Port retail's CTransition::check_other_cells (acclient_2013_pseudo_c.txt :272717-272798) into Transition.FindEnvCollisions so the foot-sphere sees walls in EVERY cell it overlaps, not just the one cell the player's center is in. Closes the Holtburg inn vestibule wall walk-through (cell 0xA9B40164 has only 4 polys; adjacent 0xA9B40157 has 23 walls that are never queried today). Architecture: new CellTransit.FindCellSet overload (preserves the candidate HashSet that FindCellList currently discards), new private Transition.CheckOtherCells method (direct port of the retail loop), one wire-up in FindEnvCollisions between the existing primary-cell BSP return and the synthesis fall-through. ~380 LOC total. Out of scope: FindObjCollisions (already landblock-radius broadphased), synthesis multi-cell search (A3's job), var_4c re-target (defer). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
17 KiB
Phase A4 — multi-cell BSP iteration (FindEnvCollisions)
Date: 2026-05-20 Status: Spec — awaiting user approval before plan-writing. Phase: ISSUES #83 fix sequence — Phase A4 of A1→A1.5→A1.6→A1.7→A2→A3→A4. Author: Claude Opus 4.7.
Summary
Transition.FindEnvCollisions currently queries the one cell the
player's center is in for wall/floor BSP collision. Retail's
CTransition::check_other_cells queries every cell the sphere
geometrically overlaps — primary cell first, then each neighbour cell
in the cell_array set built by CObjCell::find_cell_list.
Concrete user-visible symptom: the Holtburg inn vestibule (cell
0xA9B40164) has only 4 BSP polys. When the player stands in the
vestibule but their foot-sphere extends into the adjacent room (cell
0xA9B40157, 23 walls, 38 % hit rate when the player is centered
there), the adjacent room's walls are never queried — the player
walks through them.
Fix: port retail's check_other_cells loop. Reuse the existing
CellTransit.FindCellList candidate-set machinery (already builds the
right set; currently throws it away). Wire it into
FindEnvCollisions after the primary cell's BSP returns OK.
Retail oracle
CTransition::check_other_cells at
docs/research/named-retail/acclient_2013_pseudo_c.txt:272717-272798.
0050ae80 CObjCell::find_cell_list(&this->cell_array, &var_4c, &this->sphere_path);
0050ae90 if (this->cell_array.num_cells > 0)
0050aec9 do {
0050ae98 int32_t* cell = this->cell_array.cells.data[i].cell;
0050aea4 if ((cell != 0 && cell != arg2)) { // skip primary
0050aeaf result = cell->find_collisions(this); // vtable +0x88
0050aeb7 switch (result) {
0050afcf case 2: case 3: return result; // Collided/Adjusted
0050aef3 case 4: this->collision_info.contact_plane_valid = 0;
0050aef9 this->collision_info.contact_plane_is_water = 0;
0050af07 return result; // Slid
0050aeb7 }
0050aea4 }
0050aec6 i += 1;
0050aec9 } while (i < this->cell_array.num_cells);
After the loop: var_4c is the cell that NOW contains the sphere
center if it moved outside the primary cell. Retail re-targets
sphere_path.check_cell and check_pos.objcell_id accordingly.
Architecture
Three pieces, each independently testable.
1. CellTransit.FindCellSet — new overload
Current FindCellList returns a single uint (the containing cell)
and discards the candidate HashSet<uint> it builds. The new overload
preserves it:
public static uint FindCellSet(
PhysicsDataCache cache,
Vector3 worldSphereCenter,
float sphereRadius,
uint currentCellId,
out IReadOnlyCollection<uint> cellSet);
Returns the containing cell id (same as FindCellList). The cellSet
out parameter is the full candidate set (every cell the sphere
overlaps via portal/landcell-grid expansion). Existing
FindCellList callers are untouched.
Implementation: extract the body of FindCellList into a private
helper that returns both, and make FindCellList a thin wrapper that
discards the set. Zero behavior change to existing call sites.
2. Transition.CheckOtherCells — new private method
Direct port of retail's loop. Lives in TransitionTypes.cs next to
FindEnvCollisions.
private TransitionState CheckOtherCells(
PhysicsEngine engine,
Vector3 footCenter,
float sphereRadius,
IReadOnlyCollection<uint> cellSet)
{
var sp = SpherePath;
var ci = CollisionInfo;
if (engine.DataCache is null) return TransitionState.OK;
// Deterministic ordering for greppable probe logs.
foreach (uint cellId in cellSet.Where(c => c != sp.CheckCellId).OrderBy(c => c))
{
var cell = engine.DataCache.GetCellStruct(cellId);
if (cell?.BSP?.Root is null) continue; // R2 guard: stale CellPhysics
// Transform sphere into THIS cell's local space.
