VoxelForge/Standalone/Research/WorldAssemblerM3/Linux43/REPORT.md

Linux 43 — Core landmark SDF stamps

Source stamp: 18b89fb48febf887c5b61e14c699c95535b09613f23618ca57271bc5d2a455f3 (clean Clang and GCC builds).

Vocabulary and records

Stamp vocabulary v1 contains carveEllipsoid, carveArch, fillPillar, and fillDais. A resolved record carries a stable ID, strate index, position in integer micrometres, conservative bounds, half extents, smooth-CSG width, irregularity amplitude and wavelength, irregularity seed, and quarter-turn yaw. Bounds include the smoothing and roughness support.

Carves use smooth maximum with the negated SDF; fills use smooth minimum. Core keeps its density sign: negative is solid and positive is air. Bounded trilinear value noise adds seeded relief.

Placement and culling

The resolver sorts requests by stable ID and derives placement offsets and irregularity from (seed, strate, stable ID). It tries at most 512 deterministic candidates. The strict layout rejects bounds that intersect a passage mouth or walk corridor, touch a content face, or overlap a ceiling or bedrock band. No safe candidate is a placement error. The route and band negative controls were both rejected (2/2).

The stamp BVH queries conservative bounds once per tile. A tile with no candidates performs no stamp SDF evaluation. Stamp-intersecting boxes classify as mixed, and native-only readiness cannot skip them. With no requests, stamp fields are omitted from canonical JSON, preserving L42 identities and hashes.

Owner demo and hashes

--landmark-chamber-demo is off by default. On seed 0 it adds three records to the owner SurfaceWorld strate (strate 1); no Unreal files changed:

• Ellipsoid chamber landmark/sw-arrival-chamber-v1: center (-8.918, 31.950, -513.083) m, half extents (8.5, 7.5, 6.0) m.
• Arch landmark/sw-arrival-chamber-entrance-v1: center (-12.289, 37.027, -513.311) m, half extents (2.6, 4.2, 3.2) m.
• Low dais landmark/sw-arrival-chamber-dais-v1: center (-7.489, 30.688, -515.838) m, half extents (3.2, 2.8, 0.9) m.

The chamber center is 30.9 m from the arrival mouth at (9.25, 9.25, -502.608) m, off the walk route. The active layout carries vocabulary version 1 and identity resolved-layout-v3-fnv64:91eaf2f0248f2c8b.

Without stamps, the FI/SW layout IDs remain b7a7d9ecc48d8f26 / 8a52e84c26ed92d2; the 12-record owner signatures remain L42's FI 9820f4f38b189b3f7263810e7b3ff62aa502b21f476ffcfdbc7761a65cb07d58 and SW a1e9268a4bbcc77d2a1ac54095df6851a70bab65c43ae50b332ed75132617834.

At the same chamber-centered camera, W12 hashed 343 tiles. Stamp off: density 0x18d6f6cb3498deff, mesh 0x293c0ddff66b99f9, payload 0xeb07403f34d60134, topology 0x002905a3f028c9f1. Stamp on: density 0xbf8af855ee5db8fb, mesh 0xbb64c203dc533070, payload 0xd14d45e6b9faa9c0, topology 0x44f1b422df20a159. The stamp-on W1 and W12 signatures match.

Safety, cost, and tiers

With the demo on, passage walkability passed 1545/1545; the passage audit passed; endpoint probes passed 27/27; the interface audit passed with hole_columns=0 and air_samples=0.

For outside-stamp tiles, five W1 pairs over 343 tiles measured median whole-view-ready 0.457 s off / 0.466 s on (+1.97%); paired changes ranged from −2.76% to +3.08%. Five W12 pairs over 125 tiles measured 0.189 s / 0.191 s (+1.06%). These tile bounds do not intersect a stamp, so the BVH returns no stamp candidates and no stamp SDF runs. The small timing deltas sit within pair-to-pair variation. Full-tier five-pair cost versus HEAD: FI 17.342 s → 17.495 s (+0.88%); SW 1.799 s → 1.805 s (+0.33%).

Tier	Result
Fast	PASS, 161.3 s, 41 checks
Medium	PASS, 468.3 s
Full	PASS, 1551.2 s, 539 checks; clean Clang/GCC build, 30-cell determinism, cost pairs

Captures

All JPEGs passed capture_quality.py and are below 300 KB:

• Entrance, eye height — 27 KB
• Inside — 29 KB
• Overview — 43 KB

Failed runs

• An early route-control point hit the mouth guard before the corridor guard; its position was moved so the final route and band controls exercise the intended rejection checks.
• The first JSON round trip lost precision in a 64-bit irregularity seed. Seeds are now kept within the JSON number type's exact 53-bit integer range; canonical round-trip passes.
• A clip-plane capture was rejected by the renderer (open_contours=10); the final three views use the uncut mesh.
• A supplemental stamp-off, chamber-centered W1/W12 aggregate differed reproducibly (density 0x68f495ba62aace05 / 0x18d6f6cb3498deff), though its nine exact-query probes matched. The standard owner fast/full determinism checks pass, and the stamp-on chamber-centered aggregate matches W1/W12.

Orchestrator review (Claude)

Merged as infrastructure, off by default: the stamp vocabulary, records, placement guards and BVH culling are sound, the owner hashes are unchanged with no requests, and the route/band controls reject. Not accepted as a player-visible landmark yet:

• the demo chamber is buried: its top is about 4.5 m below the SW ground near the arrival mouth, with no opening to the walkable surface, so a player can never see it;
• the arch reads as a boxy notch with a sharp vertical corner;
• the entrance capture is unreadable.

Determinism finding (pre-existing, not caused by L43). The "supplemental stamp-off W1/W12 mismatch" in the failed-runs list is real. I reproduced it on committed L42 and on L37 (density 0x68f495ba… at W1 vs 0x18d6f6cb… at W12, camera (-8.918, 31.950, -513.083), --max-clip-level=0 --render-distance-chunks=0). S14 is stable there. It is a worker-count-dependent result in the production path, invisible to our tiers, which only test the spawn camera and the 24 m crossing. It is the top priority for L44.