perf(rendering): VSOut shrink, invisibility cull, volume step cut

[rulkens]

Summary

Three focused perf changes targeting the large-tier point-sprite cost plus a one-line volume optimization. Measured on M-series, large tier, fully zoomed out:

	Before	After
point-sprites	24 ms	17.5 ms
scalar-volume	8 ms	3.5 ms
Total	32 ms	21 ms

~11 ms / ~35% saved.

Points pipeline

• Drop unused dMpc and camDistMpc VSOut fields (both literally unused by either fragment per their own docstrings).
• Mark tint, intensity, axisRatio as @interpolate(flat). All three are per-instance constants — smooth interpolation was wasted rasterizer work for an identical result at every fragment.
• Fold depthFade × schechterRatio × angularDensityWeight into intensity at the vertex stage. Removes 3 varyings and 3 per-pixel select+multiply chains; mathematically identical because all four factors are per-instance constants.
• Cull invisible billboards (folded intensity < 0.005) via degenerate clip-space output, same pattern as the existing Malmquist gate. The selected galaxy bypasses the cull so the selection halo never vanishes on a faint pick. Pick fragment shares the vertex stage, so culled galaxies become non-pickable — acceptable since they're visually undetectable.

Volume raymarch

• STEP_COUNT 192 → 128. Texture-cache locality at higher step counts means the per-step cost curve flattens above this knee, so dropping 1/3 of the steps recovers ~half the cost. 128 is the sweet spot before banding becomes perceptible in sparse fields.

GPU adapter

• Request powerPreference: 'high-performance' so multi-GPU systems (older MacBook Pros with dGPU, desktops with iGPU + dGPU) pick the discrete card. No-op on single-GPU machines including Apple Silicon — shipping for cross-device benefit.

DebugPanel

• Add scalar-volume to DISPLAY_SLOT_ORDER. The slot was instrumented and timed correctly but missing from the panel's display list because it isn't part of HDR_PASSES (it runs in encodeVolumes before the HDR loop).
• Use the 60-frame rolling average for the header total instead of the single-frame sum. Makes the readout internally consistent with the per-row values and stops the header from jumping frame-to-frame at sub-ms scale.

Test plan

• Reload, large tier, fully zoomed out — verify point-sprites lands around 17–18 ms (was ~24 ms).
• Toggle Rhizome (cosmic web) on — verify scalar-volume lands around 3–4 ms (was ~8 ms) and the volume looks visually unchanged (no banding in sparse filaments).
• Click on a faint distant galaxy — verify selection ring still appears even if the galaxy would otherwise be culled by the invisibility threshold.
• Toggle through bias modes (None / VolumeLimited / VMax / Schechter / AngularReweight) — verify each renders identically to before (math has been hoisted into vertex stage but is unchanged).
• Toggle Real-only mode and Highlight fallback — verify both still behave correctly (they read isFallback which is unchanged).
• Toggle Depth fade off — verify distant galaxies stop fading (depthFade gate moved into vertex but logic preserved).
• DebugPanel: verify scalar-volume row now appears, header total reads as 60f average, all rows still update at 60 Hz.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	skymap	2846aca	Commit Preview URL Branch Preview URL	May 19 2026, 11:31 PM