// (Mirrors the primary-cell pattern at TransitionTypes.cs:1413-1485.)
var localCenter = Vector3.Transform(footCenter, cell.InverseWorldTransform);
var localCurrCenter = Vector3.Transform(sp.GlobalCurrCenter[0].Origin, cell.InverseWorldTransform);
var localSphere = new Sphere { Origin = localCenter, Radius = sphereRadius };
Sphere? localSphere1 = sp.NumSphere > 1
? new Sphere {
Origin = Vector3.Transform(sp.GlobalSphere[1].Origin, cell.InverseWorldTransform),
Radius = sp.GlobalSphere[1].Radius }
: null;
if (!Matrix4x4.Decompose(cell.WorldTransform, out _,
out Quaternion cellRotation, out Vector3 cellOrigin))
{
Console.WriteLine(/* warn — fall back to identity + .Translation */);
cellRotation = Quaternion.Identity;
cellOrigin = cell.WorldTransform.Translation;
}
var result = BSPQuery.FindCollisions(
cell.BSP.Root, cell.Resolved, this,
localSphere, localSphere1, localCurrCenter,
Vector3.UnitZ, 1.0f, cellRotation, engine,
worldOrigin: cellOrigin);
if (PhysicsDiagnostics.ProbeIndoorBspEnabled)
Console.WriteLine($"[other-cells] primary=0x{sp.CheckCellId:X8} iter=0x{cellId:X8} result={result}");
switch (result)
{
case TransitionState.Collided:
case TransitionState.Adjusted:
if (!ObjectInfo.State.HasFlag(ObjectInfoState.Contact))
ci.CollidedWithEnvironment = true;
return result;
case TransitionState.Slid:
ci.ContactPlaneValid = false;
ci.ContactPlaneIsWater = false;
return result;
// OK: continue
}
}
return TransitionState.OK;
}
3. Transition.FindEnvCollisions — wire-up
Existing structure (at src/AcDream.Core/Physics/TransitionTypes.cs:1407-1605):
indoor branch:
1. transform sphere → cell-local
2. primary cell BSP.FindCollisions
3. if non-OK: return (existing)
4. TryFindIndoorWalkablePlane synthesis ← unchanged
5. if HIT: ValidateWalkable indoor plane ← unchanged
6. if MISS: fall through to outdoor terrain ← unchanged
outdoor branch:
7. SampleTerrainWalkable + ValidateWalkable
A4 changes one thing: insert a CheckOtherCells call between steps 3 and 4 (indoor branch only — outdoor multi-cell is covered by terrain sampling already).
if (cellState != TransitionState.OK)
{
if (!ObjectInfo.State.HasFlag(ObjectInfoState.Contact))
ci.CollidedWithEnvironment = true;
return cellState;
}
+ // A4: query every other cell the sphere overlaps.
+ // Retail oracle: CTransition::check_other_cells at
+ // acclient_2013_pseudo_c.txt:272717-272798.
+ // Discard the containing-cell return — sp.CheckCellId is already
+ // authoritative for the primary cell we just queried.
+ _ = CellTransit.FindCellSet(engine.DataCache, footCenter, sphereRadius,
+ sp.CheckCellId, out var cellSet);
+ var otherCellsState = CheckOtherCells(engine, footCenter, sphereRadius, cellSet);
+ if (otherCellsState != TransitionState.OK) return otherCellsState;
// ── Synthesize indoor walkable contact plane ──────────────
bool walkableHit = TryFindIndoorWalkablePlane(...);
The synthesis call (TryFindIndoorWalkablePlane) stays primary-only.
A3 deletes it entirely; making it multi-cell is wasted work.
Mutation contract
All mutations to Transition / SpherePath / CollisionInfo from
any cell's BSP query are kept. No save/restore. Matches retail
exactly. Specifically:
- Intermediate cells return OK → BSPQuery wrote nothing concerning to Transition state (or wrote a ContactPlane via Path 5 step-up, which is valid world-space data and retainable across cells).
- The halting cell writes whatever it writes (wall normal, position adjustment) and that becomes the final state.
- Slid additionally clears
ContactPlaneValid+ContactPlaneIsWaterper retail's explicit case-4 handling.
