Files
VoxelForge/Source
Fr0zkaandClaude Opus 5 39c306b524 perf(streaming): prove sealed rock solid instead of sampling it -- crossing fix's cost recovered
4c8d06f stopped content-cutting LOD0 so the player can cross strates, at a cost of +29% static
worker CPU: LOD0 tiles in the sealed rock between strates, which the old cut skipped by accident,
were sampled in full and found solid. This proves those tiles solid instead (Codex; reviewed by me).

- UVoxelGenerator::TryProveSealedSolidTile: AllSolid only for a tile whose whole lattice lies in a
  gap/seal region. It rejects player mods (HasAnyModInChunkRange), passages plus carve pads, the
  origin spine/landing, disturbances, recipes and overrides. It is narrower than the old
  ClassifyTile, which stays off in streaming. CollisionSolidTiles / CollisionNotRequiredTiles
  bookkeeping is preserved, so the collision gate's support search still resolves.
- TryProveOutOfLayoutAirTile: a separate fail-closed constant-air proof for the zero-triangle tiles
  outside the strate layout. It never supplies collision support.
- voxel.SealedSolidProof (default 1). Mining a proven tile invalidates and re-meshes it.
- A soundness test that can fail (VoxelForgeSealedSolidProofTest): 22 candidates, 22 skips, 0
  non-solid lattice samples; a deliberately weakened diff guard produces 437 non-solid samples
  and fails.

Verified: mining tile (14,-18,-8) went from proof=1 / 0 triangles to proof=0 / 8 triangles and
re-meshed; strate-crossing test PASS (max gate 0.098 s); off/on trace 899/899 identical; export
b3e5f4c3..9b377; capability 20,830 / 10,909.

Two interleaved rounds, serialized (owner-editor count 0 before each launch):
                        baseline   crossing-fixed   + proof
  static  worker        35.5 s     41.9 s           36.0 s   (+1.4% vs baseline)
  moving  worker        78.1 s     93.2 s           78.2 s   (+0.1%)
  mesher calls static   18.078M    22.280M          18.078M
Applied triangles identical: 669,834 static / 1,331,482 moving.
Runs: one early automation probe exited 3 with no log (the corrected harness passed); one
exploratory moving run was interrupted and excluded.

Note: HasAnyModInChunkRange scans every modified chunk per candidate (cold audit #21). That is cheap
today (~20 candidates), but it grows with mining; it is queued in the correctness round.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
2026-09-14 22:08:24 +02:00
..