Replaces the single scope-global distance-sorted RetailAlphaQueue with
retail's own two independent FIFO lists (CLIP/ALPHA, capacity 3000 each,
D3DPolyRender::AddMeshToAlphaList's exact append-only/capacity-drop
behavior — Ghidra-verified 2026-09-04), routed by a new
RetailAlphaMeshRouter porting DrawMesh's five-row immediate/delayed
branch table and ConstructMesh's subset-mask formula as pure functions,
and drained at retail's four normal-world FlushAlphaList sites
(DrawBuilding/DrawBlock/PView::DrawCells/RenderNormalMode) under the
exact Ghidra-verified no-op predicate (both counts strictly below
threshold*3000). A new WalkFrameEventKind.SortCellExit /
IWalkEventSink.OnSortCellExit / IWalkFrameLeafRenderer.FlushSortCellExit
fires once per admitted land-block cell for DrawBlock's 0.75f valve,
pinned by a dedicated far/near ordering test in RetailFrameWalkTests.cs.
WbDrawDispatcher's two submit sites and ParticleRenderer's one route
through the router; since none of the three ever draws during the Sky
leaf, installs a detail surface, or sets MultiPassAlpha, rows 1/2/4/5
are provably unreachable there and the call sites assert loudly rather
than building unexercisable immediate-draw plumbing. FlushFartherThan,
RetailAlphaOrdering.ComputeViewerDistance, and every viewerDistance
argument on the submit path are deleted.
Scope note (packet s4-depth-alpha-packet.md §10): C4 (routing EnvCell's
transparent shell batches through the shared queue) was not attempted —
EnvCell draws one per-cell MultiDrawIndexedIndirect call with no
per-subset deferred-replay abstraction, and building one without visual
verification (no graphical client in this worktree) was judged out of
this bounded chunk's scope. AP-34 is therefore retired and replaced by
two narrower rows rather than deleted outright: AP-236 (the carried-
forward EnvCell-immediate residual) and AP-237 (a newly identified gap:
TranslucencyKind.AlphaBlend can arise from either retail's Alpha/
Translucent bits, mask 0x02/ALPHA, or the Translucent+ClipMap "cloud"
override, mask 0x08/CLIP — GroupKey doesn't retain the raw bit to tell
them apart, so the router always picks ALPHA; only known example is
cloud GfxObj 0x01004C35). Both are compositing-order-only divergences,
never blend/visual ones.
Mutation checks (each applied, confirmed failing, then reverted):
- FIFO drain order reversed -> 5 RetailAlphaQueueTests fail (order).
- FlushAlphaList `<` -> `<=` -> boundary/scratch tests fail (2250 case
reads drained=0 instead of 2250).
- Capacity check loosened (3000 -> 6000) -> overflow-drop test fails
(TryAppend returns true, PendingCount reads 3001).
- IsFirstForList forced true -> flag test fails once inspected on the
pre-flush two-entry snapshot (the post-flush single-survivor version
of this test was vacuous and rewritten).
- Router row 3 condition inverted -> both the hand-traced Theory (6
cases) and the 160-cell independent-truth-table brute force fail (20
mismatches).
- SortCellExit emitted before OnLandscapeCellTurn instead of after ->
RetailFrameWalkTests ordering pin fails ("SCX must immediately follow
its own cell's SC").
- Prepare-per-list instead of prepare-once-combined -> the CLIP/ALPHA
boundary batching test throws (index out of range).
Gates: Release build 0 warnings/0 errors. Hermetic lane (Lane!=Installed
Dat&...&Status!=KnownFailure) 6855/6855 passed. InstalledDat lane 249
passed / 10 failed — exactly the 4 pre-existing failures (#383 x2
LayoutImporter, TowerAscent KnownFailure, #458 Oh_doorway_still
KnownFailure) plus 6 NEW KnownFailure Facts
(AlphaFlushTranscript_*_MatchesRetailFrame2, one per capture) extending
this gate from PM/PC to AM/FL: the flush SITE+THRESHOLD sequence matches
the capture exactly for all six poses (including zero SortCellExit
drains in every capture, confirming the 0.75 valve is inert at these
scene complexities in both retail and this replay); the drained-COUNT-
per-list dimension diverges because this hermetic harness (matching the
existing PM/PC gate's own EmptyAlphaDepthWorldData design) carries no
live GfxObj/particle content, so every observed count reads (0,0)
against retail's real per-frame volume — both sequences quoted in full
per pose in the packet's new §10. Shader classes (VulkanShaderDescriptor
ContractTests/VulkanShaderManifestTests/RenderPackSpirvValidatorTests)
32/32 passed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Fixes the three-lens review blockers against 671eb3ad4 (S3 section 9.6 F1-F6).
