VoxelForge/Standalone/Research/WorldAssemblerM3/Linux139/REPORT.md

# Linux139 — Vertical Shafts on the Core planner

## Result

S3 now reports `Core.Caves.PrimitiveField` / `Caves-native`, using the shared Tunnel Network
planner. The composed acceptance gate is not green: **15/16 seeds pass; seed 2 fails the passage
shell and tile seam checks at the arrival from passage 2–3.** This worktree is not promoted as an
accepted Vertical Shafts generator.

The basal primordial route law means arrival passage 2–3 to a walkable terminal landing above the
bedrock seal. There is no departure mouth at the bottom of the owner world. The shared planner
builds that walk-only spine, its access tunnels, and landings within the existing grade contract.
The full-height open shafts and their ledges are an exploration tier: a drop or climb is not
walk-only traversal, and climb or mine access stays in the reporting tiers. `Core.Lattice` was not
used because L58's VerticalShafts lattice route looked like a corkscrew, not an open shaft.

## Plan and parameter mapping

`PlanResolvedCavernNetworkV1` is reused without a planner fork. `Core.Caves.PrimitiveField` is the
shared evaluator, extended with `shapeKind=1`: the existing room/tunnel shapes do not represent a
capped, full-height vertical bore, and the L58 lattice route was a corkscrew. Shaft records use
that cylinder primitive with ledges, short access bores, and cross-shaft links. Generic cave
branches are disabled in shaft mode.

| `FVerticalShaftParams` field(s) | Core setting |
|---|---|
| `ShaftSpacing` | Shared planner room spacing; owner voxels convert to integer micrometres. |
| `ShaftDensity` | `roomDensityQ16`, controlling shaft placement; a nonzero setting guarantees the route's initial shaft. |
| `ShaftMinRadius`, `ShaftMaxRadius` | Minimum and maximum shaft radii. |
| `CrossConnectChance` | `tunnelDensityQ16` for links between shafts. |
| `ConnectorRadius` | Equal minimum/maximum radius for access and connector bores. |
| `LedgeSpacing`, `LedgeDepth` | Vertical shelf interval and intrusion; zero spacing disables shelves and zero depth makes them nonintrusive. |
| `SurfaceRoughness` | Roughness amplitude; frequency derives as `max(0.5 m, 4 owner voxels)` because no frequency field exists. |
| `BoundarySealThickness`, `BaseDensity` | Existing resolved top/basal boundary bands via `sealAndDensity`, outside planner settings. |

## Measurements

The final source stamp was `b8603b6cb3975ca0f9669759e4e2464561c557fdb7aa6836acef123ecf0b99d1`.
The composed sweep covered seeds 0–15 with one run per seed (reported process time 320.485 s in
total; individual runs 10.296–34.015 s). Seeds 0, 1, and 3–15 passed passages, route, seams, shell,
and blocked-route controls. Seed 2's route grades/support, naturalness, mouth profile, ripple,
dispatch, and negative controls passed; its composed passage, seam, and shell summaries failed.
The shell audit reported eight `wall_missing` samples on `passage/2-3/0`, at route stations
489.627–490.321 m. The final log records positive native density at those samples and negative
wall-shell margins. See `Saved/Linux/l139/seeds/seed2-final.log` and
`Saved/Linux/l139/seeds/summary.json`.

Three geometry variants were tried against this same shell gate: starting passage clearance at
12 m, adding a guarded cubic lateral bow, and starting clearance at 8 m. The first and third still
failed wall checks; the bow failed floor checks. Per the three-failure stop rule, these were
discarded. The final binary retains the successful Q16 endpoint-normalization correction needed
for seed 1 and no seed-2 geometry patch.

