Fr0zkaandClaude Opus 5.5
9898e057fd
Linux offscreen capture (EGL/GLES pbuffer; llvmpipe in the sandbox) and an opt-in screen-space gate for LOD>=3 (normal distance projected at each LOD's closest display distance, 1 px at 1440p = 0.61 mrad). The gate rejects the S13 octave-drop and stride-2 controls, fault injections, and a vertical-vs-screen control. Finding: today's far meshes differ from the field by up to 92 px, and LOD swaps snap mesa rims by tens of px (p50 0 px), a silhouette effect of plain LOD switching. The octave cutoff was dropped: nothing qualified, and octaves are a small share of far time. Off-path cost: reviewer interleaved W12 within noise vs L4 (+0.6% median). Hashes = L4. Largest far lever measured: column work outside noise, 34% of far worker time (duplicate XY across tiles). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ED33VnqepD4YY6gDAAA832
103 lines
8.7 KiB
Markdown
103 lines
8.7 KiB
Markdown
# Linux6 report
|
||
|
||
## Result
|
||
|
||
The screen-space gate rejects today’s far meshes at every sampled LOD3–8: the largest measured same-field error is 92.0 px, and the largest sampled swap pop is 70.3 px at LOD6→5. Same-LOD regeneration is exact, the reported swaps have no cracks or floaters, and no display mesh changed in this round.
|
||
|
||
Source base: `wip/l5-far-cutoff` at `b21a19d`; L4 reference: `7770074`.
|
||
|
||
## Screen-space method
|
||
|
||
For one surface tile nearest the L1 camera at each LOD, the gate rasterizes triangles on a 65×65 XY grid, excluding skirts. It compares the mesh hit to direct `Heightfield2D` ground/cap roots; slopes use ±1-voxel finite differences. For vertical residual `Δz` and slopes `sx,sy`, it computes normal distance `|Δz|/sqrt(1+sx²+sy²)`, projects that normal onto the camera’s screen plane, then divides by the one-pixel world distance at that LOD’s closest display distance. The view uses 90° horizontal FOV, 2560 px, yaw 0°, pitch −22°; one pixel is 0.000613592 rad. Same-LOD regeneration is the mesh reference for swaps; direct field roots are the reference for true surface error.
|
||
|
||
The known failure control is a 1 m cell on an 89.427° plane, shifted sideways by 5 mm (0.005 cell): its vertical metric says 8.488 px, while its true normal distance is 5 mm and projected screen error is 0.0849 px. This confirms that vertical height can badly overstate visible displacement.
|
||
|
||
| LOD | Closest distance (m) | Field error p50 / p99 / max (px) | Max normal distance (m) |
|
||
|---:|---:|---:|---:|
|
||
| 3 | 96 | 0.81 / 4.16 / 6.00 | 0.423 |
|
||
| 4 | 192 | 1.01 / 5.12 / 7.93 | 0.952 |
|
||
| 5 | 384 | 0.79 / 16.37 / 92.02 | 25.122 |
|
||
| 6 | 768 | 0.51 / 9.75 / 70.34 | 37.644 |
|
||
| 7 | 1,536 | 0.29 / 9.04 / 42.36 | 42.714 |
|
||
| 8 | 3,072 | 0.83 / 18.03 / 47.18 | 93.175 |
|
||
|
||
| Swap | Distance (m) | Pop p50 / p99 / max (px) | Max normal distance (m) |
|
||
|---|---:|---:|---:|
|
||
| 3→2 | 96 | 0.61 / 3.77 / 6.05 | 0.367 |
|
||
| 4→3 | 192 | 0.75 / 4.63 / 7.51 | 0.896 |
|
||
| 5→4 | 384 | 0.00 / 9.58 / 31.78 | 8.139 |
|
||
| 6→5 | 768 | 0.00 / 9.75 / 70.34 | 37.644 |
|
||
| 7→6 | 1,536 | 0.00 / 9.04 / 42.36 | 42.714 |
|
||
| 8→7 | 3,072 | 0.83 / 18.03 / 47.18 | 93.175 |
|
||
|
||
The tails are not confined to cliffs. The field maximum slopes at LOD5–8 are 8.34°, 14.09°, 6.76°, and 9.33°; LOD5’s p99 is 19.8 px below 30° and 21.1 px at 45–60°. The sampled LOD6→5 maximum is at `(-472, -1432, -463.151) m`; [the marked offscreen capture](captures/lod6_to_5_max_pop.jpg) shows that location at the 768 m swap distance. This is a sampled-tile result, not a whole-world supremum.