F1 - slot key (blocking, retail). TerrainModernRenderer.DrawLandCells
normalizes every incoming landblockId to (id & 0xFFFF0000u) | 0xFFFFu
before the _idToSlot lookup: the walk hands 0xXXYY0000
(WalkLandBlock.LandblockId) but AddLandblock stores under the DAT id
0xXXYYFFFF (LandblockRenderPublisher.LandblockId) - every walk lookup
was missing and the walk path drew NO terrain. Unit-tested end-to-end
through a real RecordingGpuDevice-backed TerrainModernRenderer
(TerrainWalkSlotKeyNormalizationTests): AddLandblock(0xA9B4FFFF, ...)
is found by a 0xA9B40000 lookup, an unknown landblock is a silent
per-entry no-op, and a batch mixing a known and unknown entry submits
only the known one.
F2 - deferred cross-block batching (blocking, driver). Retail's
DrawSortCell always follows DrawLandCell (LC/SC strictly alternate,
never two LC in a row - S3 section 9 R1), so chunk 3's "merge
consecutive same-landblock LandCell events" rule never actually
merged anything; the driver review flagged batching as inert.
WalkFrameDriver.Replay now keeps ONE pending terrain batch across
landblocks ((landblockId, side, cellIndex) entries, cleared at
Replay's own start); a LandCell event only appends; every OTHER event
kind that will itself submit GPU work (StreamMark, Sky, CellShell,
PunchFan, AlphaBarrier, LandscapeFlush, ClearInteriorDepth,
ExitSeals) flushes the pending batch first; a StaticParticles/
CellParticles turn asks the new ParticleSystem.
HasRenderableEmittersInCell (an allocation-free sibling of
CopyRenderableEmittersInCell) and, when the cell has no renderable
emitter, submits nothing and does NOT flush either - the whole point
of the deferred rule. The end of Replay flushes the remainder. This
is order-preserving by construction: a flush always lands at the
exact point the unbatched draw would have, so GPU submission order -
and therefore pixels - is identical to the unbatched baseline; only
the number of small terrain draw calls shrinks.
TerrainModernRenderer.DrawLandCellRuns becomes DrawLandCells(
viewProjection, IReadOnlyList<(uint LandblockId, int SideCellCount,
int CellIndex)>) - one MultiDrawIndexedIndirect over every entry's
runs, unknown slots skipped per-entry. IWalkFrameLeafRenderer.
DrawLandCellBatch drops its separate landblockId parameter to match
(a batch can span several landblocks now) and gains
HasRenderableEmittersInCell.
Batch-count demonstration: driven through a real WalkFrameDriver
Replay (OnLandCellTurn_MergesAcrossLandblocksOverAnEmptyParticleTurn_
RealSubmissionsSplit), 4 LandCell turns across 3 distinct landblocks,
separated only by an empty particle turn, a real StreamMark, and a
building's alpha barrier, submit as exactly 3 DrawLandCellBatch calls
(2+1+1) instead of 4 - the empty particle turn's non-flush merges two
otherwise-separate cross-landblock entries. At production scale the
same mechanism is expected to cut the terrace-edge frame's ~578
individual DrawLandCell events (S3 section 9's captured transcript
count) to "tens" of submitted batches, per the contract's own
expectation: most terrain cells have no particle owner nearby, so the
strict LC/[empty-SC]/LC/[empty-SC]/... run collapses into one batch
per region bounded by real content (a building, a StreamMark-worthy
cell, or a genuine emitter) rather than per cell.
F3 - per-frame terrain diagnostic (blocking, build/test). The walk
leaf no longer brackets each batch with TerrainDrawDiagnosticsController
.Begin()/Complete() (a per-batch Stopwatch Restart/Stop pair that was
pushing one timing SAMPLE per batch, not per frame).
RetailPViewPassExecutor.DrawWalkLandCellBatch instead times its own
call with a raw Stopwatch.GetTimestamp() delta (no allocation) and
hands the ticks to the controller's new AccumulateWalkBatch;
RetailPViewRenderer.DrawWalkDrivenStatics calls the new
CompleteWalkTerrainFrame() exactly once, immediately after
driver.Replay finishes - "the end of the walk replay", where the
deleted whole-stage terrain leaf's own Begin()/Complete() bracket
used to close - which pushes ONE elapsed-time sample (even a
zero-batch frame pushes a zero sample: one sample per frame, not per
landscape turn) and publishes on the existing 5-second cadence.
