acdream/docs/research/2026-08-30-cathedral-review-datgeometry.md

Cathedral review — F418 authored DAT geometry (companion report)

Delivered 2026-08-30 by the DAT-geometry agent: fresh byte-exact parsers over the installed client_cell_1.dat/client_portal.dat (every record parsed with 0 bytes left over). Ground truth for the synthesis. Condensed; citations and caveats preserved from the agent's full text.

The landblock (LandblockInfo 0xF418FFFE)

24 EnvCells (0xF4180100–0x0117), 88 objects, 9 buildings. The cathedral complex is authored in the BUILDINGS list: 0x01001FB0 (north hall, cells 0x100–0x102), 0x01001FB2 (the 0x103–0x105 colonnade block), 0x01001FB3 (the main cathedral, cells 0x106–0x111, 13 building portals), 0x01001FB7 (south hall 0x112–0x115), 0x01001FD3 (sealed annex 0x116–0x117), plus four portal-less shells.

THE FALLS — identity corrected (slices 4/5 chased the wrong owners)

The waterfall emitters are LANDBLOCK OBJECT-LIST STABS (outdoor), not EnvCell statics and not weenies:

• Objects[0..6]: Setup 0x0200085A ×7 — EAST falls jets (x≈169–170, z=158.6), script 0x330008A6 → emitter 0x320004A4 (v=22.5 m/s jets).
• Objects[7..13]: Setup 0x02000859 ×7 — NORTH falls jets (y≈120.8), script 0x330008A5 → emitters 0x320004A3/A2 (arc + plunge, accel −40).
• Objects[14..19]: Setup 0x020003C5 ×6 — lake-base mist at z=99.9 (incl. an authored exact-duplicate pair). Their runtime ids live in the 0xC namespace (LandblockStaticEntityIdAllocator): east 0xCF418000–06, north 0xCF418007–0D, mist 0xCF41800E–13.

The probe-observed 0x4F418012–15 owners are NOT the falls — the 0x4 namespace is InteriorEntityIdAllocator; simulating the hydration filter maps them to cell 0xF4180106's OWN stabs: two 0x020005C1 TORCHES (flame emitters 0x320003AB/0x320003A5), one bare prop (0x01000DF0), one plant (0x020007A3). Slice 5's "the falls are EnvCell stab statics" was a misidentification; the cell-owners n=5 submissions were the torches.

The falls billboards' degrade tables keep them visible to 392 m — they are authored to be seen across the whole vista.

THE "STALKS" — premise corrected

Setup 0x0200085A (the PV campaign's "monster stalks") is not a visible stalk: its sole part GfxObj 0x010001EC is retail's runtime-hidden marker (GfxObj flags 0xA; DIDDegrade 0x1100002E = never drawn at any distance), shared as the invisible anchor by 13 of this landblock's setups (falls jets, torch/sound anchors). The visible dark shapes at the east cliff are these anchors DRAWN — i.e., a never-drawn-degrade skip missing on the path that renders them (the marker-degrade port fce250ec is stranded on the quarantined PV branch per the audit report). Cell 0xF418010D (the audit's stalk suspect) is actually a fully-interior undercroft/crypt with 81 candle-prop stabs and NO exit portals — its non-flooding is CORRECT behavior, not the stalk mechanism.

EXIT-PORTAL GEOMETRY — the "thin seam bands" claim REFUTED

All exit portals at the defect cells are FULL-SIZE authored openings with sills at the cell floor (z=169.80):

• 0xF4180104: four exits — two 10.72 × 10.20 m facade trapezoids (y=48 south, y=72 north) and two 24.0 × 9.2 m side ARCHES (x≈31 west, x≈41 east; the 0.57 m X-thickness in an AABB is just the arch bow, the polys stand in the Y-Z plane). One interior ceiling portal → 0x105.
• 0xF4180106: one 10.72 × 10.20 m exit (y=48 north) + two 1.67 × 5.89 m door arches (→0x10F, →0x110).

Slice 6's "authored polys are thin bands far narrower than the real opening" is therefore WRONG about the authoring — if the walk's exit views come out thin or empty (the ov==0 frames), the loss happens in OUR clipping/traversal, not in the dat.

Two real seam facts to carry instead:

1. 0x104's south exit poly and 0x106's north exit poly are COINCIDENT — the same 10.72 × 10.20 m rectangle in the same y=48 plane (buildings 01001FB2 and 01001FB3 abut), EACH marked exit-to-outside (0xFFFF) rather than cell-to-cell. This shared double-exit plane at y=48 is "the seam".
2. The openings' sills sit ~9.8 m above the outdoor terrace (floor 169.8 vs terrain 160): the lake (z≈100) only becomes visible within ~2–3 m of an opening — the downward vista is sill-clipped by geometry, not by portal truncation.

Portal flags (from all 24 cells)

Bit 0x1=ExactMatch, 0x2=PortalSide, and 0x4 co-occurs always and only with other_cell==0xFFFF (observed flags 0x5/0x7 on exits) — so the old "flag 0x4 = leads outside" lore is observationally right as a CORRELATE in this landblock even though the DatReaderWriter enum doesn't name it (the datformat report found no 0x4 member; semantic asserted from perfect correlation, not decomp).

Water

The lake/falls SURFACES are outdoor TERRAIN paint: WaterShallowSea (type 18) in the north basin (z 96–108) and terrace water court (z=160), FauxWaterRunning (type 22) on the cliff faces. No water objects besides the six mist emitters.

Cell 0xF4180104 / 0x0106 contents

• 0x104: 16 stabs — plants + FOUR looping ambient-SOUND anchors (0x0200093D/3E/3F/09AF — how the falls' roar is heard indoors; no particles) + prop.
• 0x106: 5 stabs — the two torches, sound marker, prop, plant.
• 0x105: a ceiling-linked upper cell (horizontal portals only).

Caveats (the agent's own)

The 0x4F418012–15 mapping mirrors LandblockBuildFactory's hydration filter in Python (high confidence, not executed C#); flag-0x4 semantics asserted from correlation; terrain-name indices aligned against ACE's LandDefs enum.