|
||
|
||
All same-LOD mesh comparisons at LOD3–8 pass with zero error. The S13 octave-drop and corrected stride-2 column-interpolation controls are rejected at every LOD3–8, as are missing displayed surface, 2 m floating surface, and missing fine-ring injections. The measured swap rows have zero cracks and floaters; they reject solely on displacement.
|
||
|
||
## Far worker time, W12
|
||
|
||
The L1 view has 2553 tiles. Clean-run phase totals for LOD3–8 are shown below. A separate temporary timer measured complete `FractalNoiseQ16` calls; “column outside” is clean column time minus that subtotal. The timer added 3–6% to worker time, so the split between those two column subphases is approximate; it has been removed from the final tree.
|
||
|
||
| LOD | Worker (s) | Column outside noise (ms) | Fractal/noise calls (ms) | MC (ms) | Ground refine (ms) | Cap refine (ms) | Skirts (ms) | Assembly (ms) |
|
||
|---:|---:|---:|---:|---:|---:|---:|---:|---:|
|
||
| 3 | 1.084 | 171 | 276 | 473 | 0 | 0 | 118 | 33 |
|
||
| 4 | 2.581 | 1,000 | 534 | 440 | 0 | 304 | 196 | 102 |
|
||
| 5 | 1.829 | 658 | 358 | 100 | 464 | 22 | 161 | 70 |
|
||
| 6 | 2.761 | 1,608 | 284 | 59 | 385 | 58 | 162 | 70 |
|
||
| 7 | 1.192 | 234 | 198 | 32 | 275 | 202 | 153 | 96 |
|
||
| 8 | 2.224 | 264 | 469 | 256 | 424 | 299 | 333 | 177 |
|
||
| **Total** | **11.671** | **3,934** | **2,119** | **1,360** | **1,548** | **885** | **1,123** | **548** |
|
||
|
||
Column work outside the fractal calls is the largest single phase (33.7% of far worker time); all column work is 51.9%. Ground and cap refinement total 2.43 worker-s, skirts 1.12 worker-s, MC 1.36 worker-s, and assembly 0.55 worker-s. The earlier 7–13% profile timed individual Perlin calls; the L6 subtotal times complete fractal calls, including their octave accumulation.
|
||
|
||
The listed subphases total 11.52 worker-s; about 0.15 worker-s remains in timer and per-tile residuals.
|
||
|
||
The overlap scan counted about 411,000 cross-worker neighboring duplicate XY evaluations at LOD3–8 (about 166,000 at LOD4 alone). Exact column reuse is the largest plausible lever because it targets the largest phase. I did not build it: no candidate has yet shown a screen-gated, greater-than-3% whole-view win. Stride-2 interpolation is not a safe substitute; its control fails at every far LOD.
|
||
|
||
## Cost with the feature off
|
||
|
||
Three interleaved W12 runs per binary, same L1 camera and forced generation. `whole_view_ready` is the comparison metric; the six whole-view hashes matched exactly.
|
||
|
||
| Binary | Runs (s) | Median |
|
||
|---|---|---:|
|
||
| L4 reference | 2.082, 2.180, 2.183 | 2.180 s |
|
||
| L6 cleaned tree | 2.127, 2.185, 2.223 | 2.185 s (+0.2%) |
|
||
|
||
Run-to-run spread was about 0.10 s, so the 5 ms median difference is within noise. The no-op visual cutoff contract, plan, and CLI flag were removed. The opt-in research gate retains its negative controls; skirt timing uses the existing optional stage-timing pointer, which is null on the default path. Governor stayed `powersave`; runs used CPUs 0–11, W12, 6 GiB address-space limits, and normal priority. Peak runtime RSS was 403 MiB; peak build process RSS was 874 MiB (GCC).