L80 ran after the seed sweep. The host gate was **load 1.31, CPU idle 99.4%**. It used one warm-up
per build and stratum/worker cell, then five interleaved measured runs for each of three builds at
W8 and W12 across all four strata: 120 measured runs total. Comparisons use `fa4417a`,
`assembler/world`, and the L139 binary; the established L137 A/A noise floor is ±1% ([A/A report](../Linux137/REPORT.md)).
Times are medians in seconds; deltas are L139 relative to each comparison build. Paired values in
each cell are W8 / W12.

| Stratum | LOD0–2 L139 | Δ vs `fa4417a` | Δ vs `assembler/world` | Whole-view L139 | Δ vs `fa4417a` | Δ vs `assembler/world` |
|---|---:|---:|---:|---:|---:|---:|
| S0 Surface World | 0.838 / 0.886 | +1.71% / +3.97% | +2.92% / +5.06% | 1.862 / 1.804 | +2.42% / +1.63% | −0.11% / −0.33% |
| S1 Floating Islands | 1.036 / 1.038 | +4.72% / +4.61% | +2.09% / +5.35% | 1.842 / 1.830 | +0.11% / +1.16% | −6.73% / −1.24% |
| S2 Tunnel Network | 1.105 / 1.078 | +2.44% / +3.07% | +3.57% / +7.57% | 1.876 / 1.808 | −1.11% / +0.00% | −4.33% / −1.15% |
| S3 Vertical Shafts | 0.899 / 0.884 | −3.81% / −0.94% | −0.95% / +2.00% | 1.706 / 1.657 | −6.57% / −2.53% | −6.21% / −4.77% |

Whole-view medians are numerically lower than `assembler/world` in all eight cells; the S0 W8 and
W12 differences are within the ±1% noise floor. Seven of eight LOD0–2 cells are slower than
`assembler/world`; the remaining cell is within the noise floor. This is characterization of an
unaccepted build, not a promotion claim. Raw runs and medians are in
`Saved/Linux/l139/timing/summary.json`.

## Checks and captures

- Fast tier: **PASS**, 284.0 s; the S3 dispatch line is `VerticalShafts` → `Core.Caves.PrimitiveField` v1 / `Caves-native` / `FCavesFieldAdapter`.
- Clang and GCC determinism: **PASS**, W1/W8/W12 (six cases), 13 hash lines and resolved-layout identity matched. Both compilers built the final source stamp above.
- Canonical v237 baselines added for Surface World and Floating Islands; both match their v236 values.
- Five eye-height views were captured before and after: arrival mouth, shaft access landing, looking down, looking up from a ledge, and basal terminal landing. All five L139 views pass the existing quality threshold. Two L138 baseline views fail it: the old layer has no shaft at the selected site, so the down-shaft view is flat, and the upward view is low contrast. The threshold was not relaxed. Images and camera details are in `captures/`.
- The outside views are `captures/after-whole-layer-density-cutaway.jpg` (coarse LOD density-mesh slice at y=−5 m) and `captures/after-whole-layer-side-elevation.svg` (resolved-plan projection that labels the walk route and all 53 shaft primitives). The cutaway passes the unchanged image-quality threshold; its plan companion makes the full shaft layout clearer.
- Linux `VFHostEditor` Development build: **PASS**, 27 actions, 72.58 s, run after timing with `-NoUBA -MaxParallelActions=4` and no `prlimit`. It emitted three pre-existing deprecated RealtimeMesh API warnings from `VoxelWorld.cpp`; that file was not changed.
- `git diff --check` and Python bytecode compilation of the L139 runners passed. Windows build and Windows determinism were not run.

Logs and source-stamped binaries are under `Saved/Linux/l139/`. The only changed Unreal `Source/`
file is `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp`.

## Remaining risk

The passage shell and seam break at the incoming mouth on seed 2 remain unresolved. The local
clearance/heading premise failed three geometry attempts; a new route-to-mouth design premise is
needed before another geometry change. Walk-only connectivity stays guaranteed by construction;
vertical drops, climbs, and mine access remain reporting tiers.