VoxelForge/Standalone/Research/WorldAssemblerM3/Astra3/REVIEW.md

[INFERENCE] **Improve the existing brick-proof system; don’t add a general “checked voxel” mask.** The promising savings are avoided field/composition work and passage preparation, rather than marching-cubes classification.

References below abbreviate these files: [M = production mesher](/home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Source/VoxelForge/Private/VoxelMarchingCubesMesher.cpp), [S = ShapeWorldInternal.h](/home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Core/src/ShapeWorldInternal.h), [C = VoxelForgeCore.cpp](/home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Core/src/VoxelForgeCore.cpp), [O = VoxelDensityOpStack.cpp](/home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Source/VoxelForge/Private/VoxelDensityOpStack.cpp). L-numbers identify `Standalone/Research/WorldAssemblerM3/LinuxN/REPORT.md`.

[INFERENCE] Evidence limitation: Linux128 is absent from this checkout. Its supplied 7.9 m support and slower one-owner result remain unverified here. The reports below span different revisions; their percentages must not be presented as today’s census.

**1. Where time goes**

[MEASURED] L77’s exclusive FI/TN LOD0–2 worker account:

| Work | FI | TN |
|---|---:|---:|
| Total worker CPU/view | 23.02 s | 21.50 s |
| Passage composition | 65.9% | 65.4% |
| Support | 14.2% | 13.5% |
| Base field | 3.2% | 3.3% |
| Bands / dispatch | 2.3% / 3.7% | 2.6% / 3.6% |
| MC / skirts / refinement, excluding field work | 1.6% / 0.1% / 7.5% | 1.2% / 0.2% / 8.6% |

[MEASURED] Grid fill occupied 19.058/17.377 worker-seconds, **overlapping** those phase categories. L117’s later census found support zero; its smooth-tunnel regression concentrated about 95% of added tile CPU in passage-feature tiles. Cell classification stayed flat; gradients became cheaper. Thus L77 identifies historical leverage, not current percentages.

[MEASURED] L125 preparation averaged 514/130/4.1 ms worker work at LOD0/1/2. LOD0 evaluated 3.11–8.49 million segment/sample pairs with 67–72% overlap, although the exact band occupied only 5–7% of grid samples. L126 reduced floor-profile calls 20–23% without useful readiness improvement.

[CODE] Brick skipping already exists: configurable block verdicts, required-sample masks, and active-block gradient halos (M:646,1053). MC already rejects cells without sign crossings before interpolation/gradients (M:2186). Gradient memoisation exists (M:1956). Skirts issue separate density queries (M:2491); refinement has its own stage (C:2975).

[INFERENCE] A pre-evaluation proof can save density, composition, and subsequent MC work. A sign mask constructed after density evaluation saves only downstream scanning. Classification itself costs work; no isolated current classification percentage is established by these reports.

[MEASURED] L124’s shape gate alone took 179–226 ms, with no near-tile admissions before completion. L87 additionally measured a 107–153 ms far-parent gate and no headless collision wait. Tile masks cannot directly remove these gates or explain the supplied UE/headless gap.

**2. Sound flags and hazards**

[INFERENCE] For final density \(D\), solid requires an upper bound covering the evaluator’s supremum, \(U<0\); air requires a valid lower bound \(L\ge0\). “Deep” requires an explicit margin. Unknown must stay unknown. A crossing proves an MC-active cell; uniform sampled signs do not prove an entire continuous volume surface-free.

[CODE] Reusable proofs include:

- S:5428: halo-expanded box classification.
- S:5974: exact surface-column extrema over the MC lattice.
- S:6019: floating-island density intervals.
- S:5197,6032: native cave intervals.
- S:5949: stamp interval composition with outward numerical padding.
- O:6998: operation-specific inactivity from SDF lower bounds.

