Measured (§9.7): the connected R6 soak at e10765aa showed the round-1
batching fix (§9.6 F2) was inert in practice — CPU p50 flat at 31.2 ms
and FPS flat at 32 across every outdoor destination (baseline 2.0-15.5 ms
/ 51-452 FPS), GPU p50 2-4x baseline. Root cause: nearly every admitted
land cell carries scenery statics, so a StreamMark (recorded whenever the
walk's ordered stream grows) follows almost every LandCell event and was
itself treated as a flush point, splitting the pending terrain batch back
down to ~500 submissions/frame — each paying TerrainModernRenderer's full
bind sequence plus a ring allocation and an MDI.
F10: a StreamMark no longer flushes the pending terrain batch. Citation —
the walk's ordered stream holds ONLY opaque batches
(WalkStaticStreamPopulator.cs:179, `if (batch.IsOpaque)` routes translucent
batches to the alpha list instead of the stream) drawn through the
Opaque/OpaqueAlphaToCoverage pipelines, both created with depth test AND
write ON and GpuBlendMode.None (WbDrawDispatcher.Rhi.cs:240-248,343:
`Depth = new GpuDepthState(Test: true, Write: depthWrite, depthCompare)`
with depthWrite: true); DrawOrderedRange resolves every walk-stream
command's bucket to PipelineBucket.Opaque because every command it holds
is opaque (WbDrawDispatcher.OrderedStream.cs:569-571). Opaque terrain and
opaque statics are therefore depth-resolved identically in either
submission order, so deferring the terrain batch across a StreamMark is
pixel-identical (except exact z-ties, which retail itself leaves
order-independent).
Flush points that REMAIN, because their GPU order against terrain IS
observable: PunchFan (DEPTHTEST_ALWAYS + write — the far-Z stamp this
chunk's interleave exists to order correctly), AlphaBarrier and
LandscapeFlush (translucent drains), ClearInteriorDepth, ExitSeals, Sky,
CellShell, a particle turn whose cell has a renderable emitter, and the
end of Replay. AlphaSubmitMark stays a non-flush point (unchanged — it
only enqueues into the CPU alpha list). No ordered-stream stage found
that blends or disables depth write, so no amendment to the contract was
needed; the Replay doc comment records the citation so a future stage
that does must become a flush point again.
F11: no per-batch bind latch added (tens of batches/frame is acceptable).
F12: the F4(c) driver pin is re-expressed to prove a StreamMark from a
cell with real content (not just an empty particle turn) no longer
splits, while PunchFan/AlphaBarrier/a has-emitter particle turn still do;
a new punch-order pin (LandCell(far), building turn with a PunchFan,
LandCell(near)) proves the far terrain flushes before the punch and the
near terrain starts a fresh batch after it. Hermetic App lane: 6,796
passed (6,795 base + 1 new fact), 0 failed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
896c63fe9a