Fr0zka/VoxelForge
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

passages: two-pass swept-tube tunnels (owner design, L99, v215) - pending owner visual review

Browse Source 
Route plan unchanged. Per tile, route segments whose bounded influence (tube + 4 m) touches the
tile write min-combined tube distance (base sweep); relief/floor/mouth shaping only in the shell
(detail sweep); 0.5 m edge blend into the strate field. Near-ready @W12: SW 1.01->0.80,
FI 2.23->1.43, TN 2.18->1.37, VS 1.09->0.98 s; whole view FI 3.38->2.37 s. All passage walk
audits pass (21,425/21,425 canonical, 3/3 interfaces); 24-run Clang/GCC W1/W8/W12 determinism;
0 sign flips. Change: mean 0.02 cm / 0.05 deg; max 25.9 cm (passage walls/roofs ~3.5 m from
route, LOD2) and 49.8 deg (rock walls 13-25 m from routes, LOD2). Captures with diffs in
Linux99/captures. Fast tier PASS (v215 pins).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
This commit is contained in:
Fr0zkaandClaude Opus 5.5 committed 2026-10-28 14:38:45 +01:00
1 parent 36140b573f
commit fa6363b31a
67 files changed
+2022 -45
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

Standalone/Research/WorldAssemblerM3/Linux99/REPORT.md
+114
View File
@@ -0,0 +1,114 @@
# Linux 99 — bounded swept-tube passages
Branch: `assembler/world` · evaluator version: **215** · no commit made.
## Design
The route plan and its three passage splines are unchanged. L98 and L99 resolve the same layout identity (`resolved-layout-v3-fnv64:0589300eb07dfe06`) and exact passage control points. L99 prepares each tile's sorted route-segment candidates from tube bounds expanded by 4 m, min-combines their tube distances, then applies localized wall/floor relief and mouth shaping in the detail sweep. A 0.5 m edge blend bounds the integration with the strate field. Existing tread, clearance, mouth fade, and void-crossing support remain in generation.
## Speed
L80 protocol: three interleaved measured L98/L99 pairs at W8 and W12, with warmups, for each stratum. Entries are median seconds (`L98 → L99`). Stratum labels are explicit: SW = Surface World, FI = Floating Islands, TN = Tunnel Network, VS = Vertical Shafts. Whole/near tile counts are shown to keep the world labels fixed.
| Workers | Stratum | Whole / LOD0–2 tiles | Near ready (s) | Whole-view ready (s) |
|---:|---|---:|---:|---:|
| 8 | SW | 2,850 / 1,408 | 1.179 → 0.970 | 2.605 → 2.241 |
| 8 | FI | 2,249 / 1,216 | 2.342 → 1.499 | 3.583 → 2.524 |
| 8 | TN | 2,233 / 1,344 | 2.357 → 1.506 | 3.420 → 2.449 |
| 8 | VS | 2,329 / 1,344 | 2.204 → 1.721 | 3.429 → 3.028 |
| 12 | SW | 2,850 / 1,408 | 1.006 → 0.799 | 2.114 → 1.845 |
| 12 | FI | 2,249 / 1,216 | 2.225 → 1.432 | 3.382 → 2.373 |
| 12 | TN | 2,233 / 1,344 | 2.183 → 1.369 | 3.188 → 2.237 |
| 12 | VS | 2,329 / 1,344 | 1.093 → 0.983 | 2.019 → 1.884 |
All eight cells beat L98. FI/TN approach the ~1 s target, but remain above the passage-off ceilings (0.81/0.85 s). Native Clang and GCC builds took about 11m30s and 1m38s combined; these are accounted separately from readiness times. The Unreal `VFHostEditor Linux Development` build succeeded after timing in 71.50 s with `-NoUBA -MaxParallelActions=4`.
The fast tier is **PASS**: 80 runs, 301.4 s, 269 passing checks and 5 retained historical-hash diagnostics. Both v215 owner hash pins pass.
## Passage audits and measured work
The full-view canonical audit (`max-level=-1`, W12) passes **21,425/21,425** centerline samples. All three resolved interfaces passed:
| Interface | Center / capsule samples | Blocked | Support samples / unsupported | Grade failures | Mouth content / radius | Result |
|---|---:|---:|---:|---:|---:|---|
| 0–1 | 4,151 / 261,513 | 0 | 27,206 / 0 | 0 | 2/2 · 2/2 | PASS |
| 1–2 | 239 / 15,057 | 0 | 1,309 / 0 | 0 | 2/2 · 2/2 | PASS |
| 2–3 | 1,027 / 64,701 | 0 | 5,651 / 0 | 0 | 2/2 · 2/2 | PASS |
The audit summary reports naturalness 3/3, all three interfaces passed, and zero route-binding failures. The fast-tier owner style-1 and style-2 audits also pass, as does the arrival/departure chain (`interfaces=3/3 all_walkable=true`). Capsule clearance, support, headroom, grade, floor and mouth profiles, route binding, open-air, no-long-bridge, rock-column and ripple checks pass.
