wip(assembler): round S2 (Sol) -- shared per-tile feature queries for every Core family; island placement 30x faster

Island-only worker context replaced by a shared worker-local candidate domain
(WorldAssemblyFeatureTileListV1 / ScopedWorldAssemblyFeatureQueriesV1) serving
SurfaceWorld, FloatingIslands and far-shell queries; a per-domain cost hint keeps the
full scan on small plans. Content placement prepares the same queries and skips proven
air: owner island cell p95 7.07 -> 0.24 ms; populated fixture 572 exact placements,
3.27 s -> 29 ms. Full/candidate hashes equal with dropped-feature controls on both
families; S1 default W1/W8/W16 hashes reproduced; Core-off and M2 fixtures pass.
Default contract for future families in ARCHITECTURE 8.23. Run on gpt-6.1-sol medium.

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

@@ -1138,3 +1138,59 @@ link-safety passed. No 0xC0000142 retry or new UECC folder occurred. I did not o
 editor or build his project, edit any .uasset, create a link, commit, push or stash.
 
 — Luna
+
+
+## World assembler S2 / M3 Round 6 — 2026-10-18
+
+The shared engine now owns a family-neutral, worker-local feature-query scope. I removed the
+island-only context from ShapeWorld; SurfaceWorld finite features, island groups, meshing, far-shell
+field calls, decoration cells and landmark columns use the same domain/identity/halo contract.
+S1's root-aware positive-density bounds and ordered connector composition are preserved.
+
+The measurements matter here: the staged SurfaceWorld has one finite shelf, and the staged island
+plan has four islands. Forced candidate bookkeeping gave no useful small-plan gain and regressed
+the island tile timings by 7–13%. I added a shared cost hint, resolved once per domain, so these
+small plans keep their full evaluator. Richer plans still use candidates. With the actual staged
+owner template expanded to 256 islands, forced-list p95 was 110.570 -> 7.3619 ms (about 15x), with
+identical hashes and a detected dropped-feature control. The composed surface stress (64 shelves,
+one XY height writer and one cap writer) was 0.7084 -> 0.4078 ms, again exact with its control caught.
+The S1 57x stress number used a different Round 3 template, so I do not compare those as a regression.
+
+The placement cost came from marching almost the entire strate and using a full plan at every
+step. Family-declared solid-support bounds now feed a shared column range query, alongside the
+same feature scopes. We skip only certified air and keep the old march phase, bisection,
+interpolation and transforms. Owner-palette p95 fell 7.0706 -> 0.2385 ms, with 58.84 density calls
+per cell versus S1's 7,763.38. Its sampled authored palette still places zero objects; I keep that
+limitation explicit rather than claiming those timings include spawning trees.
+
+A separate populated fixture produced 572 placements across 14 cells. All 572 full/candidate
+transforms match exactly, all canonical anchor checks pass, and the largest density residual is
+0.035156 Core voxel units. The 64-cell active batch, including merge and real HISM application,
+was 3271.147 -> 29.105 ms. All 572 instances were applied and then cleared. This is synchronous
+CPU work under null RHI; it excludes rendering, GPU upload, async tree completion and owner tree
+Blueprint logic. One run failed the new count assertion because normal region application clears
+live landmark footprints. I isolated the validation fixture from those footprints; normal content
+keeps them enabled. The corrected run passes.
+
+BuildHost SurfaceWorld passed its full actor/streaming/collision/edit checks and W1/W8/W16 with
+S1's default 25 cm hashes. The final island actor also keeps S1's entire W1/W8/W16 matrix. The final
+focused surface replay checks LOD0/1/7/8 full-versus-default tile hashes after the cost hint. M2 V1/V2
+fixture repeats, borders, MC identity and all 125 doorway phases pass. A separate Core-off run
+proves unchanged legacy density and 12.5 cm scale. All builds/link audits passed. No timeout,
+0xC0000142 or new UECC folder occurred. The failed HISM fixture run exited normally with test status 1.
+
+My next structural suggestion is an immutable per-strate family-provider table, carrying declared
+bounds, cost hints, value certificates and optional support/surface queries. It would avoid repeating
+concrete family enumeration and rebuilding metadata in each consumer. I have not rebuilt that
+boundary this round; the report lists the allocation, many-strate, hash and stale-provider evidence
+needed first. Preserve the existing bounded biome lattice and cave interval machinery meanwhile.
+
+New source files are `Round6/SharedFeatureListAudit.cpp` and `Round6/BuildSharedFeatureListAudit.ps1`
+under WorldAssemblerM3. The report lists every modified source, likely compile spots, all failed and
+replayed runs, timing limitations and the plain-words owner note. CODEMAP section 3 and ARCHITECTURE
+section 8.23 are updated. No owner project/editor was opened, closed or built. No .uasset was edited,
+and no link, commit, push or stash was created.
+
+[Full S2 report](../../WorldAssemblerM3/Round6/REPORT.md).
+
+— Sol