render: Re-evaluate virtual geometry with native Metal MDI

[hmans]

Context

Scrapbot currently uses portable GPU compaction for virtual-geometry submission on Metal because the pinned wgpu-native backend does not advertise MultiDrawIndirectCount. Fixed-count multi-draw is callable, but the current Metal backend may lower it to repeated indirect draws.

This is not a permanent Metal limitation. wgpu has native Metal Indirect Command Buffer work in progress:

• use Metal's Indirect Command Buffers for true GPU-side multi draw indirect gfx-rs/wgpu#9640 — fixed-count MDI through Metal ICBs
• Use Metal ICBs for multi-draw indirect count gfx-rs/wgpu#9679 — count-buffer MDI through Metal ICBs

Re-evaluate Scrapbot's choice once an applicable implementation is merged, released, and included in our pinned wgpu-native build.

Current evidence

A controlled Apple M4 / Metal comparison used the Sponza example at a fixed 1920×1080 resolution. Each mode was profiled twice in alternating order with 60 warmup frames and 240 measured frames.

Median	Portable compaction	Current fixed-MDI route	Change
CPU active	1.01 ms	7.23 ms	+615%
GPU frame	15.15 ms	17.45 ms	+15%
GPU culling	0.88 ms	1.61 ms	+83%
GPU shadows	1.27 ms	2.55 ms	+101%
GPU depth	0.31 ms	0.55 ms	+74%
GPU world	3.47 ms	3.37 ms	-3%

The current MDI route slightly improves world shading, but loses badly overall. The dominant CPU regression is consistent with fixed-count MDI ranges being expanded by the current Metal backend. The route also sends meshlet work through depth and four shadow cascades, increasing GPU cost.

A control with zero visible shadow casters still retained substantial MDI submission cost, showing that zero-instance indirect commands in the encoded ranges are not free under the current lowering.

Upstream dependencies

• Use Metal's Indirect Command Buffers for true GPU-side multi draw indirect
• Use Metal ICBs for multi-draw indirect count
• Apple: Encoding indirect command buffers on the GPU

Re-evaluation checklist

• Confirm the relevant upstream implementation is merged and present in Scrapbot's pinned wgpu-native artifact.
• Confirm adapter feature reporting distinguishes native fixed-count MDI from an emulated fallback.
• Preserve portable compaction for adapters without a genuinely native path.
• Profile identical project state, camera, resolution, render features, warmup, and measured-frame count.
• Run at least two alternating repetitions per path to reduce order and thermal bias.
• Compare raw median/p95/worst CPU and valid GPU timings with mise profile-analyze.
• Compare workload and counter fields, especially cluster command count, visible frontier, draw submissions, cull workgroups, depth, and every shadow cascade.
• Test both the ordinary camera path and a shadow-heavy case.
• Verify output with matching frame captures and virtual-geometry debug views.
• Update the rendering FDR/ADR and architecture inventory if capability selection or submission ownership changes.

Acceptance criteria

Adopt native Metal MDI only when it is a measurable overall improvement for representative virtual-geometry workloads without regressing correctness. Do not decide from the world pass alone: CPU encoding, culling, depth, shadows, and full GPU-frame cost are part of the decision.

If native ICB MDI is not consistently better, retain portable compaction and record the measured reason.