TerrainRenderDiagnosticFacts gains a Draws field alongside
VisibleSlots (both were the same field before); TerrainModernRenderer
tracks its own per-frame WalkVisibleSlotCount/WalkDrawCount (a
HashSet<int>/int cleared in BeginFrame, populated by DrawLandCells),
and the diagnostics source reports those whenever the walk drew at
least one batch this frame, falling back to the non-walk Draw()
path's VisibleSlots otherwise (the two paths never both run in the
same frame). The [TERRAIN-DIAG] line's meaning (cpu_us per frame) is
unchanged, so the S3 section 9.5 before/after compare stays valid.
F4 - driver pins for the LandCell position (major). Three RunFrame-
level pins replace the deleted TERRAIN:0 pins: an outdoor-root
sequence (SKY, then one LANDCELL, driving RetailFrameWalk.
DrawLandscape directly with a one-view/zero-vertex WalkPortalView so
WalkLandscape.CheckBlocks' admission stays the same deterministic
"CY-only" test RetailFrameWalkTests already relies on, while still
satisfying WalkFrameDriver's real >=1-active-view fail-loud guard);
an interior-root test with one real exit view and one populated
block (SKY, LANDCELL, LFLUSH, SEALS, SHELL...) built on the existing
RunFrame_InteriorFloodWithExitView_... fixture; and the T4 batching
pin re-expressed for the F2 rule (OnLandCellTurn_
MergesAcrossLandblocksOverAnEmptyParticleTurn_RealSubmissionsSplit,
described above). The fake leaf's DrawLandCellBatch now logs
LANDCELL:<lb>:<side>:<idx>[,...] per batch and gains
HasRenderableEmittersInCell backed by an opt-out CellsWithoutEmitters
set (default true - has-emitters - so every pre-existing pin in the
file keeps its old unconditional-submission behavior unchanged).
F5 - no code change: the walk's in-view gate is unchanged; no
whole-block terrain re-added.
F6 - minor/notes: DrawLandCells' own comment now states the walk's
CheckBlocks/landcell_check admission is the sole terrain culling
authority (retail has no separate terrain frustum test); the
HandleLandscapeTurn comment's inverted claim is corrected (a FARTHER
building's punch survived because NEARER terrain was drawn BEFORE
it, not after - the interleave now draws it after, matching retail);
the "flat/directional-shadow paths" claim is corrected to the one
actual caller, WorldScenePassExecutor.DrawFlatTerrain (a directional-
shadow receiver selects its pipeline inside the SAME DrawRhi call,
not through a second caller); the cathedral order-trace token gains
the LOD side/index (":LC<lb>/<side>:<idx>"); T2's vacuous "no
TERRAIN event" assertion in RetailFrameWalkTests is replaced by a
comment pointing at the F4 driver-level pins; and the stale
"Confirmed OH5 defect" row in oh1-construction-landscape-contract.md
is retired with "FIXED by S3 chunk 3 (commit 671eb3ad4 + fix round
1)".
App hermetic lane: 6,795/6,795 (up from 671eb3ad4's 6,786 baseline -
net +9 tests: 3 F1 slot-key tests, 2 F4a/b driver RunFrame pins, 3
TerrainDrawDiagnosticsController walk-frame tests, plus the T4->F4c
rewrite and the RetailPViewPassExecutorTests split are net neutral).
InstalledDat lane: 241 passed, the same 3 accepted failures (2
pre-existing #383 layout fixture-drift tests, 1 TowerAscent
Status=KnownFailure) - unchanged from baseline. Core Vfx tests:
109/109 (108 baseline + 1 new HasRenderableEmittersInCell lifecycle
pin mirroring CopyRenderableEmittersInCell's own add/move/remove
test).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Ports RenderDeviceD3D::DrawBlock @0x005a17c0's real per-cell order:
loop 1 (@0x005a1876) prepares shadow lists; loop 2 (@0x005a197d)
DrawLandCell(cell) @0x005a19c0 fires ONLY when the cell is in view,
STRICTLY BEFORE DrawSortCell(cell) @0x005a19e6, which fires whenever
alwaysDrawObjects (retail default 1 @0x00820ed4) or the cell is in
view. RetailFrameWalk.DrawLandscape now emits sink.OnLandCellTurn(
landblockId, side, cellIndex) at that exact point, per admitted cell,
before the existing DrawBuilding + OnLandscapeCellTurn (the
DrawSortCell half). WalkFrameDriver records one LandCell frame event
per turn and deletes the whole-stage TerrainSlice(0) emission and the
DrawTerrainSlice leaf outright — drawing all terrain before every
building let a nearer building's far-Z punch survive under farther
terrain drawn afterward, the doorway-behind-a-hill fragment bug from
the owner's G2 Holtburg screenshot; this chunk removes it by ORDER
alone, with no depth-compare change (S4's punch z-func question is
untouched, per the contract).
