# L176 — Design50 step 3 verification

## Result

Step 3 does **not** hold across all 16 seeds. The fresh sweep has zero seam mismatches, but the no-added-solid audit reports 158 more-solid density samples on five seeds. Landing pads pass 0/96; that is the step-4 no-fill site-support issue identified by the owner, so the strict clamp remains unchanged. The audit runner's strict summary is 0/16.

The final Clang and GCC builds from `HEAD` both contain source stamp `e8dba3faa06d714b1d035a02960ab894370b0c3fd3b3d76669f2209290808d81`. All 16 sweep rows were rerun against that stamp; none resumed pre-crash output.

## Design50 edits retained

The composition order is layer content, air-only bore plus floor shell where host density is air, Design49 clamp stamp masked by the bore, then edits. L176 removes the chamber records and composition, slab return, corridor exemptions, shell-over-air fills, `floorProfileIncludesLandingChambers`, both floor blends, the Surface World plane fit, and chamber audits. The pad stamp and `LandingPadAudit` controls remain. No verification fix was applied; the strict clamp is kept.

## Fresh 16-seed gates and attribution

The sequential sweep window was Nov 4, approximately 16:23–17:50 local.

| Gate | Result |
|---|---|
| Source stamp | 16/16 match |
| Tile seams | 104,945 pairs; 128,557,625 shared samples; 0 mismatches; 0 exemptions. Every seed covers 6/6 landing discs and 6/6 end runs. |
| No added solid over host air | 190,941 host-air samples; 11/16 seeds have zero changes; 158 changes in seeds 5, 6, 9, 13, and 15. The planted-wall control was detected on 16/16 seeds. |
| Floor shell | Aggregate: 154,127 shell samples for 154,190 host-air samples, 0 missing. Caveat: 23 passage rows had zero host-air samples and per-row `FAIL`, despite the aggregate summary passing. |
| Passage interior | 4,863,267 capsule samples; 0 solid samples; controls 48/48. |
| End runs | 48/48 pass. |
| Tunnel Network routes | 14/16 pass (seed 2 blocked 28 samples; seed 13 blocked 18). This is one seed better than L175's 13/16. |
| Vertical Shafts routes | 16/16 pass. |
| Blocked-route controls | 32/32 pass. |
| Landing pads | 0/96 pass; negative controls 383/384. Seed 6 missed the 2% tilt control on the upper `passage/1-2/0` pad. |

The 158 no-added-solid samples count any density shift toward solid, not only sign changes. The first sample reported for each failing seed is:

| Seed | Samples | First point (m) | Composed / host density |
|---:|---:|---|---:|
| 5 | 13 | (-17.652, 5.395, -744.064) | -1.744629 / +2.301025 |
| 6 | 49 | (12.114, -11.141, -744.111) | +0.964031 / +1.529659 |
| 9 | 87 | (15.018, -11.921, -744.112) | +2.023779 / +2.342343 |
| 13 | 5 | (16.552, 0.321, -744.061) | +0.045202 / +1.734494 |
| 15 | 4 | (13.073, -4.463, -744.300) | +1.460241 / +2.570065 |

These points cluster near z = -744 m. Their proximity to passage `1-2` is inferred from the saved layouts; the audit emits coordinates and densities, not a source primitive, so the composition cause remains untraced. Seed 0's L130 proof passes (reach 1.053 m, 29,856 samples, 0 proof/feature misses, 0 pad-stamp misses).

The pad audit remains a step-4 placement failure: the no-fill surface landing lacks support from existing ground. All pad checks stay strict. The step-2 census was not rerun: L174 recorded 6 overlap pairs (seeds 1, 10, 11), 23 end-turn violations (seeds 1, 3–10, 13–15), and 0 landing-column violations.

## Compiler hashes and fast tier

All 18 canonical hash runs passed at the final stamp during Nov 4, 18:04–18:09 local. Clang and GCC match at W1, W8, and W12 for owner4 surface-world, owner2 surface-world, and owner2 floating-islands. The two owner2 v242 baselines were written; each differs from v241 in 3 of 12 canonical rows. Results are in `Saved/Linux/l176/hashes/summary.json`.

`verify.py fast` ran Nov 4, 18:09–18:10 local. It passed `streaming-admission-order-mid-pass-refill`, then stopped at its first owner audit, `owner-floating-islands-style2-clang` (exit 10; stamp, recipe, and style checks pass). That fixture reports 249 added-solid samples, 25 unsupported probes, 57 tube-overlap pairs, and one end-turn violation. The later fast-tier stages did not run. Summary: `Saved/Linux/l176/verify-fast/fast-summary.json`.

The seam negative control was also rerun on the final build. It injected one 1e-6 density mismatch; the seam audit reported `mismatches=1`, `injected=DETECTED`, `status=FAIL` as expected (exit 10). See `Saved/Linux/l176/controls/seam-injected.retry.log`.

## Time split — one seed and layer

One measured L176 run used seed 0, layer 0's camera, W8, and `--profile-lod-work`. The run window was Nov 4, 18:16 local; pre-run 1-minute load was 0.56 and 2-second CPU idle was 99.0%. This is a single diagnostic, not an L80 timing matrix.

| Stage | Measured result |
|---|---:|
| World assembly plan (`WorldAssemblerPlanTiming`) | 16.295 s |
| Shape compile (`StreamPipeline shape` / `ShapeStage`) | 0.192 s |
| First tile ready | 16.424 s |
| LOD0–2 ready | 17.108 s |
| First tile to LOD0–2 readiness | 0.684 s |

The LOD0–2 worker counters sum to about 0.770 worker-seconds of column computation, 1.170 worker-seconds of marching-cubes work, 0.860 worker-seconds of skirts, and 0.012 worker-seconds of mesh assembly. Those counters accumulate work across workers and overlap; they are not wall-time slices. The full-detail LOD0–1 tile summary reports 0.774 worker-seconds column compute, 0.777 classification, and 2.758 meshing across 960 tiles.

The dominant measured interval is planning, not shape compilation or the final tile-generation/meshing window. The plan log shows 14,529 route candidates and terrain grids of 11,449, 46,225, and 184,041 samples at 1.0, 0.5, and 0.25 m. The two explicit cave-planner timers total 0.900 ms. It is an inference that the Surface World route search accounts for most of the plan time; the harness does not time those search sub-stages separately. No passage/seam audit or debug capture was enabled in this timing run.

The `assembler/world` e3359d2 comparison binary built successfully in `/tmp` with stamp `74aa3923e38272bc37e4c2fe3b4529b7c6f03ecc4e1b8e8cdb3e764b65dd229e`. The same seed-0 harness failed before streaming: all 14,520 finite-strate route candidates were rejected by the composed-field capsule/support sweep. There is no comparable mainline timing.

## Checks and remaining risks

- Fresh Clang and GCC builds completed; both binary stamps match the source stamp above.
- The complete seed sweep, 18-run hash matrix, injected seam control, `verify.py fast`, and one-source timing run completed. `verify.py fast` failed at the first owner passage audit as described.
- `git diff --check` passed.
- The mainline benchmark did not reach streaming; no Unreal build or Windows compile was run.
- No `Source/` Unreal module files changed. The L176 standalone harness source is `Standalone/Source/StreamBenchmarkPipeline.h`.
- Remaining failures are the five-seed no-added-solid gate, step-4 landing support/pad gate, and parked step-2 census. The 23 zero-coverage floor-shell rows and one missed pad negative control limit audit confidence.