The dense-Arwic re-measure crashed with VK_ERROR_DEVICE_LOST: between
ordered ranges the walk leaf draws (terrain, shells, sky, punch fans)
and RetailAlphaQueue flushes rebind the SAME set-0 storage slots to
their own sections, so the bind-once latch made the next range draw
against foreign buffers - out-of-bounds instance reads and a GPU
fault. Sections now re-bind on every DrawOrderedRange call, exactly
like the proven DrawPreparedAlphaBatchRhi; the once-per-frame ring
WRITES in PrepareOrderedStream (the actual measured cost) are
unchanged. The bind-once referee test flips to assert per-range
rebinds with unchanged draw coverage.
Suites: full Release build 0 warnings; hermetic 6,758/0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The FW3.4 dense-Arwic pair triggered the +/-20% stop rule (+33.5% CPU
p50, 14x frame allocation). This slice removes the three measured
costs without changing GPU command order (the referee suites assert
identical recorded call sequences):
- WalkFrameDriver: Collect (ONE walk per frame - no GPU work; leaf
calls and flush points become a recorded event list; the driver
absorbed the renderer collection pass and exposes the visited sets)
+ Replay (prepare the whole stream once, then replay events,
interleaving DrawOrderedRange with leaf calls in the exact recorded
order). RunFrame = Collect+Replay for existing callers.
- WbDrawDispatcher: SubmitOrderedStream split into PrepareOrderedStream
(all sections + commands + merge runs uploaded once per frame) and
DrawOrderedRange (bind-once latch; per-run pipeline + DrawIdOffset +
DrawIndirectRangeRhi). Load-bearing correctness catch from the
implementation round: merge runs take FORCED BREAKS at the recorded
event marks - whole-stream merging must not fuse two segments that
retail separates with a leaf GPU call (shell, punch); the straddle
assert stays as a dead-code safety net.
- WalkProductionWorldData: WalkFrameStaticRecords carries an
ArraySegment into a per-frame grow-only arena; the per-cell
fresh-array copies (the 1.9 MB/frame alloc p50) are gone - zero
steady-state allocation after warmup.
Suites (lead-verified): full Release build 0 warnings; hermetic
6,758/0; Walk lane 209/1; InstalledDat Walk conformance 40/1
untouched. Next: the dense-Arwic re-measure against the same-session
baseline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The retail frame walk now drives every production static draw. In
RetailPViewRenderer.DrawInside, when the concrete executor + the
packed product + the FW3.1 walk registries are all wired (all
production compositions - anything less throws):
- A pre-walk events-only collection pass (the shadow sink generalized
to WalkVisitedSetCollector) gathers the frame's visited cells,
buildings, and landscape-cell turns; the visited cells union into
prepareCells so EnvCellRenderer prepares every shell the driver
draws.
- DrawWalkDrivenStatics runs the WalkFrameDriver over the production
world data (WalkProductionWorldData over RenderSceneQuery + the
building registry): sky, terrain slices, outdoor statics at their
landscape-cell turns, buildings (alpha barrier -> punch/look-in
passes -> shell) in retail order, interior clear+seals as leaf
closures (the old tail block's drain reasoning moves with them),
flood cells shell-then-contents. Landscape/cell-stage particle
owners re-source from the walk's visited sets - retail gates
particles per cell turn (ShouldDrawParticles @0x0050FE60), which
this is; the old sphere filter was the approximation.
- DrawLandscapeDynamicsPhase + DrawBuildingLookInDynamics carry the
dynamics-only remainder (LookInObject now dynamic-classified,
late outside-dynamics + weather, particle unions); DrawDynamicsLast
and the outdoor flush are unchanged.
- The product builder stops emitting LandscapeOutdoorStatic /
LandscapeBuildingShell / CellStatic (methods deleted, dead index
tracking removed); LookInObject loads cells with
includeStatics: false.
The old static path survives ONLY behind !walkActive for the
standalone/diagnostic executor-fake path that keeps 15 retail-ordering
regression tests exercising the barrier/punch/seal machinery; no
production composition can reach it. Its deletion is FW4 scope (the
plan's "deleting the patch apparatus") - recorded in the plan.