Index-run arithmetic (S3 §9.1 R3): the terrain mesh is cell-major
(LandblockMesh.Build: cy outer, cx inner, 6 indices/cell, 384/land-
block). A retail LOD cell (side n, LOD coords X,Y) covers cx in
[X*8/n,(X+1)*8/n), cy in [Y*8/n,(Y+1)*8/n) — one contiguous run per
covered cy row: side 8 -> one run of 6, side 4 -> two runs of 12,
side 2 -> four runs of 24; side 1's single coarse cell covers every
row contiguously so its 8 per-row runs collapse into ONE run of all
384 indices. TerrainModernRenderer.AppendCellIndexRuns (pure, no GPU,
no baked table) and DrawLandCellRuns (resolves the landblock's slot,
builds one DrawElementsIndirectCommand per run, reuses the existing
DrawRhi bind-and-submit path) implement this; TerrainModernRenderer
.Draw(...) is untouched and keeps serving non-walk callers (directional-
shadow receivers, the flat terrain path).
Order-preserving batching (S3 §9.2 B2): WalkFrameDriver.Replay merges
consecutive same-landblock LandCell events with no intervening event
into ONE DrawLandCellBatch leaf call; any other event splits the
batch. In production this rarely fires because retail's own
DrawSortCell (AlwaysDrawObjects=true) always interposes an object-
list turn between one cell's LandCell event and the next's — see
perfNote in the task report for the resulting command-count increase.
Deleted as dead: the _walkTerrainInViewLandcells field and
SetWalkTerrainInViewLandcells setter on RetailPViewPassExecutor (fed
only DrawWalkTerrainSlice's inViewLandcells filter, which no longer
exists — the walk's own per-cell CellInView admission is now the
sole terrain-visibility authority) and its two call sites in
RetailPViewRenderer.DrawWalkDrivenStatics.
Weather placement (S3 §9.1 R5): confirmed unchanged. GameSky's
weather pass (RenderWeather, gated on is_player_outside) already
runs after every LandCell event for both root kinds — for an
outdoor root, DrawLandscapeDynamicsPhase is called directly after
driver.Replay() completes (which processes the whole per-cell
_events list first); for an interior root with ov>0, it fires via
the LandscapeFlush leaf (RetailPViewRenderer.FlushWalkLandscape ->
_walkPreClearDynamics), and OnInteriorFloodDrawTurn only emits
LandscapeFlush AFTER DrawLandscape's per-cell loop has fully run
and recorded every LandCell event ahead of it in the same _events
list Replay walks in order. No code change needed; verified by
reading the call sites.
Tests: WalkEvents/RetailFrameWalk/WalkFrameDriver's existing pins
updated (every "TERRAIN:0" expectation deleted, matching the deleted
event); new coverage for T1 (WalkLandCellOrderTests — LandWalkOrder +
WalkLandscape.CalcDrawOrder + WalkLandscapeAssembler
.SideCellCountForRing reproduce both cathedral captures' frame-2
LC/SC sequences byte-for-byte, 533/698 arrival and 531/757 leak, cross-
verified against the retail ring-to-LOD table), T2 (a synthetic two-
block landscape with a real WalkVisibilityMath-driven out-of-view
column, proving LC-before-building/statics, far-to-near, no-LC-but-
keeps-SC for the excluded cells, and no TerrainSlice event of any
kind), T3 (TerrainLandCellIndexRunsTests — the index-run arithmetic
for every side/cell, disjoint and exhaustive over the 384-index
landblock), T4 (batching merge/split), and the RetailPViewPassExecutor
CompiledCallGraph pin retargeted at DrawWalkLandCellBatch /
DrawLandCellRuns. WalkOracleTrace gains LC/SC event kinds and a
Load(root, name) overload for the OH capture directory — the one
parser change this chunk needs, no validator, no other infrastructure.
