- Structure volume caches read through concurrent maps and publish only current builds
- Natural terrain receipts reuse a per-thread deflater with the platform gzip framing
- Tests pin receipt and saved biome bytes across reused compressors
- Terrain batch scratch marks reset in finally blocks so a failing kernel cannot grow the scratch
- Generator evaluation counters only count when generation counters are enabled
- Core tests run with generation counters enabled
- Bake capacity derived from the heap (13 KB per entry against 1/32 of maxMemory, 2048 to 32768 entries); detached runtimes use the floor
- Pregeneration profile no longer raises the global noise cache size or the bake; it only raises burst parallelism
- Removed the hydrology bake reservation, noise cache budget, expansion machinery and the tile cache terrain preparation hook
- Bake thread recents compare a generation counter; clear empties retired entries
- Removed unused column bake, bake registry and bake accessors
- Object placement bakes split into one per-object layout (palette, tiles, canopy) and per-variant offsets
- Variants beyond the per-object limit are built transiently instead of clearing the variant map
- High-variant golden and no-rebuild tests
- A caller whose submission window is full runs the next component itself instead of parking
- Pass barriers run still-queued component tasks in reverse order before waiting
- Queued component tasks reopen the submitting generation's runtime scope on whichever thread runs them
- Helpers only run queued tasks bound to their own runtime or when they hold none
- A finished submission no longer reports a task as dropped once a helper started it
- Burst worker scope test keeps one stage on a worker while the caller helps
- Cell lattice passes gather every missing corner sample of the pass and evaluate base, detail and modules through NoiseProgram.fillPoints3D
- Warp B samples and single-row warp A samples are evaluated as one point batch per pass; longer warp A runs keep fillColumn3D
- Module bands compact their points before the batch call and scatter back in the original accumulation order
- Lattice batch buffers and their NoiseScratch live in a per-thread CaveLatticeBatch chain that nested carves acquire separately
- Lattice test compares pass-wide point batches against scalar plane density bit for bit
- Hydrology planner height reads reserve bake capacity up to the hydrology cache target from the shared budget
- Planner reservation is released when the complex closes
- Log transaction writes in parallel arrays instead of mutation records
- Replay source plans straight from their packed cells without decoding records
- Build source plans from a log slice
- Journal commit originals in a packed-key set instead of a linked map
- Buffer cave object writes in parallel arrays
- Drop the now-unused import in the runner
- Hydrology planner height and natural samples memoize columns in the bake like other reads
- Batched planner base heights publish into the base plane
- Removed the direct-mapped sparse base height memo
- Only bounds lattice biome reads stay out of the bake
- Memoize the last subterrain protection verdict in object destination transactions
- Key buffered cave object blocks by packed positions and check each position once at commit
- Build every free source plan before waiting on plans other threads are building
- Reuse TreeBlockMaterial values per block state
- Layer and generator programs sample single points through NoiseProgram.sample2D
- Scalar interpolator group means use the compiled programs when the seed matches
- Single-point program heights are property-tested against the generator path
- Add SectionBuffer: chunk palette of written instances and short[4096] ids per section in y, z, x order
- Column tops are indexed per x-slice and padded so fan-out slices do not share a cache line
- Add ColumnRunHunk and forward column runs through the listening hunk and block update listener
- Terrain painter emits column runs and evaluates ores only inside the column's ore bands
- Bukkit generation writes into SectionBuffer and applies it through ChunkDataOutput
- Remove ChunkDataHunkHolder and the probe's ProductionBlockHunk copy
- Probe production buffers use SectionBuffer
- Bake rotated block offsets, palette indexes, tiles and canopy bits per object and spin kernel
- Resolve decay, rotation, waterlog and air facts once per palette entry per placement
- Skip the destination read for waterloggable states that cannot become waterlogged
- Move SpinKernel to its own file
- Facade reads reuse the thread's last chunk entry instead of probing the bake map
- Stores and full-chunk fills refresh that entry from the map
- Surface lattice lookups reuse the previous point's cell indices
- Rarity stream picks are property-tested against the VolmLib rarity selection
- Pregen capacity growth scales the bake by the same factor as engine start
- Cell kernel passes fill a per-pass density lattice in field-major batches before classifying rows
- Lattice density columns are shared between profiles with identical density fields
- Warp samples are shared between profiles with the same warp field and batch along lattice columns
- Added lattice read, sample and warp counters
- Added bit-exact lattice column and field sharing tests
- Rarity selection streams pick from precomputed cumulative tables
- Implode and surface biome fixes take primitive coordinates
- Child selection reuses one derived child seed instead of a new RNG per level
- Base biome selection reads bridge and inferred streams directly
- Planes fill lazily per column with ready bits and batch only missing columns on full-chunk requests
- Natural fills batch only the base heights Terrain3D still needs and reuse the sparse memo
- Bounds lattice and hydrology planner reads stay out of the bake
- Bake capacity is eight times the noise cache, covering the measured working set
- Generator evaluation counters and a bake census print from the probe
- Parity test compares batched chunk base heights and evaluation counts with the scalar path
- Generation pool sized to performance cores
- Pregen burst raise sized from performance cores
- Hydrology planner pool sized to performance cores, at least four
