Render policy and device capabilities for SplatMetalView

[Xget7]

Summary

• SplatMetalView.renderPolicy and deviceCapabilities, with SKRenderPolicy, SKDeviceCapabilities and -[SKSplatEngine applyRenderPolicy:reason:warnings:].
• The shared engine (resolveRenderPolicy, SplatEngine::setRenderPolicy) re-validates each request on the render thread.
  Metal applies the sort depth with GPU sort and the sub-pixel threshold under tight culling; every other field falls back with a warning, and an invalid request keeps the previous policy.
• Removes SPLATKIT_METAL_MIN_PIXEL_RADIUS and SPLATKIT_METAL_DEPTH_KEY_BITS; set renderPolicy instead.
• Benchmarks log 30-second windows and final p99 frame and GPU times, and reject durations outside (0, 3600] seconds.
• Adds CHANGELOG.md; the duplicate packages/splatkit-ios/distribution/ copies of README, CONTRIBUTING and AGENTS are removed, since their links resolve only from the root.

Exported from the SplatKit monorepo at 6237bb8 with scripts/export-ios-source.py --platform ios.

Validation

• Engine tests (176) and the Metal suite (215) passed on a Mac host.
• Python harness and export tests passed.
• Not run on an iPhone yet; the React Native iOS adapter compiles against the simulator SDK only.

Before merging

• Run the dev app on an iPhone with a policy change.
• Package.swift still pins the 0.1.0-alpha.2 XCFramework, so renderPolicy stays unreleased until an alpha.3 binary is published.

[Xget7]

iPhone run

iPhone 17 Pro (A19 Pro), physical device, Release build of apps/ios-dev against static libraries built from monorepo 6237bb8 plus a dev-app log of the applied policy and capabilities.
World kitchen_500k.spz, fixed pose, capture after 8 s, 1206x2622.

Launch args	Applied policy	Drawn of 500,000	Capture vs baseline
(default)	sortDepth 32, subpixel 0.5	96,029	repeat run: 0.002% of pixels differ, max 3/255
--depth-key-bits 16	sortDepth 16	96,029	PSNR 43.1 dB, 0.89% of pixels over 8/255, edge ordering only
--metal-culling 1	subpixel 0.5	73,906	reference for the rows below
--metal-culling 1 --min-pixel-radius 1.2	subpixel 1.2	73,893	PSNR 65.2 dB
--metal-culling 1 --min-pixel-radius 3	subpixel 3	73,648	PSNR 64.0 dB
--metal-culling 1 --min-pixel-radius 8	subpixel 8	70,556	PSNR 42.1 dB

• The renderer logged Metal policy: 16-bit sort keys and each sub-pixel radius, so both applicable fields reconfigure Metal on device.
• Capabilities report sortDepth supported and subpixelThreshold supported only with the culling experiment on, matching the changelog.
• World upload took 140 ms and the first frame was ready at 180 ms; no policy rejections or GPU errors were logged.
• Not covered on device: rejected and preparation-failed outcomes (engine tests only) and sustained frame timing.

[Xget7]

iPhone heavy run: ISS

iPhone 17 Pro (A19 Pro), same build as above.
iss_10M-sse-depth10-sh1.lodsplat (10M splats, 11,744,642 resident nodes), 2.2M selection capacity, SH 1, 1206x2622, horizontal 45 m orbit at 4 degrees/s from 90 degrees, --metal-culling 1 --tile-raster 1, 30 s after the first frame, back to back.
The sort depth and sub-pixel radius now come from renderPolicy, not the removed environment variables.

Run	Policy	FPS mean (min-max)	GPU ms	Sort ms	Drawn mean	First frame
1	sortDepth 16, subpixel 1.0	22.3 (19.0-29.0)	66.4	23.3	1,958,282	3.69 s
2	sortDepth 32, subpixel 1.0	19.2 (17.5-25.3)	86.2	39.8	1,959,609	1.92 s
3	run 1 repeated	23.3 (21.3-30.8)	57.5	21.8	1,957,768	3.18 s
4	sortDepth 16, subpixel 0.5	23.2 (21.3-30.6)	63.8	22.7	2,021,273	1.93 s

• 16-bit keys cut sort time from 39.8 ms to 22-23 ms and raise the mean from 19.2 to 22.3-23.3 FPS, in line with the 2026-09-12 environment-variable A/B (18.9 to 23.4 FPS).
• Lowering the sub-pixel radius from 1.0 to 0.5 px draws 63,000 more splats.
• Every run logged the requested Metal policy line and stopped on duration-complete.
• Sampled peak footprint was 2,827,848,480 to 2,828,651,296 bytes across runs (2026-09-12: 2,827,044,672); footprint after each run stopped was 1.08 GB, or 1.35 GB in the 32-bit run.
• Thermal state stayed fair, and no GPU errors or policy rejections were logged.
• The signal 14 after each run is the test harness's timeout closing the console, after the app had already stopped.
• Short runs, not sustained benchmarks; the 30 FPS target at 2.2M is still not met.