L77 near-LOD passage CPU share (three repeats, measured mean):
| Stratum | L98 | L99 |
|---|---:|---:|
| SW | 42.83% | 19.75% |
| FI | 56.22% | 34.78% |
| TN | 55.58% | 28.81% |
| VS | 30.21% | 8.56% |
Measured LOD0–2 sweep counters (W8):
| Stratum | Bounded segment candidates | Segment/sample pairs (outside-bound rejects) | Grid samples (no influence) | Base / detail samples |
|---|---:|---:|---:|---:|
| SW | 892 | 8,568,233 (4,699,672) | 14,011,370 (11,949,427) | 2,061,943 / 210,393 |
| FI | 1,439 | 15,261,585 (8,730,625) | 21,565,975 (18,604,965) | 3,377,667 / 347,335 |
| TN | 1,560 | 14,723,215 (8,313,402) | 17,969,538 (15,233,648) | 3,152,585 / 332,547 |
| VS | 1,558 | 14,092,586 (7,348,748) | 20,749,529 (18,066,064) | 2,683,465 / 273,636 |
| **Total** | **5,449** | **52,645,619 (29,092,447)** | **74,296,412 (63,854,104)** | **11,275,660 / 1,163,911** |
The profile covered 478/629/614/699 near tiles (SW/FI/TN/VS). Counts are measured. The bounded-reject and no-influence counts mean those segment tests and tube samples do not proceed to distance composition; the remaining 43,354,950 prepared-sweep lookups had 11,996,257 hits and 31,358,693 misses.
## World change and captures
Carve-reach comparisons covered every desired near-set tile: SW 1,408/1,408, FI 1,216/1,216, TN 1,344/1,344, VS 1,344/1,344. Every stratum had zero sign flips and zero edge-crossing mismatches.
| Stratum | Max / mean displacement | Max / mean normal change | Moved crossings | Pairing failures |
|---|---:|---:|---:|---:|
| SW | 20.11 / 0.0180 cm | 21.35 / 0.0874° | 9,214 | 21 |
| FI | 19.80 / 0.0185 cm | 29.91 / 0.0477° | 10,238 | 114 |
| TN | 21.74 / 0.0387 cm | 41.79 / 0.0291° | 11,370 | 116 |
| VS | 25.88 / 0.0178 cm | 49.82 / 0.00731° | 5,087 | 32 |
The audit's `FAIL` status is only its existing 1 cm displacement / 2° normal-change diagnostic threshold; those thresholds were not an L99 gate. Maxima and their coordinates are in the capture index. Paired L98/L99 captures include the top five displacement sites, top five normal-change sites, both mouths of all three passages, and a midpoint view for each passage: **19 matched views / 38 before-and-after PNGs** in [`captures/INDEX.md`](captures/INDEX.md), plus ten amplified hotspot diffs.
## Determinism and version
All 24 whole-view runs passed: Clang/GCC at W1/W8/W12 across SW/FI/TN/VS. The 20 peer comparisons had zero tile-count, per-tile hash, or whole-hash mismatches. Representative Clang W1 whole hashes (tile count followed by four components):
| Stratum | Whole hash |
|---|---|
| SW | `2850, a962f8cb29383f0e, 3deecb1a8f537726, fa367c6df153636f, 0b0a55b24243bd1e` |
| FI | `2249, 1d867417832047f9, 3f5bf03b78621a43, 601921b84046915d, c00d44660831d266` |
| TN | `2233, 3c0f10f94c18944d, e213da2d3bc34fe1, 318aaa3a18a997be, 7a1cd0183cc57c07` |
| VS | `2329, 3005885856f8adef, 2e3181ca8152f93c, 0cddc16aa45946bb, 69eb58192bab4302` |
Version 215 pins: [`owner-hash-floating-islands-style2-v215.baseline`](../../../../Tools/verify/baselines/owner-hash/canonical/owner-hash-floating-islands-style2-v215.baseline) and [`owner-hash-surface-world-style1-v215.baseline`](../../../../Tools/verify/baselines/owner-hash/canonical/owner-hash-surface-world-style1-v215.baseline). Source stamp: `0d31d588192e30bd`.
## Failed runs
- The first fast-tier attempt stopped at stale pre-L99 hash pins; v215 pins were added and the full rerun passed.
- A zero-radius smoke configuration aborted with `std::bad_alloc` on both L98 and L99; it was an invalid diagnostic setup. Capture probes first rejected an overly strict mouth-anchor assumption, then hit open-contour clip-cap errors; a camera-coordinate bug also produced an empty render. Those probes were corrected, and the final matched views rendered successfully without clipping.
- The displacement/normal diagnostic returned exit 14 because its legacy 1 cm / 2° thresholds were exceeded. This reports the measured world change above, not a passage walkability failure.
All measured workload processes were capped below 30 minutes and run on CPUs 0–11; headless processes used the 6 GiB address-space limit. Logs and summaries are under `Saved/Linux/l99/`. No install, commit, push, or stash was made.
