell hole leaves those silhouettes un-occluded, Step 4 paints the
whole outdoor world through them. There is no constraint tying a deeper cell's
exit portal to the portal chain (here: the narrow stairwell) through which its
cell became visible.**

This is a flat-vs-constrained masking gap. Not a depth bug (occluders emit and
render), not the Step-2 punch, not the camera-side filter (the cellar camera is
geometrically on the interior side of a ground-floor window's plane, so the
per-portal filter passes it).

## Phase 4 — fix options

1. **Camera-cell-scoped mask (minimal, conservative).** Mark only the camera
   cell's own exit portals. Cellar (0 exit portals) → empty mask → no leak;
   windowed room → marks its own windows. **Risk:** loses daylight through an
   *adjacent* cell's window seen across a doorway in multi-cell ground-floor
   rooms (e.g. the inn) — a visible-but-minor regression, and the flat approach
   was wrong there anyway.

2. **Vertical-portal-aware scoping (targeted).** Don't propagate exit-portal
   marking across a floor/ceiling (vertical-normal) portal. The cellar→ground
   stairwell is a horizontal-plane portal; suppressing inheritance across it
   stops the cellar from marking ground-floor windows while preserving
   same-level multi-cell rooms. Needs per-portal polygon-normal classification.

3. **WB recursive/constrained portal masking (faithful, largest).** Constrain
   each deeper portal's stencil to the screen region of the portal chain leading
   to it. Correct for all cases (cellar + multi-cell rooms) but a substantial
   port of WB's recursive RenderInsideOut.

**Recommendation:** option 2 is the best correctness/effort trade — it fixes the
cellar without the inn regression risk of option 1, and is a principled scoping
rule (don't inherit a different vertical level's exterior openings) rather than a
band-aid. Option 3 remains the eventual faithful target if cross-level portal
visibility ever needs to be exact.

## Reproduction / verification assets

- `tools/A8CellAudit` `portals` mode (added this session) dumps any cell's
  `CellPortals` offline. `A8CellAudit buildings <lb> <radius>` dumps
  building→cell grouping. These make the whole investigation re-runnable in
  seconds with zero launches.