Out of scope (intentional)
FindObjCollisions. Already broadphases by landblock-radius viaShadowObjects.GetNearbyObjects(currPos, queryRadius, ..., landblockId, ...), not cell-keyed. No change needed for shadow-object collision.var_4cre-target ofcheck_pos.objcell_id. Retail's lines 272760-272797 re-target the sphere's check cell mid-frame if it moved outside the primary cell. OurPhysicsEngine.ResolveCellIdruns on the next tick to re-resolve cell membership; A1.7 closed the "no cell contains player" gap. Defer the var_4c port — would only matter mid-frame, and we lack evidence it's needed.- Synthesis (
TryFindIndoorWalkablePlane). Primary-cell only. A3 deletes it entirely; making it multi-cell would just be wasted work. - A2 (PHSP inversion fix). Orthogonal — A2 fixes per-poly precision at the tangent boundary; A4 fixes which cells get queried. Land them separately.
Edge cases
| # | Case | Handling |
|---|---|---|
| E1 | Cell set has only the primary cell | Loop iterates zero times, returns OK. No-op. |
| E2 | Candidate cell's BSP.Root is null |
continue (skip). Matches handoff R2 guard. |
| E3 | Candidate cell has 0 polys but valid BSP root | BSPQuery returns OK (no leaves intersect). Continue. |
| E4 | Primary cell returns Slid before A4 runs | Existing if (cellState != OK) return halts. CheckOtherCells never invoked. ✓ |
| E5 | Outdoor primary with indoor building stab nearby | FindCellList's outdoor branch handles via AddAllOutsideCells + CheckBuildingTransit. Cell set already includes both. No new code. |
| E6 | Same shadow object referenced by two cells | N/A for A4. Shadow dedup is FindObjCollisions's concern, already handled. |
| E7 | CellPhysics.WorldTransform doesn't decompose |
Same fallback as primary cell: warn + identity rotation + .Translation. |
| E8 | Indoor primary, sphere extends through exit portal to outdoor | FindCellList's exitOutside branch adds adjacent outdoor landcells. They have no BSP → E2 skip. Outdoor terrain collision still flows through the existing fall-through after CheckOtherCells returns OK. |
| E9 | Performance — indoor doorway with 7 cells in set | Each iteration ≈ 50 µs BSP + 3 µs transform = ~370 µs per ResolveWithTransition cycle. At 30 Hz = ~11 ms/sec. Acceptable. |
Risks
- R1 — iteration order ambiguity. Deterministic sort by cellId is arbitrary; on close calls a different cell could be "first to Collide" depending on geometry. But Collided is Collided; the player stops either way. Acceptable.
- R2 — outdoor candidate cells in the set with no BSP. Skipped by
E2 guard. They don't contribute BSP collisions. Outdoor terrain
continues to flow through the existing post-
CheckOtherCellspath. - R3 — BSPQuery state mutation across cells. Matches retail's no-save/no-restore behavior. Intermediate OK results don't corrupt state because BSPQuery's mutations on OK are either no-op or valid (world-space CP via Path 5).
Testing strategy
Unit tests (~280 LOC)
tests/AcDream.Core.Tests/Physics/CellTransitFindCellSetTests.cs (~80 LOC, new file):
Sphere_FullyInsidePrimaryCell_ReturnsOnlyPrimary— sphere doesn't touch portal plane → set = {primary}.Sphere_StraddlingPortal_ReturnsBothCells— synthetic 2-cell fixture; sphere overlaps portal plane → set = {primary, neighbour}.Sphere_AtExitPortal_AddsOutdoorNeighbours— indoor cell with exit portal; sphere straddles → set includes adjacent outdoor landcells.
tests/AcDream.Core.Tests/Physics/TransitionCheckOtherCellsTests.cs (~120 LOC, new file):
EmptyCellSet_ReturnsOk— single-cell case, no-op.AllCellsReturnOk_ReturnsOk— three cells, all OK → final OK, no halt mutation.FirstCollidedHalts_KeepsMutations— second iterated cell returns Collided; verifyCollidedWithEnvironment = trueand remaining cells NOT queried.SlidClearsContactPlane— cell returns Slid; verifyContactPlaneValid = false+ContactPlaneIsWater = false, remaining cells not queried.NullBspRootIsSkipped— cell in set hasBSP.Root == null→ skipped, iteration continues.
tests/AcDream.Core.Tests/Physics/FindEnvCollisionsMultiCellTests.cs (~80 LOC, new file):
- End-to-end integration: synthetic primary cell (empty BSP) + adjacent cell (one wall poly); sphere center in primary, sphere extends into adjacent → returns Collided with adjacent cell's wall normal.
