Ports retail ACRender::polyClipFinish (0x006b6d00, pc:702749) near-eye
semantics into PortalProjection.ProjectToClip - the fundamental fix for
the in-plane portal clip family (climb strobes, tower-top roof/floor
flap while turning; live-corroborated this session: [viewer-diff]
0xAAB30108 strobing 27x mid-climb, whole interior dropping at the top).
Pseudocode: docs/research/2026-06-11-polyclipfinish-w0-clip-pseudocode.md.
Three legs, all decomp-driven:
1. ProjectToClip clips at w >= 0 EXACTLY (was EyePlaneW=1e-4), with
retail's any-negative-w gate. Boundary intersections land at w == 0
(homogeneous directions), so a portal the eye is CROSSING yields the
correct unbounded half-region that the bounded view-region clip cuts
to the screen. A w=0 vertex cannot survive a bounded region clip
into the divide (direction fails some edge of any bounded convex
region); the measure-zero corner case is guarded non-finite->empty.
2. CellView.CanonicalKey keys ALL-COLLINEAR (zero-area) views as their
snapped segment ("L:" + extremes) instead of rejecting them - retail
PROPAGATES degenerate views (ClipPortals decomp:433651-433711
forwards any count!=0 GetClip output, no area gate anywhere), keeping
the cell behind an exactly-in-plane portal in the draw list (cells
draw whole; onward floods die naturally). Rejection dropped the
whole chain for the frame - the parked-eye knife-edge band. Finite
key space unchanged -> dedup + strict-growth convergence intact.
3. The EyeInsidePortalOpening rescue is DELETED (the T2-documented
compensation for the 1e-4 divergence) along with EyeStandingPerpDist
+ PointInPoly2D. Empty clip = no flood, period (retail's rule).
CornerFloodReplay - the gate that REFUTED the previous deletion
attempt - passes WITHOUT the rescue under the W=0 port.
Harness criterion corrected to retail's rules (it codified the rescue):
cells fully BEHIND the camera are not required (all-behind portals clip
empty in retail); monotone area holds per root regime; the two
manufactured exact-on-plane steps assert root-only (boundary root pick
is ambiguous; the in-plane portal there is ~perpendicular to the gaze =
genuinely off-screen). Build_CollapsedInteriorPortalNearEye test
inverted to pin the retail empty-clip rule (it pinned the rescue).
New pins: eye-crossing portal -> w==0 boundary verts + half-region (not
sliver); gaze-along-plane degenerate view accepted + segment-key dedup;
non-finite guard. Replay harnesses (CornerFloodReplay, Issue120,
TowerAscent, HouseExit, Issue127) all green.
Suites: App 246+1skip / Core 1430+2skip / UI 420 / Net 294.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The live capture pinned it end to end. BuildInteriorEntitiesForStreaming
lifts the render-side cell transform +0.02 m Z (shell z-fighting vs
terrain - a DRAW concern) and passed that LIFTED transform to
BuildLoadedCell, so every plane in the visibility graph sat 2 cm high.
The portal side test's in-plane window is +-10 mm: an eye standing ON a
floor containing a HORIZONTAL portal (the tower's deck lip 010A->0107,
stair landings, cellar mouths) sits 0-10 mm above the TRUE plane = 10-20
mm BELOW the lifted plane -> outside the window -> the cell behind the
portal side-culled out of the flood. Captured live at the stair top:
root=0xAAB3010A eye z=126.803 vs the portal plane at 126.80, flood=1,
0x0107 (the whole tower interior incl. the staircase) dropped WHILE THE
GAZE LOOKED STRAIGHT AT IT - "stairs disappear and you can walk on
them", and the roof/edge flap as the gaze swung the marginal admissions.
Vertical doorways were immune (the lift slides their planes along
themselves) - exactly why this hit stairs/decks/floors and not doors.
Chase chain (the apparatus did all the work): [viewer] print-on-change
probe with eye@mm -> the user's climb capture -> [viewer-diff] naming
the dropped cells per flip -> headless replay of the exact captured
(eye,fwd) frame: healthy UNLIFTED, reproduces ONLY with the production
lift -> gate-by-gate diagnostic (side test dot=+0.003 unlifted vs
-0.017 lifted; clip + rescue exonerated; knife-edge z-sweep all-stable,
killing the float-chaos theory).
