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>
212f5a12e5