Renderer: extend two-phase GPU HZB occlusion culling to the Deferred path #486

Problem

The two-phase GPU-driven HZB occlusion cull shipped in #431 (GPUDrivenOcclusionPass)
is Forward / Forward+ only. In the Deferred path, occlusion-culled instanced
static batches stay on the legacy single-phase cull: they are frustum + HZB tested
against the previous frame's pyramid at submission and drawn through the ScenePass
G-Buffer bucket. That is correct but not disocclusion-corrected — on fast camera
moves / disocclusions, an instance that was occluded last frame but is visible this
frame can pop for one frame, exactly the artifact the two-phase scheme removes for the
forward path.

The hardened occlusion shader (InstanceOcclusionCull.comp) is already shared, so the
deferred single-phase path benefits from the UE5.8-conformance work (point-block gather,
prev-transform reprojection); what's missing is the second pass (rebuild HZB from
this frame's depth → re-test the reject list) in the deferred context.

See docs/analysis/ue5.8-hzb-occlusion-analysis.md
§7 row 17 — the one remaining deliberate scope cut from #431.

Why it matters (Tier 2 — performance ceiling)

Deferred is the path the dense-instancing genre content (foliage / debris / props on
buildings + terrain) will actually run in. Bringing it to parity with the forward
two-phase removes the one-frame pop there too, and makes "HZB occlusion culling" a
path-agnostic claim rather than a Forward-only one.

Proposed approach

Extend GPUDrivenOcclusionPass (or a deferred sibling) to draw the phase-1 survivors
and phase-2 recoveries into the G-Buffer, mirroring the established
DeferredOpaqueDecalPass integration (the precedent for a pass that mutates the
G-Buffer between ScenePass and DeferredLightingPass):

• Hold the GBuffer (per the pipeline wiring), target GetFramebuffer() /
  GetSamplingFramebuffer() correctly for the MSAA vs resolved cases.
• Declare the G-Buffer attachment writes (albedo / normal / emissive + their MS
  variants + velocity-MS + depth-MS, like DeferredOpaqueDecalPass::Setup) so the
  graph orders DeferredLightingPass after this pass and the hazard validator passes.
• Build the mid-frame HZB from the G-Buffer depth; run DispatchPhase2 against it.
• Re-export SceneDepth (depth-only is acceptable to start; the G-Buffer normal RT is
  packed, so a normal re-export is a follow-up nicety).
• Drop the Path != Deferred gate on the submission reroute
  (Renderer3D::SubmitGPUCulledInstanced); confirm deferred instanced batches resolve
  to the G-Buffer shader (DrawMeshInstanced currently does not swap to
  PBRGBufferShader for instanced — verify / fix as part of this).
• Register the pass before DeferredOpaqueDecalPass / DeferredLightingPass in the
  deferred topology.

Spike first (confidence 0.5): prototype the G-Buffer draw + a deferred visual test
before committing — the MSAA / per-sample-resolve interaction and the deferred instanced
shader selection are the unknowns that can make this "tests green, screen wrong."

Acceptance criteria

• Two-phase occlusion runs in Deferred: phase-1 vs previous HZB → reject list → mid-frame
  HZB → phase-2 recovery, drawn into the G-Buffer before lighting.
• A deferred full-pipeline visual-evidence test (RendererAttachedTest, Path = Deferred, dense instanced field behind an occluder) proves no corruption / no holes
  and a measurable survivor reduction — mirroring OccludedInstanceFieldScene.
• No regression to the forward two-phase path or to non-instanced deferred rendering;
  render-graph hazard validation stays clean.
• AO/SSR see the instanced depth in Deferred (depth re-export), matching the forward
  Stage-3 behaviour.

Score

capability: 3
craft: 3
stability: 2
decay: 1
effort: 5
confidence: 0.5
learning: 3
fun: 3
kano: performance
blocked_by: []
blocks: []

_{Rated per issue-scoring · score = confidence × (capability + craft + stability + decay) / effort, derived by the picker.}