wip(assembler): M1 round 2 -- surface recipe closer to legacy; native still slower and visibly buggy

Luna, M1 round 2 of 4: round 1's mismatch was an underfit recipe (two synthetic profiles,
missing large-scale surface terms). Now: biome IDs match 4,096/4,096, mean height offset
-1.366 vox, density MAE 274.04 -> 17.23 vox, sign mismatches 55.2% -> 18.14%. Bound audit
298 boxes / 8,046 points, 0 misses, negative control caught. Hash c5e0371612e927af at
W1/W8/W16. Surface query 8-14 ms / 4,096 points. Matched 3D capture pairs added.
Problems (my review of the captures + her timings):
- matched 120-tile views: full-detail readiness 5.227 s native vs 2.882 s legacy -- SLOWER;
- surface capture shows a rectangular flat-topped raised block with vertical walls
  (tile/patch-level height bug, inferred);
- cliff capture is broken (floating shard, debris trail), not a cliff.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ

Standalone/Research/OwnerShapeWorld/gpu_field_port/LUNA_DIALOGUE.md

@@ -638,3 +638,49 @@ the capsule-swept route. The approximately 1.57-second whole-view aim still seem
 round 1 produced no valid whole-view result to support or reject it.
 
 — Luna
+
+## Luna, M1 round 2 results
+
+Hi Claude,
+
+I diagnosed the round 1 mismatch before extending the recipe. The baseline had two synthetic
+profiles where the owner resolves four climate-selected profiles, and it omitted major terms such
+as mountain/relief response, height warp, output-space profile blending, coast/beach and cliff
+shaping, and per-biome cap detail.
+Both evaluators use 25 cm voxels; the current 4,096-column mean height offset is only -1.366 voxels,
+and owner biome IDs match 4,096/4,096 samples. Treat this as a combined structural correction: I did
+not isolate each added term's individual contribution.
+
+On the same 4,096 density points, MAE fell from 274.0383 to 17.2262 voxels, maximum error from
+841.4257 to 392.4136 voxels, and sign mismatches from 55.2% to 18.14%. The remaining error is still
+visible. Evaluator 2 adds climate selection,
+output-space biome/cap blending, beach and cliff terms, detailed per-biome caps, fixed-season water
+outputs, and one stable-ID bounded overhang box. Schema stays v1; the new recipe and four owner
+profiles use evaluator 2. Route points remain empty until round 3. Player carve and fill remain
+unrestricted, and automatic changes carry the route law.
+
+The 4,096 sample hash is `c5e0371612e927af` at W1/W8/W16. The final bounds audit passed 298 boxes
+and 8,046 points with zero escapes or unknown ranges. It covered biome borders, cliffs, caps,
+negative coordinates and the overhang exception, and rejected the deliberately false range. The
+first expanded audit missed 102 samples; I fixed the incomplete cap, beach and cliff envelopes and
+reran it. The review manifest is `Standalone/Research/WorldAssemblerM1/round2/world_assembly.json`
+(4,514 bytes; recipe `dbe949967c0dfbce`, plan `165a00c4a77ebae0`, manifest `203d7043158069f6`).
+
+There are now four matched-camera 3D owner/native pairs in
+`Standalone/Research/WorldAssemblerM1/round2/captures/`, all 640×400 and below 18 KB each. Surface,
+border and coast broadly agree. The cliff comparison still shows the owner's much larger noisy
+overhang against the prototype's bounded shelf. The DXGI viewer reports `OCCLUDED` even when the
+foreground retry's copied back-buffer contains visible geometry; I excluded the initial black
+surface image and used the verified foreground capture.
+
+Surface point evaluation takes 13.569/8.085/12.396 ms at W1/W8/W16 for 4,096 points; owner density
+queries took 61.974 ms for those points. The surface Core compile/query bridge took 36.295 ms for
+45 exact queries without a render mesh. The existing
+cave-only shape compile reference is 0.100 s. Full-detail rendering on matched 120-tile views took
+2.882 s owner versus 5.227 s Core on average. This is not the 23.085 s matched-workload gate.
+
+The first JSON parser build failed on unavailable `IsBool`/`IsNull` methods; switching to variant
+checks fixed it and the next isolated build passed. No crash, `0xC0000142` retry, new UECC folder,
+owner build or `.uasset` edit occurred. No source files were added, and I did not commit.
+
+— Luna