VoxelForge/Standalone/Research/WorldAssemblerM3/Linux144/REPORT.md

Linux144 — Core.Liminal.Field v1

Variant (seed)	Arrival at eye height	Wide view
Flat field + monolith (0)
Repeating hills (1)
Road over cloud floor (2)
Pillar grid (3)

Result: stopped at the mouth-join gate; this is not ready to ship. The new Core.Liminal.Field v1 recipe plans and draws all four seed-selected looks between fixed mouths. In the final composed probe, three of four seeds failed one of their two passage interfaces, and all four failed the liminal walk audit. The requested eight-seed-per-variant matrix was not run after the stop condition was reached.

What changed

• Added one data-driven recipe, Core.Liminal.Field v1. seed % 4 selects flat field with monolith, repeating hills, road over cloud floor, or pillar grid.
• The liminal plan binds the existing passage/0-1/0 arrival and passage/1-2/0 departure as fixed endpoints. It draws an 8-point route deck and structural supports; the road variant uses that deck as its route. The pillar variant resolved 221 grid pillars after route clearance.
• Added a three-strata fixture: Surface World → liminal → Tunnel Network. BuildL144Fixtures.py and CaptureL144.py reproduce the fixture and the eight captures.

Composed probe

The final composed audit window was one run for each seed 0–3 at W1, using Clang 19 source stamp 8acbb7c3dfdb0484100d3d55036db20743ce76d74097b69362bb7d1a1c1d6302. These are audit counts, not timing comparisons.

Variant	Passage interfaces passed	Liminal walk audit	Passage shell / seam audit	Liminal negative control
Flat field + monolith, seed 0	1/2 — FAIL	222 blocked samples, 7 grade failures — FAIL	PASS	altered departure rejected — PASS
Repeating hills, seed 1	2/2; composed summary FAIL	1 blocked sample — FAIL	PASS	altered departure rejected — PASS
Road over cloud floor, seed 2	1/2 — FAIL	3 blocked samples — FAIL	FAIL, 1 floor violation	altered departure rejected — PASS
Pillar grid, seed 3	1/2 — FAIL	289 blocked samples, 9 grade failures — FAIL	PASS	altered departure rejected — PASS

The three partial passage-interface failures reached the requested stop threshold. Every liminal walk probe reported zero unsupported samples and zero structural-support failures; support alone therefore does not explain the blocked routes. The top Surface World route audit passed all four seeds. Passage naturalness and mouth-profile negative controls also detected their injected faults in all four logs.

Captures and limits

All eight 1600×900 images passed capture_quality.py. Arrival cameras are 1.6 m above the fixed arrival mouth and look along the route. Capture exports used --capture-skip-walkability to preserve render evidence despite the failing composed audits; this flag was not used for the audit runs. The wide pillar view shows the overhead deck and one support rather than a legible field of grid columns, and the other images also expose the current route/deck defects.

No additional geometry change was attempted after the three-interface-failure threshold. Earlier diagnostics found the short terminal secant exaggerated the mouth grade; widening the tangent probe interval did not make the final composed walk pass. It remains unproven which combination of terminal connector geometry and nearby field solids accounts for the blocked capsule samples.

Checks

• Clang 19 headless build completed; the final audit and capture executables report the source stamp above.
• Four composed probes ran at W1, one seed per variant. Planning/drawing and the per-audit results are in Saved/Linux/l144/logs/probe8-seed00.log through probe8-seed03.log.
• Eight capture quality checks passed. git diff --check passed.
• Not run after the stop gate: eight seeds per variant, GCC or W8/W12 determinism, owner canonical-hash comparison, fast tier, and one near-ready run. No performance timing window was collected.

Changed Unreal Source/ files: none. Changed standalone sources include StreamBenchmarkHeadless.cpp, StreamBenchmarkPipeline.h, and WorldAssemblyPrototype.h.