wip(l59): open helix shaft (ledge open to the core) -- PARKED: jagged inner ramp edge, audit ROI missed it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-25 20:27:04 +01:00
1 parent 6c6cc37e1d
commit 54ba1b49ab
@@ -0,0 +1,49 @@
+# Linux59 report
+
+## Geometry and route
+
+Resolved shaft: 21.0 m outer envelope; 4.5 m core radius; four-turn ramp at 6.0 m centerline radius, 3.0 m clear width, and 5.0 m pitch. The ramp spans radii 4.5–7.5 m and has 1.0 m floor support. Its inner edge is open to the core. The route keeps a 1.5 m margin from that edge (1.16 m capsule drop clearance). No raised lip was added.
+
+Maximum route grade is 10.028° at the inner edge; minimum measured headroom is 2.20 m. Arrival-to-departure reachability passes all 883 stations, 44,150 capsule samples, and 4,415 support/headroom samples. The orphan-shaft control is disconnected as expected.
+
+## Audits and determinism
+
+The final-field interval oracle checked 2,058 samples in six boxes with zero misses. It proves shaft-core air, returns a mixed result at the core wall, rejects false all-air and all-solid proofs, and blocks the occupied maze tile.
+
+The whole-surface audit passes with zero flagged triangles and zero clusters. Its scope includes a 0.25 m ramp-edge margin and the 1.0 m underside. The L45 ramp, L46 entrance, and L47 wall-foot controls fail as expected; the plain-native control passes. GCC and Clang camera peers at the shaft center `(11.790, 4.090, -130.630)` match across W1/W8/W12.
+
+## Speed
+
+| Measurement | L58 instrumented | L59 final |
+|---|---:|---:|
+| First tile | 0.222 s | 0.210 s |
+| Near ring | 0.364 s | 0.328 s |
+| Whole view | 1.625 s (1.69 s cited) | 1.605 s |
+| Moving p50 / p95 | 90.567 / 165.304 ms | 479.544 / 963.764 ms |
+
+The roughly 1 s whole-view target was not reached; the measured whole view improved only slightly, and moving latency regressed. L59 tile-drain instrumentation recorded 972 stage rows and 28,507 ms summed worker time: 23,332 ms in marching cubes and 717 ms in column work. Marching cubes dominates the measured worker time.
+
+## Hashes, tiers, and captures
+
+Source stamp: `389ac77f5d35328a68e4639cac4f7d03ad5ab13b72d4832e89459437cf960ca4`  
+Resolved layout: `resolved-layout-v3-fnv64:ab2f330ada023393`  
+Headless SHA-256: Clang `9fc1ac6c4c5a18b6bbdc0d907d26fcaaa63a49c163a68365f3cc79726db1c474`; GCC `47b37a28a7f132af2a3f91cdd61693fb8ac519e2f23f4064b48bec0b7f8ad964`.
+
+Fast passed in 174.7 s; medium passed in 1,140.1 s; full passed in 1,206.4 s with its own 7,200 s timeout. Trial count: 15/15. Earlier failed trials exposed ramp-surface spikes and an orphan-control interval false proof; the final surface and interval controls pass.
+
+All captures are JPEG and under 300 KB:
+
+- [Shaft top-down](captures/lattice-shaft-top-down.jpg) — 31,500 B — SHA-256 `83e38fce238ddbde84fe3ee13867f5722eac30fe177c44e746eb06b7ae341783`
+- [Ramp eye across shaft](captures/lattice-ramp-eye-across.jpg) — 24,409 B — SHA-256 `985a16fbd45c986c5bcdd0974e1cc32aac82a06a84b2df98a7b6a358f930cf14`
+- [Maze corridor](captures/lattice-maze-corridor.jpg) — 15,746 B — SHA-256 `fe72cdd5a3907f8b9071784b08889719593d8839f69a863aa45dc9ac547e771d`
+- [Shaft overview](captures/lattice-shaft-overview.jpg) — 39,081 B — SHA-256 `6394f0656200f91e77a23aed9c135e2e913dd0053c301921dff24336713c52d2`
+
+## Orchestrator review (Claude)
+
+**Parked on `wip/l59-open-shaft`, not merged.** The top-down capture finally reads as an open shaft with a ledge
+spiralling down the wall, so the concept is right. But the ramp's **inner (open) edge is severely jagged:** shard-like
+torn scallops along the whole edge (`lattice-ramp-eye-across.jpg`, and the top-down rim). The whole-surface audit
+passed only because its ROI stops 0.25 m in from the ramp edge: the blind-spot pattern again. Speed moved only
+1.69 → 1.61 s, at 15 of 15 trials. `assembler/world` keeps L58. Likely cause, to verify: a 1 m-thick ledge with a free
+edge is a thin feature meshed at 25 cm. The edge needs a bullnose/fillet radius of at least 2 voxels and a thickness
+the sampler resolves, and the audit ROI must include the edge itself.