Fix: BuildLoadedCell receives the PHYSICS (unlifted) transform; the
drawn shells keep their lift. The seal/punch fans (which read the
visibility LoadedCell's WorldTransform) now stamp TRUE depth - MORE
consistent with the unlifted terrain they protect.
Pins: CapturedTopOfStairs_MainCellStaysInFlood - arm 1 (unlifted =
post-fix production) asserts the main cell admitted at the captured
frame; arm 2 (lifted) is the mechanism canary asserting the drop, with
instructions if it ever starts passing. Plus the gate-by-gate
diagnostic + knife-edge sweep as the investigation record.
Also this session: Issue127FloodFlipReplayTests (the captured 4 cm
outdoor flip pair replays STABLE across fovs/pre-gate arms - the
outdoor churn is NOT the flood math; remaining #127 = distant-building
admission churn, lower priority now that the tower-cell drops are
explained by the lift), and the [viewer-diff] probe (per-flip added/
removed cell naming - keep, it found this).
Suites: App 242+1skip, Core 1422+2skip, UI 420, Net 294.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause (pinned by the new deterministic exit-walk harness, NOT guessed):
under an interior render root, the exit-portal SEAL stamps the door fan at
TRUE depth after the gated full depth clear, and T1's "ALL dynamics last"
pass then drew the outdoor-classified player depth-tested - every fragment
beyond the door plane z-failed against the seal across the whole aperture.
Harness measured the full window: from the moment the sphere center crosses
the plane until the eye follows (~2.6 m of camera lag, ~2.2 s at walk speed)
the player is invisible; while straddling, the beyond-plane body half clips
at the plane. The handoff's three cone-level candidates are all EXONERATED:
the cone walk passes every step; (eye, ViewerCellId) come from the same
SweepEye call with camera-update-before-visibility-read in the same frame;
the side-test window is sub-epsilon under healthy resolution.
Retail oracle (grep-named-first): PView::DrawCells 0x005a4840 runs
LScape::draw FIRST (pc:432719), then the gated depth clear (pc:432731-32)
and the exit-portal seals (pc:432785-86); outdoor cell objects draw inside
the landscape stage (DrawBlock 0x005a17c0 -> DrawSortCell pc:430124), and
an object draws once per overlapped shadow cell (pc:430056-64) - the
straddling body composes from both stages, neither half clips.
Fix: RetailPViewRenderer assigns dynamics to the OUTSIDE stage under an
interior root when outdoor-classified OR sphere-straddling an exit-portal
plane of their flood-visible cell (DynamicDrawsInOutsideStage - pure, the
harness drives it as the ordering contract); they ride the landscape slice
draw (pre-clear, seal-protected) with the same per-slice cone test as
outdoor statics. Indoor dynamics keep the last pass (retail loop C);
straddlers draw in both (retail shadow dual-draw). Outdoor roots keep
all-dynamics-last - the BR-2 punch-after-dynamics lesson (88be519) stands.
Apparatus: HouseExitWalkReplayTests - dat-backed corner-building exit walk
driving the production stack headlessly (RetailChaseCamera damping ->
healthy-sweep viewer resolution -> PortalVisibilityBuilder.Build ->
ClipFrameAssembler -> ViewconeCuller -> the DrawDynamicsLast predicate +
a CPU seal-depth model). 5 tests: cone pin, seal-depth pin, straddle
dual-draw pin, per-step table, stale-root window quantifier (#118 cand 2).
Suites: App 232 (227+5), Core 1416+2skip, UI 420, Net 294.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Investigation: retail's growth propagation RECURSES natively too
(AddViewToPortals -> FixCellList -> AdjustCellView -> AddViewToPortals,
Ghidra 0x005a52d0/0x005a5250/0x005a5770, no depth guard) - the in-place
recursion shape is faithful; retail's safety is fast convergence. Our
depth-128 firing means slow/non-saturating growth (each lap of a portal
cycle nests one recursion level), not necessarily a true infinite loop.