## Captures (fixed)
The first hotspot views exported LOD0, while these maxima come from LOD2. The fixed views export the matching LOD2 audit tile, put the exact audit point at image center, and place the camera 3.25 m from its L98 surface (3.25–3.29 m from the audit point). Each before-and-after surface triangle is within 0.09 m of the reported point in the matching LOD2 tile. Displacement views look obliquely along the passage; normal-change views use normal-color shading. Pixel counts below are measured decoded RGB pixels with any changed channel; each diff image is absolute RGB difference amplified 64×. All ten pairs differ.
Strata labels: SW = Surface World, FI = Floating Islands, TN = Tunnel Network, VS = Vertical Shafts. Passage distances are measured from the site to the nearest route centerline; surface labels use the nearest LOD2 surface and passage-region class.
| Site | Strate · change · focus (m) | Surface · nearest passage | Changed pixels | Before / after / diff |
|---|---|---|---:|---|
| `top-displacement-01` | VS · 25.88 cm · (18.000, -2.000, -817.035) | passage mouth surface · passage/2-3/0, 3.17 m | 557,549 | [L98](captures/top-displacement-01-before-l98.png) / [L99](captures/top-displacement-01-after-l99.png) / [Δ×64](captures/top-displacement-01-diff-amplified.png) |
| `top-displacement-02` | TN · 21.74 cm · (26.000, -4.000, -692.765) | passage upper wall/roof · passage/0-1/0, 3.47 m | 565,587 | [L98](captures/top-displacement-02-before-l98.png) / [L99](captures/top-displacement-02-after-l99.png) / [Δ×64](captures/top-displacement-02-diff-amplified.png) |
| `top-displacement-03` | SW · 20.11 cm · (-23.127, 4.000, -490.000) | passage upper wall/roof · passage/0-1/0, 3.45 m | 760,804 | [L98](captures/top-displacement-03-before-l98.png) / [L99](captures/top-displacement-03-after-l99.png) / [Δ×64](captures/top-displacement-03-diff-amplified.png) |
| `top-displacement-04` | FI · 19.80 cm · (16.000, -17.133, -668.000) | passage upper wall/roof · passage/0-1/0, 3.47 m | 729,645 | [L98](captures/top-displacement-04-before-l98.png) / [L99](captures/top-displacement-04-after-l99.png) / [Δ×64](captures/top-displacement-04-diff-amplified.png) |
| `top-displacement-05` | SW · 19.01 cm · (20.000, 15.067, -500.000) | passage wall · passage/0-1/0, 3.51 m | 372,637 | [L98](captures/top-displacement-05-before-l98.png) / [L99](captures/top-displacement-05-after-l99.png) / [Δ×64](captures/top-displacement-05-diff-amplified.png) |
| `top-normal-01` | VS · 49.82° · (39.696, 6.000, -820.000) | other strata rock wall outside passage envelope · passage/2-3/0, 24.47 m | 423,185 | [L98](captures/top-normal-01-before-l98.png) / [L99](captures/top-normal-01-after-l99.png) / [Δ×64](captures/top-normal-01-diff-amplified.png) |
| `top-normal-02` | TN · 41.79° · (34.000, 5.849, -892.000) | other strata rock wall outside passage envelope · passage/2-3/0, 19.85 m | 539,369 | [L98](captures/top-normal-02-before-l98.png) / [L99](captures/top-normal-02-after-l99.png) / [Δ×64](captures/top-normal-02-diff-amplified.png) |
| `top-normal-03` | TN · 35.51° · (38.000, 3.770, -844.000) | other strata rock wall outside passage envelope · passage/2-3/0, 23.26 m | 738,068 | [L98](captures/top-normal-03-before-l98.png) / [L99](captures/top-normal-03-after-l99.png) / [Δ×64](captures/top-normal-03-diff-amplified.png) |
| `top-normal-04` | VS · 35.05° · (-6.234, 28.000, -824.000) | rock overhang/underside near passage · passage/2-3/0, 13.06 m | 770,128 | [L98](captures/top-normal-04-before-l98.png) / [L99](captures/top-normal-04-after-l99.png) / [Δ×64](captures/top-normal-04-diff-amplified.png) |
| `top-normal-05` | FI · 29.91° · (-39.904, -16.000, -690.000) | rock overhang/island underside outside passage envelope · passage/0-1/0, 24.12 m | 291,241 | [L98](captures/top-normal-05-before-l98.png) / [L99](captures/top-normal-05-after-l99.png) / [Δ×64](captures/top-normal-05-diff-amplified.png) |
The earlier LOD0 pairs and a first LOD2 placement did not include the audited tile; the latter placed displacement-01's nearest surface 22.73 m away and failed the close-framing check. The final LOD2 tile keys match all ten audit rows and all ten image pairs pass the raw-pixel comparison. Export and render logs plus exact camera, triangle, and pixel-diff metadata are in `Saved/Linux/l99/captures-fix/`.