App hermetic lane: 6,786/6,786 (6,765 baseline + 21 new). InstalledDat
lane: 241 passed, 3 accepted failures (2 pre-existing #383 layout
fixture-drift tests, 1 TowerAscent Status=KnownFailure) — unchanged
from baseline. Core terrain tests: 116/116 (Core untouched by this
chunk).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
RetailFrameWalk.DrawInside now passes _interiorPView.OutsideView.ViewCount
into IWalkEventSink.OnInteriorFloodDrawTurn. WalkFrameDriver's own
implementation reproduces PView::DrawCells @0x005a4840's exact gating
(0x005a4852-0x005a49eb, all inside `if (outside_view.view_count > 0)`):
the landscape flush (retail FlushAlphaList(0f) @0x005a4872 plus the
pre-clear dynamics hook), the device-stamp advance @0x005a4886, a GATED
depth clear (pc:432731-432732), and the exit-portal seals
(pc:432785-432786) — all four skipped entirely when outsideViewCount == 0.
The depth clear is gated on a new driver field, PortalsDrawnCount, which
models retail's D3DPolyRender::portalsDrawnCount (uint16 @0x008719b4):
read-then-zeroed at the interior root's own flood turn
(@0x005a489c-0x005a489e), and re-armed at Replay by the count
IWalkFrameLeafRenderer.DrawExitSeals now returns (the SAME portal
enumeration RetailPViewPassExecutor.DrawPortalDepthWrite already performs
— OtherCellId==0xFFFF, >=3 vertices). The field persists across frames
(never cleared by BeginFrame/AbortFrame/EndFrame/Replay), reproducing
retail's documented quirk: a fresh driver's first ov>0 frame never clears;
every later ov>0 frame clears because the previous frame's own seals armed
the counter.
_skyDrawnThisFrame — the proxy for outside_view.view_count != 0 that used
to gate the stamp re-arm — is deleted; its "second Landscape turn in one
frame" fail-loud guard moves to a frame-scoped counter
(_landscapeTurnsThisFrame). RetailPViewRenderer.ClearWalkInteriorDepth
splits into FlushWalkLandscape (pre-clear dynamics + FlushLandscapeAlpha)
and ClearWalkInteriorDepth (the Z clear only), both wired through the new
IWalkFrameLeafRenderer.FlushLandscape leaf and the WalkLeaf production
adapter.
Tests: flipped the ov==0 pin to expect no landscape-flush/clear/seals at
all (T1); added the two-frame first-frame-no-clear / armed-clear pin plus
a no-exit-portal-never-clears pin (T2); added a look-in-neither-arms-
nor-consumes-the-counter pin (T3); added RetailFrameWalk's two-PView
draw_landscape wiring pin and a WalkPView.ConstructView reset pin (T4);
updated every direct OnInteriorFloodDrawTurn caller to pass the
outsideViewCount it models (T5). The four per-category leaf-contract pins
(whole-once shell, Boolean sphere admission, portal-polygon-only clip,
local-player repeated submission) already existed and needed no additions
(B4).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The offline degrade probe proved the mechanism: every Holtburg building
carries PORT nodes ONLY in its level-0 GfxObj (out to ~24-48 m); every
degraded level has zero. Retail walks the CURRENT degrade level BSP
(part->gfxobj[deg_level]) - that is what limits look-in punches to the
nearest buildings. WalkBuilding gains the degrade ladder +
SelectDrawingBsp (band pick; UpdateViewerDistance hysteresis is a port
TODO), the walk selects per viewer distance, the adapter builds
per-level BSPs, and the stab-list load rule (CLandBlock::init_buildings
@0052fd80: a full-res block loads exactly its buildings portal stab
cells) replaces load-everything in the landscape builder. The sweep now
shows clean rosters with all far-building punches gone; remaining
deltas: the near buildings 001a/0022 (50 m/28 m center distance vs the
48 m band edge - sphere-adjusted distance/hysteresis to port) and the
one ring-1 frustum boundary pair (aab50002/a9b3003c).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
RetailFrameWalk composes the ported machinery into retail frame shapes:
SmartBox::RenderNormalMode rooting (camera-cell low word < 0x100),
PView::DrawInside + the DrawCells event half (interior pview, the
traces pv=009d4a80), LScape::draw (per-view visibility, blocks
far-to-near, per-block cells far-to-near, buildings at their cell turn)
and RenderDeviceD3D::DrawBuilding (BLD at entry before the degrade
check; two-pass BSP portal walk per active view on the outdoor pview,
the traces pv=009d4b08). WalkLandscape ports the block grid +
draw_check_blocks/landcell_check visibility (192 m/24 m pitch,
viewer-relative, union across views, never downgrade). Six composition
tests: outdoor far-to-near emission, degrade-entry event, interior
ov=0/ov=1 shapes, rooting, complete view unwind.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Erik