Concurrent Chunk Management Engine - Forge

C2ME-Forge 0.2.0-forge.6 (1.20.1)

THE PERFORMANCE RELEASE — r1000 pregeneration on an 8C/16T machine: 118s -> 65s
(3.5x vanilla, was 1.93x; n=3: 65/66/64). Same worlds, verified block-for-block and
light-for-light against unmodified generation.

PERFORMANCE
- Recalibrated default worldgen worker count. The old formula left multi-core machines
  mostly idle (6 workers on 16 threads, ~46% CPU use). Out of the box: -15%; the value
  remains overridable via globalExecutorParallelism in c2me.toml.
- Aquifer fast path (concept-port of upstream PR#559): blocks provably above every nearby
  fluid level skip the aquifer entirely; per-grid-cell candidate reuse; lazy third-nearest.
  -6.5% wall; aquifer CPU share 18% -> 3%.
- Worldgen heightmap update skip (port of upstream PR#553): provably redundant per-column
  updates during the noise fill are skipped. -2.5%.
- Structure terrain adaptation (Beardifier) skips provably-zero piece/junction
  contributions via precomputed influence bounds (port of upstream PR#552, radii
  re-derived from 1.20.1). Chunk height limits cached at construction (PR#555).
- Noise cell drivers no longer allocate a lambda per block (~21.5% of all worldgen
  allocation eliminated; GC runs half as often during pregeneration).
- Adaptive worldgen feed controller (from the 1.18.2 branch): tunes Chunky's working count
  live from GC pressure and heap headroom; heap-aware seeding; protects small heaps.

SAFETY
- Per-mod feature lanes: mods that are (or depend on) libraries with a history of unsafe
  cross-chunk decoration writes (Blueprint, Lost Cities) get their features serialized
  behind a dedicated lane; all other mods keep fully parallel feature generation. Clean
  packs are unaffected. (1.18.2's blanket kill-switch, replaced by surgery.)

DIAGNOSTICS
- RAM advisor: the log now tells you whether RAM or CPU is your bottleneck during heavy
  generation and what would actually help.

All changes verified world-identical against unmodified generation via block-level and
light-level comparison at scale. Upstream ports used under MIT with attribution.