Two dat-backed sweeps over the corner-building cell set could NOT
reproduce the T5 firing:
- PortalPlaneCrossings_InPlacePropagationConverges: +/-6cm eye sweep
across every portal plane, seeded from both sides.
- InCellDirectionSweep_InPlacePropagationConverges: 3024 builds, in-cell
eye grid x 8 yaw x 3 pitch (the walking-and-turning regime).
Both pass with 0 firings -> production-only ingredients suspected (full
lookup graph - one T5 firing was 0x0162, another building - and/or the
real camera path).
Armed: PortalVisibilityBuilder.ConvergenceTripwireCount (test
observable, both Build + look-in sites) + DumpPropagationChain - on the
next firing the log carries root cell, eye, per-cell frequency summary,
and the 24-entry chain tail, so the cycle's structure (A<->B ping-pong
vs 3-cycle laps) reads directly off the output. Both sweeps stay as
regression pins.
App tests: 227 green (was 225; +2 pins).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE BUG (pinned deterministically by the new CornerFloodReplayTests harness — real
Holtburg cells, captured corner-press scenario): a smooth 2 cm/step monotonic eye
sweep across the 0172↔0173↔0171 doorway produced a NON-monotonic flood — on ~10 of
61 steps the player's room (0172) vanished from the flood entirely or collapsed to
a sub-pixel sliver, taking its downstream chain (016F, the outside view) with it.
Live, those isolated frames are the §4 background strobe: openings/passages flash
the clear color during transitions, and the corner press shows background at the
angles that park the eye near the doorway plane.
TWO root causes, both fixed:
1. ApplyReciprocalClip ran the reciprocal portal polygon through the legacy
divide-first ProjectToNdc + 2D Intersect path, justified by "the reciprocal is
never near the eye." That assumption is exactly false at doorways/corners: the
reciprocal IS the same opening whose plane the eye presses against (2-60 cm).
ProjectToNdc's MinW=0.05 eye-clip + side-plane clip + divide is knife-edge
there — 2 cm eye moves flipped its output between a no-op and a duplicated-
vertex hairline that ground the healthy region down to <3 distinct vertices.
FIX: route the reciprocal through the SAME homogeneous pipeline as the forward
clip (ProjectToClip + ClipToRegion) — which is what retail does:
PView::OtherPortalClip (decomp:433524-433563) runs the reciprocal through the
very same GetClip(finish=1) → ACRender::polyClipFinish homogeneous clipper.
Also ported retail's skip: exact_match portals (CCellPortal.exact_match,
acclient.h:32300; PView::ClipPortals :433689) bypass the reciprocal clip —
both sides share the same polygon, so re-clipping is redundant.
2. CellView.CanonicalKey missed COLLINEAR re-emissions: the homogeneous region
clipper legitimately inserts intersection vertices ON a subject edge when a
region edge grazes it, so BFS re-clip rounds re-emit the SAME geometric region
with 1-2 extra collinear edge vertices — keyed as distinct, defeating the
dedup and accumulating duplicate polygons (this was the real mechanism behind
the historical "float drift defeats the dedup" rationale that had parked the
reciprocal on the unstable path). FIX: canonicalize away collinear snapped
points (exact integer cross-products on the 1e-3 NDC grid) so the key is
purely a function of the region's corners.
Conformance: CornerSweep_FloodIsCompleteAndMonotone pins the fixed behavior —
61-step monotonic eye sweep ⇒ full flood every step, outside view always reached,
player-room region monotone (was: clean shrink 4.000→2.879 with zero drops, vs
~10 glitch steps before). Diagnostic facts (trace diff, hop microscope, primitive
scratch) retained as the apparatus.
Suites: App 223 green (incl. Build_AppliesReciprocalOtherPortalClip, now passing
with proper tightening AND dedup), Core 1377 green + the 4 pre-existing #99-era
failures + 1 skip, UI 420, Net 294. Visual gate pending: corner press, room↔room,
cellar↔floor, indoor↔outdoor transitions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Erik