Fr0zkaandClaude Opus 5.5
bd1b2a8f1d
FI/TN/VS (legacy compiled-analytic) skip ZERO tiles at every LOD and evaluate ~44k density points per LOD0/1 tile; SW skips 359/512 at LOD0. The whole-tile skip requires a native tile with a uniform verdict, and the legacy paths return Mixed/unknown (R39 notes attached). Opt-in stage timers, sampled field timer, per-LOD classification and composer counters. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
7.4 KiB
Linux 73 — whole-view breakdown
No generator fix this round. The timed data identifies density/grid work as the broad FI/TN cost, but does not isolate one composer term or show how many tiles a safe interval proof would remove.
Whole view and worker breakdown
Three interleaved clean/profile pairs per strate. Clean medians are the readiness measurements; wall time gates readiness. Worker-seconds sum tile-worker durations and show aggregate work. Stage values below are profiled worker-seconds. field is a sampled estimate (one in 256 calls); it overlaps grid/MC. MC includes grid fill, and normals are a subset of MC, so these columns must not be added. Other is ground+cap refinement, connector shell, and mesh assembly. Payload is separately timed.
Profile whole-view medians (wall / worker s) were SW 2.427/16.250, FI 6.531/49.151, TN 5.593/41.435, VS 2.698/18.206. The clean/profile pairs retained one plan and layout identity throughout.
| Strate | Tiles | Clean wall / worker s | Classify s | Field est. s | Grid fill s | MC s | Normals s | Skirts s | Other s | Payload s |
|---|---|---|---|---|---|---|---|---|---|---|
| SW | 2,850 | 2.324 / 15.044 | 0.425 | 10.135 | 7.871 | 8.395 | 0.113 | 1.013 | 3.541 | 0.011 |
| FI | 2,249 | 6.120 / 45.967 | 0.004 | 39.912 | 36.137 | 40.576 | 0.306 | 1.079 | 2.861 | 0.012 |
| TN | 2,233 | 5.259 / 38.815 | 0.003 | 33.248 | 29.372 | 32.634 | 0.246 | 1.091 | 2.777 | 0.012 |
| VS | 2,329 | 2.548 / 16.929 | 0.004 | 11.656 | 9.349 | 10.095 | 0.200 | 0.640 | 3.305 | 0.012 |
Per LOD classification and density calls
A/S/M/U = all-air / all-solid / mixed / unclassified tiles. Skip is whole-tile skips. Grid is tiles with at least one density sample (it does not claim every possible lattice point was filled). Calls/tile is the measured density point-call count divided by tiles at that LOD.
| Strate | LOD | Tiles | A/S/M/U | Skip | Grid | Calls/tile |
|---|---|---|---|---|---|---|
| SW | 0 | 512 | 256/103/153/0 | 359 | 71 | 3,952 |
| SW | 1 | 448 | 122/173/153/0 | 295 | 77 | 3,025 |
| SW | 2 | 448 | 154/114/116/64 | 268 | 126 | 1,733 |
| SW | 3 | 384 | 64/42/230/48 | 106 | 271 | 4,507 |
| SW | 4 | 320 | 12/0/84/224 | 12 | 308 | 6,260 |
| SW | 5 | 192 | 0/0/0/192 | 0 | 192 | 7,527 |
| SW | 6 | 144 | 0/0/0/144 | 0 | 144 | 11,820 |
| SW | 7 | 96 | 0/0/0/96 | 0 | 96 | 8,621 |
| SW | 8 | 306 | 0/0/0/306 | 0 | 306 | 8,660 |
| FI | 0 | 512 | 0/0/256/256 | 0 | 512 | 44,121 |
| FI | 1 | 448 | 0/0/224/224 | 0 | 448 | 43,322 |
| FI | 2 | 256 | 0/0/160/96 | 0 | 218 | 6,286 |
| FI | 3 | 272 | 0/0/176/96 | 0 | 272 | 7,073 |
| FI | 4 | 272 | 0/0/0/272 | 0 | 272 | 7,342 |
| FI | 5 | 192 | 0/0/0/192 | 0 | 192 | 7,802 |
| FI | 6 | 96 | 0/0/0/96 | 0 | 96 | 13,996 |
| FI | 7 | 48 | 0/0/0/48 | 0 | 48 | 9,773 |
| FI | 8 | 153 | 0/0/0/153 | 0 | 153 | 9,852 |
| TN | 0 | 512 | 0/0/128/384 | 0 | 512 | 44,076 |
| TN | 1 | 448 | 0/0/112/336 | 0 | 448 | 43,315 |
| TN | 2 | 384 | 0/0/112/272 | 0 | 344 | 6,550 |
| TN | 3 | 256 | 0/0/48/208 | 0 | 256 | 7,185 |
| TN | 4 | 192 | 0/0/0/192 | 0 | 192 | 7,466 |
| TN | 5 | 144 | 0/0/0/144 | 0 | 144 | 7,836 |
| TN | 6 | 96 | 0/0/0/96 | 0 | 96 | 13,996 |
| TN | 7 | 48 | 0/0/0/48 | 0 | 48 | 9,773 |
| TN | 8 | 153 | 0/0/0/153 | 0 | 153 | 9,853 |
| VS | 0 | 512 | 0/0/0/512 | 0 | 512 | 43,824 |
| VS | 1 | 448 | 0/0/0/448 | 0 | 448 | 43,230 |
| VS | 2 | 384 | 0/0/0/384 | 0 | 384 | 7,050 |
| VS | 3 | 256 | 0/0/0/256 | 0 | 256 | 7,163 |
| VS | 4 | 192 | 0/0/0/192 | 0 | 192 | 7,290 |
