wip(assembler): round L33 -- SurfaceWorld can be an upper strate; FloorProfileBlend measures seams

Route planning for an upper SurfaceWorld failed on long candidate chords over steep relief; chords
are now capped to the arrival/departure radii on an 8 m centre grid with an origin-proximity cost
(21.5 m walk, about 4 deg). The departure passage starts at the SW ground and tunnels down through rock.
FI -> SW -> TN fixture: all 4 styles pass every passage audit on both compilers. FloorProfileBlend
now checks endpoint height/tangent seams (step and tangent-step negative controls detected). Owner
world hashes unchanged; 30/30 determinism; cost within noise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX

Standalone/Research/WorldAssemblerM3/Linux33/REPORT.md

@@ -0,0 +1,58 @@
+# Linux33 report
+
+## Upper SurfaceWorld route
+
+The old fixed candidate set used 64 m chords, 2 m waypoint spacing and 16 m centre spacing. On the refreshed upper-stack fixture, the legacy replay measured 366/648 candidates above the 15° grade limit; its worst segment was 79.971° at XY (−84,−48) to (−82,−48). These were long paths over steep terrain relief near the outer search area. L23’s earlier fixture reported 648/648 grade rejects; the refreshed fixture gives a different count. The route Z samples are terrain-ground heights (about −503 m at the chosen mouth), not a target at the bottom of the 499 m strate. There was no unit or bedrock-reference error.
+
+The planner now caps chords at 58.5 m for the 24 m arrival and 37.5 m departure radii, searches six lengths and places route centres on an 8 m grid, with a proximity cost favouring the origin. It found a 21.5 m walk from (−2.75,0) to departure (18.75,0), average grade about 4.2°. The SW passage starts at the ground floor (−503.166 m), with a 1.5 m portal blending to a 3.5 m tunnel over 9 m, then descends through SW rock. The first tunnel point is at (32.5,0), floor −505.637 m for style 0.
+
+## Upper-stack fixture audits
+
+Fixture: FloatingIslands → SurfaceWorld → TunnelNetwork; recipe stamp `floating-islands`; layout identity `resolved-layout-v3-fnv64:fc4e73619037f88c`. Clang and GCC passed every row for every style.
+
+| Passage style | Floor chord deviation (voxels) | Passage and route audits | Open air | Naturalness | Bridge / columns | Bindings / walkability |
+|---:|---:|---|---|---|---|---|
+| 0 | 2.081 | 2/2 each pass | 0 violations | 2/2 pass | pass / 0 failures | 0 failures / pass |
+| 1 | 10.901 | 2/2 each pass | 0 violations | 2/2 pass | pass / 0 failures | 0 failures / pass |
+| 2 | 14.822 | 2/2 each pass | 0 violations | 2/2 pass | pass / 0 failures | 0 failures / pass |
+| 3 | 16.374 | 2/2 each pass | 0 violations | 2/2 pass | pass / 0 failures | 0 failures / pass |
+
+All rows are recipe-stamped and passed on both compilers. Capsule blocking, unsupported samples and grade failures were zero; mouth-content and mouth-radius checks were 2/2. This includes open-air, naturalness, no-long-bridge, rock-column wall/support, route binding and walkability audits.
+
+## FloorProfileBlend
+
+The old 2-voxel check measured maximum curve-to-straight-chord deviation, which grows with valid curvature and does not measure a seam. The check now compares endpoint height and tangent residuals. All styles have 0.00000000 voxel height residual and 0.00000000 voxel/m tangent residual (limits 0.010 and 0.001). Negative controls detected a 3-voxel step and a 0.25 voxel/m tangent step. Verdict: check corrected; no floor geometry change was needed.
+
+## Captures
+
+Both 1600×900 JPEGs are under 300 KB. The first shows the SW ground mouth; the second is a cutaway through the rock entry.
+
+Sizes and SHA-256: mouth 26,939 B, `561d3a1e334ea3a3115fb781a1130ec850d8777b74c1be3704c70b865878640a`; rock entry 37,572 B, `f85d150bf44cd2cdae1ee49de023ad9e4c657b8fa7e8d8f8b93abc86586ad6e5`.
+
+![SurfaceWorld departure mouth](captures/sw-departure-mouth.jpg)
+
+![SurfaceWorld departure rock entry](captures/sw-departure-rock-entry.jpg)
+
+## Hashes, determinism and cost
+
+Owner-world hashes match L32 exactly: all 12 hash lines per run across Clang/GCC and both recipes passed the owner preflight; the L25 and 24 m crossing matrix also passed at W1/W8/W12 (24/24 owner rows). W12 whole-view hashes: SurfaceWorld density `953eb1df1705f277`, mesh `d8006cb43f44aa5b`, payload `9217bff766970de5`, topology `83e3aa6f3a063e60`; FloatingIslands density `806537cdd98f8afb`, mesh `a4b06dd4d260943d`, payload `763a99aa04a73eae`, topology `baa37d80d3620d87`. The upper fixture was stable across W1/W8/W12 on both compilers. Total determinism: 30/30. Invariant audits: 16/16.
+
+Five interleaved W12 cost pairs per recipe against L32: median whole-view ready time changed +0.99% for SurfaceWorld and +0.41% for FloatingIslands; paired medians were +0.57% and +0.41%, within run noise. Recipe/evaluator versions remain 2/20. No Unreal files changed; no Unreal build was run. No version bump was needed because the owner hashes are unchanged.
+
+## Failed and retried runs
+
+- CLI setup rejected `--compiler clang`; the accepted form was `--compiler=clang++-19`. One shell quoting attempt ended at EOF. A FloatingIslands fixture initially had a `surface-world` recipe stamp and was regenerated with `floating-islands`.
+- A translated ring centre failed the strict placement disk and was reverted. Fixture audit iterations: r1 had style 1–3 failures (TN capsule pinch, route undercut, naturalness); r2 had 141 open-air violations plus route/naturalness failures; r3 had strict ring-placement and route-support failures; r4 had four edge open-air samples. The broad column-window fix passed passage audits but changed owner hashes. Its first scoped retry dropped supported bracket samples and failed SurfaceWorld hash preflight on both compilers; restoring the original non-SW bracket bounds fixed it.
+- An intermediate Clang build failed on `coreVoxelSizeMetres` and `gradeSegments`; both names were corrected. Final Clang 19 and GCC 14 `--fp-contract=off`, `-j12` builds succeeded. An early determinism run on stale binaries was stopped after 21 rows and showed the interim hash mismatch; the final binaries passed all 30 rows.
+- The first capture export included the far ring and grew to 2.49 GB; it was stopped before rendering. The LOD0 streamed export reduced the scene to about 2.2 MB. Capture framings r5–r8 rendered successfully but obscured the mouth; r9 is the inspected final framing.
+
+No run exceeded 30 minutes and no step remained stuck for 15 minutes. Logs and exports are in `Saved/Linux/l33/`; no commit, push or stash was made.
+
+## Orchestrator review (Claude)
+
+- Merged: the owner world is unchanged (hashes equal L32); SW-as-upper now plans its route and passes the full audit set
+  on the FI→SW→TN fixture. The FloorProfileBlend check now measures seams (endpoint height/tangent residuals) with
+  step/tangent negative controls. Accepted: chord deviation measured curvature, not a defect.
+- Captures: `sw-departure-mouth.jpg` is out of focus and unreadable; `sw-departure-rock-entry.jpg` shows a natural opening
+  but **horizontal step-like ledges inside**, reminiscent of the L23 terraces. L34 checks passage mouths and entries for
+  stepping across styles and fixes the blurry capture setup.