[Performance] Batch one: skip the empty Gaussian pass, animation pause flag, lazy error names, unfair lock on component ids

[miogds]

First optimization batch from docs/proposals/PerformanceMeasurementResearch.md, stacked on #53 so the benchmark can measure it. Fork-only; nothing goes upstream until the whole effort is done.

Changes

• Skip the Gaussian pass when the frame has no splats. With an empty working set the pass still copied the opaque depth, cleared and stored the full-resolution coverage target and ran the tile initialize and post-process stages. On Apple Vision Pro (2048×1984 per eye, two eyes) that was 0.7 to 4.9 ms of GPU time per frame in scenes without a single splat, the single largest fixed cost in the first headset run. gaussianCoverageWritten stays false, so the composite and anti-aliasing passes do not read the stale map. Verified on macOS that the pass and its depth copy appear only in the Gaussian scene.
• AnimationSystem. The update loop read each entity's pause state through isAnimationComponentPaused(entityId:), which walks the entity's subtree allocating a Set and an array per node with a scene.get per node, every frame, to answer a Bool the loop already holds. It now reads animationComponent.pause; the public API is unchanged.
• ErrorHandlingSystem. The entity-taking handleError overloads formatted the entity name before the logger's level check. The message is built inside Logger.logError's autoclosure.
• ComponentPool. The component type-to-id map, read on every scene.get, sits behind an OSAllocatedUnfairLock instead of an NSLock.

Measured on macOS (Mac16,5, 1920×1080 at 120 Hz, release, three runs each side, aggregated by minimum, before the Gaussian change)

Scene	update	scene graph	culling	encode	mean frame
primitives-1k	−12.8 %	−14.5 %	−14.9 %	−6.7 %	0 (display-paced)
primitives-10k	−15.1 %	−15.8 %	−4.5 %	−6.1 %	−7.0 %
lights-64	−9.6 %	−14.1 %	−13.4 %	−8.6 %	0 (display-paced)
postfx	−10.7 %	−12.1 %	−12.5 %	−6.0 %	0 (display-paced)

The light scenes move within the run-to-run noise of a mostly idle machine. Most of that gain is the component-id lock, which every ECS access hits.

What the first Apple Vision Pro run said (baseline in #53, perf/baselines/RealityDevice17_1)

Every scene missed the compositor deadline on almost every frame. The light scenes (animation-16, stadium, gaussian) are GPU-bound at 8.6 to 9.7 ms mean GPU time for 14 to 160 visible instances: the fixed per-frame cost of full-screen passes at two eyes of 2048×1984 (Gaussian pass 3.6 to 4.9 ms, G-buffer + light 3.3 to 4.5 ms even with a dozen draws, pre-composite 1.1, output transform 0.8 to 1.9, post-processing bypass 0.4 to 2.0, environment 0.5 to 1.0, HZB and depth copies about 1). The heavy scenes are CPU-bound: 24 ms of encode for 1264 opaque plus 1407 shadow draws (primitives-1k) and 48 ms for the 10k scene, about 9 µs per draw on the headset CPU, with the eye-dependent graph executed twice. Those numbers are the case for the next items in the note: layered stereo with vertex amplification, foveation, the draw list and the scene-access refactor.

Verified

swift build; animation, camera, component-registry and custom-system core tests; GaussianRenderingTest (all but the PSNR comparison, which needs cv2 locally), GPUPassTimerTests, UpdateXRRenderingSystemTest, WholeRenderGraphValidationTest, EngineStatsMonitorTests; the benchmark runs above.