VoxelForge/Standalone/Research/WorldAssemblerM3/Linux67/REPORT.md

L67 — four-strate passage walkability

Cause and fix

The FI, TN, and VS paths use the compiled analytic base evaluator and then the shared strate composer. The “native” 5/5 result was the same composed field with resolved passages removed; it was not a raw-generator probe. The band stage is therefore not the difference. The passage carve and the support/capsule gates are.

• 0–1 SW→FI: at the failed probe (21.366, -11.480, -726.367), passage 0 first restored support, then passage 1’s bore changed density to +1.613 (carve_sdf=-0.292), removing it. The shared composer now defers that support only when a physically overlapping later bore can reach it, then applies it after the later carve. The column’s natural relief is predictor-driven and tapers at the root; relief rose from 0.172 m to 0.219 m.
• 1–2 FI→TN: at (2.419, -13.433, -735.870), support restoration was rejected by an adjacent route segment’s player-capsule test (clearance=-0.245 m, required 0.045 m). The segment-specific overlap candidates now keep that test local to physically relevant route geometry. The reported 0.2498 m step came from the audit switching to a neighboring zero crossing on its 0.25 m scan. Refining and following the same root shows a continuous floor; the step audit now reports 0.000 m.
• 2–3 TN→VS: at (-1.365, -15.673, -897.088), the same capsule gate rejected support next to route segment 81 (clearance=-0.329 m). The shared segment-overlap handling fixes this link too.

The L66 passageIndex>0 / strates.size()>2 switches are gone. “Internal chain” now means both mouths lie inside their endpoint strata content bands and neither endpoint is a SurfaceWorld landing. Lower-mouth body overlap uses actual 3D distance to the mouth and tread, independent of plan length or passage index.

Audit

All existing thresholds are unchanged. The L66 baseline binary reproduced the original failures.

Passage	L66 negative control	L67 result
0–1 SW→FI	59 unsupported; column relief 0.172 m < 0.20 m	0 unsupported / 0 blocked / 0 grade failures; wall/floor ripple 1.49 / 0.06 cm; column relief 0.219 m; wall relief 0.267 m
1–2 FI→TN	3 unsupported; wall/floor ripple 6.78 / 2.90 cm; apparent 0.2498 m step	0 / 0 / 0; wall/floor ripple 2.29 / 1.23 cm; refined step 0.000 m; wall relief 0.257 m
2–3 TN→VS	2 unsupported	0 / 0 / 0; wall/floor ripple 1.48 / 0.47 cm; wall relief 0.248 m

All three now pass support, blocked, grade, mouth-binding, ripple, and naturalness gates. The detailed final audit is in owner-four-strata-route4.log; the measured composition trace is in l66-root-refined-audit.log.

Controls, hashes, and builds

• The L66 build remains a negative control: it reproduces 59 / 3 / 2 unsupported and the 0.172 m column relief failure. The expected audit exit is 10.
• The pinned fast tier passes: 269 PASS, 5 diagnostic, overall PASS. The prior L52 surface-world hash comparison is diagnostic because the route output changed; the changed style-1 baseline is pinned at v212. See the fast summary.
• Evaluator version is 212 (from 211), required because shared tread composition and route selection changed generated output. Source stamp: d49aa8d941fb390818569243d7783f3e2799795244a8d7ca48e3e640d4a2876f.
• All four strata produced 12 equal hash rows at each of W1/W8/W12 (determinism summary). Clang and GCC builds both pass.
• Changed implementation files are Standalone/Core/include/VoxelForgeCore/VoxelForgeCore.h, Standalone/Core/src/ResolvedLayoutInternal.h, Standalone/Core/src/ShapeWorldInternal.h, Standalone/Source/OwnerWorldLoader.h, Standalone/Source/StreamBenchmarkPipeline.h, and Standalone/Tools/verify/verify.py. No Unreal Source/ files changed.

Cost

Three interleaved whole-view pairs per strate, L66→L67 median seconds (summary):

Strate	L66	L67	Change
0	1.101	1.121	+1.8%
1	1.575	1.552	−1.5%
2	0.581	0.602	+3.6%
3	0.216	0.212	−1.9%

Strate 2’s samples overlap (0.582–0.600 s) and paired results are within observed run variation. The route chooser now selects the shortest candidate within 0.10 m of maximum sampled tread clearance, with tangent alignment as a tie-breaker. It also skips an extra closure turn when mouth separation exceeds the tread envelope; the FI→TN route is 114.231 m rather than the earlier 165.799 m candidate.

Captures

Seven JPEGs are in captures/; each is under 300 KB. The TN→VS endpoint meshes are sparse in these offscreen views, so those two frames show the nearby cave/strate surface more clearly than the tread.

• 0–1 arrival 26,342 B, departure 40,012 B
• 1–2 arrival 34,298 B, departure 42,450 B
• 2–3 arrival 23,937 B, departure 23,380 B
• Mid-tunnel, 1–2 22,392 B

Failed runs

• The expected L66 negative-control audit exited 10 after reproducing the failures.
• One p0 clipped arrival-camera trial exited 3 because the clip plane left two open contours. It was replaced by the successful uncut SurfaceWorld entrance view.
• A p2 clipped arrival-camera trial also exited 3 on open contours. The final endpoint image uses the no-clip interior view. Both cap failures are recorded in failed-capture-trials.log.
• No final capture, Clang/GCC build, determinism run, fast-tier gate, or cost run failed. No run exceeded 30 minutes.

Audit controls (follow-up)

This section supersedes the earlier wall-ripple values; the support, blocked, grade, mouth, and column results remain as reported. The residual-only gauge omitted pose.radius and hid an authored-radius beat:

Field / gauge	Old gauge (authored radius, linear detrend)	Residual-only gauge
L66 four-strata field	6.78 cm	2.21 cm
L67 four-strata field	14.52 cm	2.29 cm

On the old gauge, the geometry change raised 1–2 by 7.74 cm. On L67 geometry, the residual-only gauge then removed 12.23 cm; the reported 6.78→2.29 cm drop therefore masked the radius beat. The final gauge keeps sampleRadius + 0.5*(left+right) and cubic-detrends each 32 m window. The 0.18 m is the probe depth inside the authored radius; its former standalone subtraction is removed. The cubic baseline leaves the 0–1 and 2–3 taper profiles below 0.01 cm, while 1–2 retains a 4.77 cm authored-radius beat at 7.06 m. Final ripple results are 1.45 / 7.27 / 0.92 cm: 1–2 fails the unchanged 3.5 cm wall limit. Support, blocked, grade, and mouth-binding checks still pass for all three; overall passage audit is 2/3. See audit-control-l67-final-owner.log and audit-control-authored-radius-all.log.

Negative control under the updated gauge	Measured	Result
L35 wall ribs	9.75 cm at 6.50 m; limit 3.50 cm	Detected
Synthetic wall spheres / floor beads	5.98 / 5.98 cm; limits 3.50 / 2.50 cm	Detected
Synthetic signed-distance floor step	0.25 m step; 2.97 cm ripple vs 2.50 cm limit	Detected
L36 washboard field, five-step root refinement	6.46 cm; limit 2.50 cm	Detected
L23 terrace profile	2.00 m flat then 0.25 m step	Detected

No control passed, so no old audit behavior was restored. Logs: L35 ribs, L36 washboard. The L67 synthetic controls are in the final owner log above. One direct link of the current headless code against archived Core libraries failed route setup and was discarded; the control runs used source/Core-matched L35 builds. The follow-up changed only Standalone/Source/StreamBenchmarkPipeline.h; Clang Linux build succeeded. No timing run, commit, push, or stash was made.