Transitional risks recorded in code/report: the two-pass walk cost
(FW3.4 measures), the interior slice-count reconciliation between the
old clip assembly and the walk's own exit-view survival, and the
outdoor merged-flood punch coverage now riding the walk's own
building-BSP punches (retail-faithful per FW1; the owner visual gate
verifies).
Suites (lead-verified): full Release build 0 warnings; hermetic
6,750/0 (baseline minus the three deleted route tests); Walk lane
201/1; InstalledDat Walk conformance 40/1 untouched. The two
IL-branch tests the implementation round reported failing pass in
every lead run - the recurring parallel-load flake pair.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The piece that turns walk-visited static content into draws, with no
production frame wiring (FW3.2b roots the frame):
- TryClassifyBatch: ONE shared per-batch classify core (the #426
untextured gate, #188 opacity promotion, texture resolve, foliage
classification, in the exact original order) extracted from
ClassifyBatches; the classic and packed classifiers now call it -
behavior-identical, proven by the full hermetic + InstalledDat +
Core Wb suites.
- ClassifyEntityForWalk / WalkClassifiedBatch: the per-entity seam
yielding per-batch keys + instance data WITHOUT InstanceGroup
bucketing, plus the per-part selection data (picking stays alive on
the walk path - the survey's unlisted-consumer fix).
- WalkStaticStreamPopulator: per-entity walk-ordered opaque appends
(under depth Less, opaque order is pixel-relevant only for coplanar
surfaces, which retail resolves first-drawn-wins in ITS order -
never material-grouped), translucent instances to the SAME
RetailAlphaQueue via SubmitWalkAlphaInstance (identical viewer
distances; walk-order submission improves retail's tie fidelity),
selection parts published per entity.
- SubmitOrderedStream now owns _orderedDrawCullModes, retiring the
FW2-recorded alpha-scope interleaving constraint;
DrawIndirectRangeRhi takes an optional cull array (all existing
call sites unchanged). The referee test was verified to FAIL
against the old shared-scratch behavior.
- WalkDrawStage.OutdoorStatic added for the landscape turn.
Suites: full Release build 0 warnings; Walk lane 195/1 skip;
hermetic 6,747/0 (the two failures the implementation round reported
were transient - both pass in isolation and in the full run).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The walk-order submission layer over the existing RHI (plan section FW2):
- OrderedDrawStream: append-only walk-ordered draw commands
(GroupKey + transform + per-instance data + WalkDrawStage + cell
provenance), struct-of-arrays with one lockstep Reset (#193 shape).
The PortalPunch stage exists but has no FW2 submission path - the
submitter throws on it; punch emission lands with FW3 wiring.
- WbDrawDispatcher.OrderedStream partial: per-instance-first emission
(the deferred-alpha shape - command i owns instance i, walk order
survives into the indirect array), each SSBO section written once,
then one DrawIndirectRangeRhi call per maximal merge run. Runs are
built by pure-CPU BuildOrderedMergeRuns and may never span a stage,
pipeline-bucket, or cull boundary; ValidateMergeRun re-checks every
emitted run and throws (the campaign fail-loud rule). Nothing is
sorted, reordered, or dropped: N commands in, N indirect commands
out, covered exactly once.
- WorldDepthContract: retail world depth verified verbatim from the
decomp - Render::zfuncVal @0x00820e1c = 0x2, SetDepthBufferMode
@0x005a2d10 writes the enum directly as D3DRS_ZFUNC so the value IS
D3DCMP_LESS, applied by the surface-state applier @0x0059c80a with
Z-write toggled by blend; the LESSEQUAL sites are GameSky::Draw-local.
Seven world pipeline sites now cite the named constant (no value
changes).
- Plan updated: FW1 status block + gate amendment (the ten pose-stamped
retail traces supersede re-expressing the old-builder replay
fixtures; those retire with the old builder at FW4 and their
scenario classes re-verify at the FW3/FW4 connected gates).
Known FW2 scope notes recorded in the code: the building-detail
overlay replay is production wiring (FW3); the _drawCullModes scratch
may not interleave with a mid-flight RetailAlphaQueue scope (FW3
sequencing constraint). The pixel A/B equivalence proof rides FW3's
cutover toggle where a walk-driven scene first exists.
Suites: full Release build 0 warnings; Walk lane 154/1 skip;
hermetic 6,714/0 (+27 new).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