[CODE] These proofs have guards: passage overlap returns unknown (S:5687); unsupported stamp composition returns unknown (S:5655); stratum crossings are rejected (S:5900). The mesher distinguishes uniform candidates from `bSkipSafe` certificates (M:715,756), and otherwise validates candidates against actual samples (M:1838).

[CODE] For gated modifiers, `Identity` can mean `Sdf >= T`, not `abs(Sdf) >= T` (O:1496,6998). Noise bounds cannot assume nominal `[-1,1]`: FBM and unclamped cellular bounds are explicitly discussed at O:848,301.

[INFERENCE] Additional hazards: confusing internal SDF with final density; certifying base terrain before caves, tunnels, shells, stamps or edits; using lattice proofs for off-grid queries; ignoring rounding or gradient halos. A below-`TerrainZ` solid rule applied to 3D caves deletes caves. “Touched by tunnel” should mean conservative candidate overlap; only **proven untouched** authorizes omission.

**3. Granularity**

[INFERENCE] Storage per flag for 32³ cells:

| Granularity | Storage | Construction/tradeoff |
|---|---:|---|
| Per cell | 4 KiB | 32,768 decisions; expensive unless derived during necessary work |
| 4³ bricks | 64 B | 512 proofs; tighter boundaries, more proof/halo overhead |
| 8³ bricks | 8 B | 64 proofs; cheaper, more mixed bricks |
| XY columns | 128 B | 1,024 decisions; useful mainly for heightfield-specific proofs |

[INFERENCE] These exclude corner/halo samples, metadata and additional flags. A column bit cannot represent multiple cave intervals. Brick runs preserve contiguous X sampling and SIMD better than scattered per-sample branches; bit packing alone does not establish a speedup.

[CODE] The existing implementation uses byte arrays and preserves exact samples around active bricks (M:1053–1083). Replacing densities with arbitrary same-sign constants near active cells would corrupt interpolated positions or normals.

**4. “Already evaluated”**

[CODE] Passage preparation precedes meshing (C:2747). Its valid sentinel is initialization, not completion: subsequent segments update floor selection, shell minima and round-distance winners, with deterministic tie-breaking (S:11520,11615,11688).

[INFERENCE] Therefore first-visitor wins is invalid. Memoisation is valid only for identical queries against identical seed/layout/evaluator/edit revision and query semantics; cache the value, not just a bit.

[MEASURED] L77 counted 8.6–8.9% halo points, without measuring duplicate coordinates. L125 verified 4,900 shared samples across four neighbour pairs and still recorded substantial off-grid skirt/midpoint fallbacks.

[CODE] Shared-grid reuse and gradient validity caches already exist (M:1031,1956). [INFERENCE] Skirt/seam queries are reuse candidates only when coordinates and semantics actually match; “same voxel” is insufficient.

**5. Other uses**

[INFERENCE] Edit-dirty bricks can localize regeneration, with invalidation expanded through normals, seams and collision dependencies. Collision-only servers can skip certified empty regions, but still need boundary geometry. Streaming can prioritize potentially occupied bricks; nav and lighting need clearance/connectivity or visibility data beyond occupancy. Keep masks derived locally from replicated seed + layout + diff, versioned with that state.

**6. Recommendation**

- [INFERENCE] **Tune existing 4³/8³ certificates**, caching proofs and emitting contiguous required-sample runs. Plausible incremental near-ready saving: **0–5%**, with regressions possible. Measure proof time, samples avoided, halo expansion, total worker CPU and UE readiness.
- [INFERENCE] **Pass required-region information into passage preparation**, rejecting whole segment/brick regions conservatively. Plausible near-ready saving: **0–10%**, highly uncertain because passage overlap currently vetoes proofs. Kill it if pair reduction fails to reduce preparation and end-to-end time.

[INFERENCE] Both require seam/off-grid validation, adversarial bound checks, edit invalidation tests, and cross-machine/worker determinism. Preserve canonical tie-breaking; old-version byte identity need not constrain a sound redesign.