Summary

Renderer changes to make output correct and consistent across all resolutions. The pre-fix code worked at 1080p but silently broke at 1920×1200, 1440p, 4K, and 8K — and even where it produced output, the same world-space scene rendered very differently at different resolutions.

What was broken

1. Silent key overflow above 8192 tiles. The pair-sort key packed (tileIdx << 19) | splatIdx into a single u32. At any image with more than 8192 total tiles (≥ 1920×1200), tileIdx << 19 overflowed u32 and aliased tile indices — silent corruption with no error.
2. Hard tile-edge artifacts at higher resolutions. Splat pixel radius scales with image height (focal length × world-radius / depth). At 8K a splat that fit comfortably in the per-splat tile-coverage cap at 1080p projects to 16× more tiles. The cap was previously enforced by clamping coverage and dropping tail tiles, producing visible hard square edges throughout the image.
3. Resolution-dependent radius fade. RADIUS_FADE_START_PX = 1024 / RADIUS_FADE_END_PX = 2048 were absolute pixel values. World-space splats whose 1080p pixel radius was 256-512 px (well below 1080p's 1024-px fade-start) projected to 1024-2048 px at 8K and were silently faded out. Result: large patches of grass, sky, and foreground content present at 1080p went missing at 8K.

What changed

• Key + value radix sort. Switched from packed-key u32 sort to ComputeRadixSort.sortIndirect's key+value mode (initialValues param). tileKeysBuffer holds tile indices, splatValuesBuffer holds splat indices; both get full 32-bit range. Deletes SPLAT_IDX_BITS/SPLAT_IDX_MASK from config and shrinks sort passes at 1080p from 8 to ~4 (numBits auto-derived from numTiles).
• Sub-frame split with no truncation. Images larger than 120×68 tiles (= 1080p) are partitioned into a grid of ≤ 1080p sub-frames. Each sub-frame renders independently into a sub-frame-sized running state buffer, then is stitched into the final image. The global CPU depth sort is shared across sub-frames so there are no inter-sub-frame seams. maxCoveragePerSplat is set to the sub-frame's full tile area (rounded up to a power of two), which geometrically bounds every splat's bbox-in-sub-frame — so the project shader's coverage clamp never bites and emit-pairs writes every splat's full bbox. No hard tile edges anywhere.
• Image-height-relative radius fade. Replaced RADIUS_FADE_START_PX / RADIUS_FADE_END_PX with RADIUS_FADE_START_FRAC / RADIUS_FADE_END_FRAC. The project shader computes the per-render threshold as frac × imageHeight, so the SAME world-space splats fade at every render resolution. After this fix the 8K-downsampled-to-1080p image matches 1080p direct at SSIM 0.93.
• Right-sized running-state and output buffers. Buffers are sized on groupTilesX × groupTilesY × TILE_SIZE² instead of groupSizeTiles² (the bounding-square of the larger axis). Saves ~25 MB at 1080p, more at non-square aspect ratios and at higher resolutions.
• Dropped multi-slot plumbing. numSlots / Slot[] / per-slot shader names removed; rasterizer holds a single PipelineBuffers. The "group" abstraction is retained — it's the hook the sub-frame split uses (beginGroup / dispatchChunk / finishGroup with a group origin and tile rectangle, the project shader's group-AABB cull, and group-pixel-origin uniforms).
• Indirect-dispatch slot pool bumped to 256 K. PlayCanvas resets the device's indirect-dispatch slot counter at frame-end only; in an offline pipeline the counter monotonically grows across (chunk × sub-frame) iterations and at 8K can hit ~70 K slots. Each slot is 12 bytes so 256 K = 3 MB, cheap.
• Centralised render tunables. AA_DILATION_COV, SIGMA_CUTOFF, DISCRIMINANT_FLOOR, OPACITY_CAP, MIN_ALPHA, MIN_TRANSMITTANCE, JACOBIAN_LIMIT_FACTOR, TILE_SIZE, the radius-fade fractions, the pair-buffer memory budget, and the MAX_COVERAGE_PER_SPLAT_AT_1080P baseline now all live in src/lib/render/config.ts with documentation. gaussian-aabb.ts reads SIGMA_CUTOFF from there too so the BVH extent and the projected splat radius stay in sync.
• Golden-image regression tests. Added test/render-golden.test.mjs with tiny (320×240) and mid (640×360) fixtures shared between the runner and the regenerator via test/render-golden.cases.mjs. Byte-exact compares against checked-in webp goldens.

Verification

• npm run build clean, no TypeScript warnings.
• node --test test/render-golden.test.mjs — both fixtures pass byte-exact (tiny and mid).
• npm run lint — 22 errors, all pre-existing on main; no new errors introduced.
• Windmill scene (17.9 M splats) rendered at 1080p, 1440p, 4K, and 8K. All produce visually correct output with no tile artifacts, no seams, no missing content. The 8K-downsampled-to-1080p matches the 1080p direct render at SSIM 0.93 (the remaining ~7 % gap is residual SSAA sharpness from AA_DILATION_COV / DISCRIMINANT_FLOOR still being in absolute pixel units, which is reference-3DGS behaviour).

Performance and memory

• 1080p windmill: 1.08 s → 5.9 s render. Slower because chunkCap is forced down by maxStorageBufferBindingSize once maxCov covers the whole 1080p sub-frame area (16 384 tiles). Peak GPU memory unchanged at 2.20 GB.
• 8K windmill: 54 s render; previously this resolution silently produced corrupt or content-missing output. Peak GPU memory 3.31 GB (dominated by the image-wide 530 MB running state, 130 MB output, and per-chunk pair buffers).

Caveats / known follow-ups

• AA_DILATION_COV and DISCRIMINANT_FLOOR are still absolute pixel units. Making them resolution-invariant would close the residual 7 % SSAA gap but is an "exceed reference 3DGS" decision and a separate change.
• Project shader re-runs per sub-frame. Sharing projection across sub-frames within a chunk (project once, bin per sub-frame) would cut 8K render time by roughly 4–8× — a sizeable refactor, deferred.
• Prefix-sum is single-workgroup. Per-thread element budget scales linearly with chunkCap; current scan is the placeholder from the rework plan's Open Question 1. Two-level Blelloch is the natural follow-up if chunkCap ever grows past ~1 M.
• No seam-N regression fixture. The rework plan called for a synthetic scene engineered to span the original 128-px tile-group boundaries; not added. The current tiny + mid fixtures plus byte-exact passing under forced sub-frame split (verified locally with 7×5 split — still byte-exact) cover the boundary case adequately for now.
• FAR_PLANE_NEAR_FACTOR is unused after the BVH removal; left in config for now.

Full reasoning trail and design alternatives considered are in the saved review at ~/.claude/plans/review-the-changes-committed-mutable-tiger.md.