|
||
|
||
## Determinism and hashes
|
||
|
||
Clang 19 `-ffp-contract=off` W1/W12 and GCC default W1/W12 produced the same hashes. LOD0–2, exact queries, and all whole-view hashes equal L4; no recipe bump is needed.
|
||
|
||
- Recipe / plan / manifest: `83b616a3f4fe781d` / `1ca84619c8600ee5` / `e55991385bf0ca29` (recipe version 1).
|
||
- Whole view (density / mesh / payload / topology): `c10697a8f48380c7` / `e75c9119a1498cd5` / `49a9a8d1835a6c8c` / `edcb23288b5e8724`.
|
||
- Full detail (density / mesh / payload / topology): `0ee0579bd15cbbe4` / `32e5d4964d610fee` / `1932b96ee4b1d910` / `69743c6bad3447dc`.
|
||
- Exact queries (density / unclamped / detail / collision): `e4d4a8eb95b35933` / `e4d4a8eb95b35933` / `efcba20c0c70c415` / `e604823a249029bf`.
|
||
- SHA-256 L4 binary / Clang L6 binary / GCC L6 binary: `6efb8f19a41ca3cd23aca9294ceee1e9cdd1d3dc8da23b3ee61ef8ae943aa763` / `3f5a4f88b0d911e8e66af9ef72e294926a5507a73d44fcfe868df27c666bf01e` / `de872e8fdaa9f40e8a6c096700fd998bcf35f232b06f2b4e35a374b4b0e9fa07`.
|
||
- Capture JPEG SHA-256: `483b434f3ad752dd09b82cd1c78fe81ee8cd31803ecf4dcd3c2918a8ae5d66b1` (156 KiB).
|
||
|
||
## Failed or retried runs
|
||
|
||
- First Clang build failed on a missing view argument in `LinuxScreenSpaceGate.h`; fixed and rebuilt successfully.
|
||
- First gate pass paired extra mesh hits incorrectly and reported same-LOD floaters at LOD5–8. The reference pairing was corrected; the final gate reports zero same-LOD displacement and no floaters.
|
||
- A `/usr/bin/time -v` wrapper attempt exited 127 because that binary is unavailable; the runs were retried with `/proc` RSS monitoring.
|
||
- The final quality gates exit 6 by design because measured displacement exceeds the 1 px rule. Builds, captures, determinism runs, and cost runs completed successfully. EGL could not write `~/.cache` in this environment, disabled shader caching, and rendered successfully on llvmpipe.
|
||
|
||
## Note for the owner
|
||
|
||
Today’s `assembler/world` remains at L4; this sampled view measures up to 92.0 px from the field and a 70.3 px LOD6→5 pop at the tested display distances, with the largest point marked in the capture. This round changes no default mesh or recipe output, removes the no-op cutoff, and leaves W12 cost within run-to-run noise.
|
||
|
||
## Orchestrator review (Claude)
|
||
|
||
- Cost re-run by me (interleaved W12, L1 view): L4 2.098/2.138/2.171 s vs L6 2.116/2.150/2.197 s
|
||
(median +0.6%, within noise), density `c10697a8f48380c7`; the L6 binary SHA matches the report.
|
||
Merged into `assembler/world` with L5 squashed in.
|
||
- The marked LOD6→5 maximum sits on the **rim of a flat-topped hill (mesa edge)**, not on open gentle
|
||
ground: the "8–14° slope" is the field slope at the sample, on the plateau top beside the drop. The
|
||
coarse mesh rounds the rim by about a cell (16–32 m at LOD5–6), so the pop is a real **silhouette snap
|
||
at sharp rims** (tens of px at 768 m), while p50 stays at 0 px. That is ordinary LOD-switch behaviour
|
||
without geomorphing or cross-fade, so a fix belongs in rendering (Unreal), not in generation. The
|
||
owner decides whether it is visible enough to matter.
|
||
- Next speed lever, per the far breakdown: exact cross-tile column reuse (column work outside noise is
|
||
34% of far worker time; ~411k duplicate XY evaluations at LOD3–8).
|