VoxelForge/Standalone/Research/WorldAssemblerM3/Linux51/REPORT.md

# Linux 51 — FloatingIslands speed pass

Date: 2026-10-24. Final source stamp: `8ce895ccb91f7ffd3717068b9c0b867bda5f1e5e8a894c557afbea882a3f4894`.
Final fast, medium, and full tiers pass. Final trials used fresh processes, W12, `taskset -c 0-11`,
`prlimit --as=6442450944`, and `TMPDIR=Saved/Linux/l51/tmp`.

## Measurements

The five interleaved W12 cost pairs in the full tier measured:

| Recipe | Before median | After median | Change |
| --- | ---: | ---: | ---: |
| FloatingIslands | 16.736 s | 7.355 s | -56.05% |
| SurfaceWorld | 1.770 s | 1.756 s | -0.79% |

The FI pair changes were -54.80%, -55.91%, -56.05%, -56.30%, and -55.79%.
The earlier five-pair stream timings also gave 15.425 s to 7.115 s for FI and 1.629 s to 1.699 s for SW.

| Recipe | First tile before → after | Near ring before → after |
| --- | ---: | ---: |
| FloatingIslands | 1.273 → 1.245 s | 1.563 → 1.417 s |
| SurfaceWorld | 0.514 → 0.519 s | 0.528 → 0.538 s |

Moving-run entering-tile latency, measured submit-to-completion:

| Recipe / step | Before p50 / p95 | After p50 / p95 |
| --- | ---: | ---: |
| FI / 8 m | 874 / 1,419 ms | 446 / 661 ms |
| FI / 24 m | 1,924 / 3,276 ms | 763 / 1,304 ms |
| SW / 8 m | 23 / 46 ms | 22 / 42 ms |
| SW / 24 m | 58 / 104 ms | 59 / 104 ms |

The baseline mesh-coverage run found 1,295 nonempty FI tiles of 2,370, so 1,075 tiles (45.36%) were all-air.
SW had 676 empty tiles (28.52%). Before the proof change FI had 25,536 unknown boxes and no whole-tile skips.
After it, the no-audit run reported 1,959 air boxes, 18,857 mixed, 240 unknown, and 216 whole-tile skips.

## Suspects and cuts

- **Loose FI intervals: confirmed and fixed.** Added conservative ranges for island bodies, root lobes, anchor
  buttresses, connectors, surface detail, and smooth-min budgets. The shared box classifier now admits a pure FI
  field to the same native interval route used by other generators. LOD0–1 worker time fell from 116.28 to
  48.16 worker-seconds after the interval and X8 cuts.
- **Small-plan per-tile candidate pruning: no benefit on the owner's four-island plan.** Turning it on added no
  tile skips and raised profiled LOD0–1 work from 79.87 to 86.11 worker-seconds. The existing `>4` gate remains.
- **Duplicate X8 preflight: confirmed and removed.** The native FI lane samples are now reused by the X8 result
  path, with composition, edits, and counters preserved.
- **LOD7–8 far-shell work: confirmed and cut outside FI support bounds.** Baseline FI LOD7–8 worker time was
  47.32 seconds versus 4.25 for SW. FI LOD8 alone spent 31.26 seconds in connector-shell work and 20.96 seconds
  in column evaluation. The shortcut now checks per-tile island/connector influence bounds with a two-voxel halo
  and preserves top/bottom seal regions. FI LOD7–8 worker time fell to 7.55 seconds; connector field calls fell
  from 8.02 million combined to zero. The far shell geometry changed at LOD7–8.

The small-plan gate was kept off for this plan. `perf record` could not run because `perf` is not installed; no
installation was attempted.

## Proofs, output, and verification

The FI uniform-block audit checked 21,504 certified blocks and 28,621,824 samples: zero fallbacks, unknowns, or
violations. The SW control reported zero violations. The support-bound far-shell gate is separate from the audited
uniform-block classifier; it retains the existing conservative plan bounds and boundary-seal checks.

The 14 canonical passage PNGs are byte-identical to L50. Medium also passed walkability, passage, transition,
ring, capture, and hole controls. Full passed Clang and GCC W1/W8/W12 determinism, the random-camera sweep, and
30-cell determinism.

The far-shell output change bumped `farHorizonVersion` from 4 to 5. FI full-detail mesh/payload/topology hashes
remain `a818654be89790ac / 370ff80451105a39 / 06e507557aa70173`; the FI full-detail density aggregate changed
because 70 empty LOD0 rows now carry the empty FNV density hash. The new FI whole-view signature
`density/mesh/payload/topology` is:

`55f117f15cd44103 / d987bfc3a5888d04 / f04ce3c0b238a8c4 / a4601550e27a18a4`.

SW geometry is unchanged; its LOD4–8 hashes use v5 salt. Its new whole-view signature is
`7dc75e2c09302258 / 7bcc2f81b3cce165 / 5199d76ac540ff54 / f1dd43d2c10c78bf`.

Final verification: fast PASS (152.3 s), medium PASS (471.2 s), full PASS (1,061.6 s). The old L46 hash identity
check is now diagnostic per the owner's correction; worker/compiler determinism remains gated.

Failed or unusable attempts: an early fast run lacked verifier-named binaries; a medium run used a trial output
directory without those binaries, then rejected stale GCC; one CLI attempt used unsupported `--data=` instead of
the positional input. These were corrected before the final tiers. Two intermediate trial builds were also
launched without the outer CPU/address-space wrapper because `VOXELFORGE_LIMITED_BUILD=1` bypassed the script's
wrapper; they are not part of the final build or tier results. All final builds and final benchmark/verification
processes were constrained. The initial L46 hash gate was updated to a diagnostic after reading the shared owner
correction.

## Remaining to reach about one second

FI is still 7.36 seconds for the whole 2,370-tile view, with a 1.42-second near ring. The first tile is about
1.25 seconds and the shape/setup stage is still close to 0.9 seconds. LOD0–1 mesh work and 7.55 worker-seconds
remaining at LOD7–8 are the next measured targets. SW remains about 1.76 seconds whole-view.

## Orchestrator review (Claude)

Accepted. The stamp matches the final Clang and GCC builds (`8ce895cc…`). FI whole view 16.74 → 7.36 s (−56%; all five pairs between −54.8% and −56.3%). FI moving p50 halves at 8 m and 24 m; SW is unchanged. Uniform-block proof audit: 28.6 M samples, 0 violations. The far horizon changed outside island support (farHorizonVersion 4 → 5), which is allowed under the owner rule (peer determinism, no visible regression, version bump). Next targets per the report: LOD0–1 meshing on mixed boxes, and the roughly 0.9 s shape/setup.