wip: L174 Design50 steps 0-2 in progress (interrupted for a power cut)

Audits collect gates and run seams (landing discs + end runs, LOD0/1), tagged uncapped failures, separate exempt
counts, self-overlap census; level end runs; overlap/turn constraints. Partial sweep: end runs pass 3/3 per seed;
1-2 end-turn violations per seed; 50-110k tagged seam mismatches per seed (real vs audit artefact not yet decided).
Not verified; resume L174.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1

Standalone/Research/WorldAssemblerM3/Design50/README.md

@@ -0,0 +1,104 @@
+# Design50 — why the tunnel landings fail, and the plan (review 2026-11-03)
+
+A read-only review (Claude Opus 5.5, extra-high effort) of `wip/combined-cave` (b097283), with the L133–L173 history and
+the L173 logs. The orchestrator verified R6 and R7a in the code.
+
+**Headline:** one wrong assumption, that a tunnel end is a separate 2-D "landing chamber" owned by the tube, the layer
+and the audits at once. The plan replaces it with a level end run of the same spline plus the Design49 stamp,
+applied last.
+
+## Root causes
+
+- **R1:** the chamber plane and the tube's 1-D h(s) disagree off the centreline (up to 1.3 m at the 7.5 m rim on FI).
+- **R2:** the departure tube descends from the disc centre and undercuts its own landing. Unsupported samples: upper
+  chambers 316, lower 0 (L173); L159 upper pads 96/96 failing, lower 2/96.
+- **R3:** the passage composer overrides the host:
+  - a forced-solid chamber floor slab (`ShapeWorldInternal.h` ~1665);
+  - shell walls filled over host air (~1869–1898);
+  - per-layer corridor exemptions only (Surface World, planned tunnels);
+  - the liminal route stamp re-applied after the carve (~1997, the L146 refill).
+- **R4:** planner self-overlap. The FI helix pitch (3.2–3.9 m) is less than the bore height (~5.8 m), with 2,738
+  overlapping control-point pairs. Upper and lower mouths share XY, so 32/32 final junctions turn 107–178°
+  (curl-back).
+- **R5:** the passage picks the landing height, not the layer. FI departures land inside the FI→TN transition band.
+- **R6 (bug, verified):** `floorProfileIncludesLandingChambers` (set for every passage) makes
+  `PassageFlatFloorCenterZSubvoxels` return early, skipping both blends.
+- **R7 (audits):**
+  - (a) **verified:** the composed audit returns 10 before the seam audit, so seams were never measured;
+  - (b) audits re-model chambers (L159's interior-air 0/16 was a phantom);
+  - (c) exempt probes count as supported; zero-coverage gates pass;
+  - (d) only 8 failures are printed;
+  - (e) the TN and VS sweeps are the same 4-strata world.
+
+## Model
+
+1. **One spline, one floor:** h(s) is level and straight for the first and last a ≥ R+1 m.
+2. **The landing = the end run:** a flat disc radius R at P = h(end); the descent starts outside the disc.
+3. **One composition order:** layer content (including liminal decks) → bore carve (air only, plus a floor shell under
+   the bore where the host is air) → Design49 clamp stamp last, masked by the bore → edits. No walls or roofs over
+   host air; no corridor exemptions.
+4. **Who sets P:** fixed-ground layers (SW, FI, liminal) propose the site and P; for carved layers (TN, VS) the passage
+   picks P and the cave planner connects.
+
+**Keep:**
+- the spline, h(s) and the grade audit;
+- the interior-air gate;
+- cave binding to the end anchors;
+- the L148 envelope, the L156 band fit and the L130 bounds;
+- `l159-pad-stamp.patch` (stamp, bounds, LandingPadAudit plus 4 controls).
+
+**Delete:**
+- the chamber records and their composition, and the slab return;
+- the corridor exemptions and the shell-over-air fills;
+- the flag and both floor blends;
+- the SW plane fit;
+- the chamber audits.
+
+## Steps
+
+- **0:** audits that measure:
+  - collect gates, exit at the end;
+  - seams over discs and end runs at LOD0/1;
+  - tagged failing samples, with no 8-sample cap;
+  - exempt probes counted separately, with a coverage minimum;
+  - a self-overlap census;
+  - Core-exported predicates.
+- **1:** level end runs.
+- **2:** no self-overlap or hairpins:
+  - pitch ≥ bore height + 2 shells + margin;
+  - lower mouth on the ring;
+  - end turns ≤ 60°;
+  - the helix kept out of the landing columns.
+- **3:** the composer and the audits together, with the stamp last.
+- **4:** site selection for fixed-ground layers.
+- **5:** verify (fast tier; TN/VS/native-FI sweeps 16/16; versioned hashes; Clang/GCC W1/W8/W12; captures).
+
+**Owner look decisions:**
+- the disc radius, the end-run length, the cone angle (12°) and reach (8 m);
+- no walls on open-air landings;
+- headroom underground;
+- whether a FI landing may float or must sit on an island;
+- wider spirals (~6 m or more);
+- landings may move.
+
+## Owner decisions (2026-11-03)
+
+- Open-air landings: **no walls**.
+- Floating Islands landings: **add a landing island to the FI plan** at each tunnel end: top at P, footprint wider
+  than the pad, normal island shaping at the edges. It's planned from the seed before generation, and the FI route
+  connects to it. (Not pathfinding the tunnel to an existing island.)
+
+## Owner decision: uniform layer planners (2026-11-03)
+
+- **Planning side: a `LayerPlanner` parent with one child per layer type:**
+  - `Plan(seed, settings)`;
+  - `OnTunnelArrival(end)`, the landing hook: nothing (stamp only), a plan record (FI landing island, liminal walkway,
+    cave connection), or a content request for Unreal;
+  - `ProvideRouteAnchors()`;
+  - `MakeDrawer()`.
+
+  Planning runs once per world, so virtual dispatch costs nothing there.
+- **Drawing side:** keep the `std::variant` adapters with a shared `densityAt` shape (no virtual call per sample;
+  `ShapeWorldInternal.h` ~121), and choose the drawer once per tile.
+- **Migration:** one planner at a time; owner canonical hashes identical after each move (refactor only); the landing
+  hook then implements step 4.