| VS | 5 | 144 | 0/0/0/144 | 0 | 144 | 7,536 |
| VS | 6 | 144 | 0/0/0/144 | 0 | 144 | 11,617 |
| VS | 7 | 96 | 0/0/0/96 | 0 | 96 | 8,316 |
| VS | 8 | 153 | 0/0/0/153 | 0 | 153 | 9,855 |
Per-LOD stage timers, proof-cell counts, grid sample totals, field-timer sample counts, and composer counters are in lod-breakdown.csv; medians and paired overhead are in strate-summary.csv. The tile counter distinguishes whole-tile skips from tiles with any density samples; it does not prove every lattice point was filled. The field timer samples aggregate evaluation, not individual composer terms. FI native-component attribution stayed zero on this compiled-analytic path. Composer counters are counts only: at LOD0 FI recorded 54.5M relief calls, 63.4M carve visits, 20.4M support visits, and 24.6M mouth-route checks; TN recorded 46.9M, 54.2M, 15.5M, and 20.4M respectively. They show work volume, not time share.
Skip verdict
FI has zero all-air/all-solid skips at every LOD. At LOD0/1 it classifies half the tiles mixed and half unclassified, yet evaluates about 44k/43k density points per tile. TN also has zero skips; its LOD0/1 tiles are mixed or unclassified and have the same roughly 44k/43k calls per tile.
The active FI owner export has no matching native IslandPlan/field adapter. The FI interior therefore falls back to Mixed/unknown; the mesher skips only AllAir or AllSolid tiles. Its inactive native-FI interval cannot prove the compiled-analytic field. Any future finite-support bound must cover the true FBM supremum (1.5) and downward root tails, then include shared solid writers.
TN is still on the legacy compiled-analytic path. Its nearby-primitive checks are vetoes; the fallback evaluates the box center, not a whole-box interval, and does not feed a tile skip. The shared composer can force Mixed for local partial passage/band intersections, but is not forcing every tile: FI/TN bands are 0.5 m thick against 240 m/128 m interiors, and tiles are unclassified across LODs, not only at strate boundaries. A box wholly inside a band can be AllSolid.
Next options (no fix selected)
- FI finite-support interval for the active compiled-analytic field. Highest potential: many proven-empty tiles could remove a large share of FI's roughly 40 worker-seconds of field evaluation. Gain is unquantified until proof-hit rate and classifier cost are measured. Bound roughness, roots, connectors, and shared solid writers.
- TN interval for the active compiled-analytic cave field. Second potential: prove host-rock tiles outside cave influence and skip them. Could save substantial TN work if most blocks miss cave bounds; proof coverage is unknown.
- Composer-aware uniform verdicts. Carry air proofs through carve-only passage intersections and resolve boundary/support writers by polarity. Expected smaller and local to overlap tiles; count the rejected-Mixed reasons before implementing.
No before/after generator comparison, passage audit, fast-tier run, or world version bump applies because no generator change was made. No output tile hashes were collected: --stream-hash disables tile skips, which would invalidate this measurement. No Unreal build was needed. All 24 final clean/profile generator jobs exited 0. Two preliminary reduction attempts failed only while parsing zero-sample profiler counters; the reducer was corrected.
Identity
- HEAD:
d970d75be80fd31f16e2dc41006456bc52ad180d(assembler/world) - Build source stamp:
ab7f0e95f415c0dc4eb81ae3b9a355fe98e5ceb89faf4aefc54b65d936d09fdf - Plan hash:
bcc0d0d33ea39435; resolved layout:resolved-layout-v3-fnv64:0589300eb07dfe06 - World assembly schema/evaluator: 2 / 213; owner export schema: 1. No world content or generator version changed.
- Profiling self-cost (median paired profile-minus-clean): SW +0.089 s wall / +0.732 worker-s; FI +0.503 / +3.840; TN +0.073 / +0.437; VS +0.151 / +1.210. These differences varied across the three pairs.