Baseline regression
dotnet build -c Debuggreen.- All existing physics tests pass (1129-test baseline holds, per the
2026-05-21 handoff). Specifically the indoor-BSP world-origin fix
(Bug B, commit
de8ffde) and the A1-A1.7 series must remain green.
Visual acceptance (the real test)
Launch with light-probe set (ACDREAM_PROBE_INDOOR_BSP=1,
ACDREAM_PROBE_CELL=1, ACDREAM_PROBE_CELL_CACHE=1).
- Vestibule wall block (primary acceptance). Walk to Holtburg
inn cell
0xA9B40164(the 4-poly vestibule). Approach a wall in adjacent cell0xA9B40157. Wall must BLOCK when the foot-sphere extends across the cell boundary even though the player's center is still in the vestibule. Probe must show[other-cells] primary=0xA9B40164 iter=0xA9B40157 result=Collided. - Stairs walk-through (secondary, likely closes). Walk up the inn stairs. Riser walls in the adjacent cell from where the player's foot-sphere center is must block, not pass through. If this fails: file a separate sub-issue for per-cell stair physics-poly coverage. Do NOT extend A4's scope to chase it.
- No regression in existing fixes:
- "Thin air" inside cottages still absent (A1 still works).
- Walls outside buildings still solid (A1.5 still works).
- Doorway exit still transitions correctly (A1.7 still works).
- Free-fall through floor at threshold does NOT regress (synthesis untouched).
Stop rule. Per CLAUDE.md: three failed visual verifications = write a handoff, not a fourth attempt.
Probe
Reuse the existing ACDREAM_PROBE_INDOOR_BSP env var (already wired to
PhysicsDiagnostics.ProbeIndoorBspEnabled). Add one log line per
iterated cell:
[other-cells] primary=0x<id> iter=0x<id> result=<state>
Low volume — only fires when CheckOtherCells actually iterates a
non-primary cell. Suppressed when the flag is off.
File map
| Change | File | Approx LOC |
|---|---|---|
New FindCellSet overload |
src/AcDream.Core/Physics/CellTransit.cs |
+30 |
New CheckOtherCells private method |
src/AcDream.Core/Physics/TransitionTypes.cs |
+60 |
Wire-up in FindEnvCollisions |
src/AcDream.Core/Physics/TransitionTypes.cs |
+6 |
| Cell-set enumeration tests | tests/AcDream.Core.Tests/Physics/CellTransitFindCellSetTests.cs |
+80 |
| CheckOtherCells halt-semantics tests | tests/AcDream.Core.Tests/Physics/TransitionCheckOtherCellsTests.cs |
+120 |
| End-to-end multi-cell integration | tests/AcDream.Core.Tests/Physics/FindEnvCollisionsMultiCellTests.cs |
+80 |
| Total | ~380 LOC |
Acceptance criteria
dotnet build -c Debuggreen.dotnet testgreen; 1129-test physics baseline holds, no regressions in A1/A1.5/A1.6/A1.7/Bug B suites.- Three new test files pass with all listed cases.
- Visual acceptance #1 passes — vestibule wall blocks at Holtburg inn.
- Visual acceptance #3 passes — no regressions in prior fixes.
- Roadmap updated with A4 shipped status.
docs/research/2026-05-21-open-items-pickup-prompt.mdupdated to reflect A4 closed; remaining items (stairs verify, A2, A3, lighting) re-ordered if A4's outcome changes priorities.
Retail decomp anchors (verbatim)
acclient_2013_pseudo_c.txt:272717-272798—CTransition::check_other_cells(A4 oracle).:323725-323939—BSPTREE::find_collisions(already ported; the call we make per cell).:309085-309092—CObjCell::find_cell_listsphere-path variant (already ported asCellTransit.FindCellList; A4 adds the set-returning overload).:308742-308783—CObjCell::find_cell_list(Position, n_spheres, CSphere*)variant (parent of the sphere-path variant).
Why A4 doesn't fix everything
A4 is a wall-collision fix. It does NOT:
- Synthesize a floor plane (that's
TryFindIndoorWalkablePlane's job — unchanged by A4, removed by A3). - Handle the doorway "no floor poly at this XY" free-fall (that's Bug A's actual root cause from the 2026-05-20 session — A3's job to investigate further, after A4 + A2 land).
- Fix the tangent-boundary
polygon_hits_sphererejection (that's A2).
A4 ships the iteration foundation. A2 and A3 build on it. Stairs walk-through is a known unknown — A4 should fix it via the same multi-cell BSP path, but if a stair's geometry is laid out as a single cell (no boundary to straddle), the BSP itself may have a coverage gap that needs separate investigation.