- Every getChunk(..).use() site calls Mantle.useChunk, which cannot fail because a save closed the chunk in between
- Covered-chunk cleanup, custom block and static object layers take their chunk through useChunk
- Discovered flag writes go through Mantle.flag instead of an unpinned chunk
- Test mantles stub useChunk the way the real one behaves
- A demanded owner missing from the owner cache is restored from its planned tile (resident or prepared-plan store) instead of drafted again
- Same restore path as persisted tiles; plans identical
- Plans real pack tiles headless and prints a digest per tile with planning CPU
- Optional pregen radius plans them under a pregeneration scope and prepared-plan store
- Expected-digest file fails the run on any changed plan
- ChunkBlockUpdateListener remembers the last state with no update or custom effect
- Plain terrain runs skip the per-block isUpdatable and isCustom checks
- Test replays random writes against the per-write catch semantics
- saved biome capture resolves each quart column through one engine call
- column-invariant height, region, surface and cave fallback resolved once
- cavern reads use one pinned mantle chunk per column
- saved biome cells interned per capture
- engine getters borrow the bound route runtime for other chunks instead of opening a stage
- Pregeneration scope limits the lower-rank neighbour drafts an owner starts early to tiles the area composes from
- Undemanded ring drafts (and the early neighbours they started) are no longer computed; demanded ones still resolve on demand
- r2000 tile set: 80 -> 54 drafts, planning CPU -32%, plans identical
- dimension stack metadata clearing and continuation restore rewrite cavern cells after sCave, so they drop the context snapshot and decoration re-reads the mantle
- test that decoration reuses only its own chunk's snapshot
- boundary columns build runs directly instead of voxel lists, tree sets and hash maps
- palette sorted once per column with a direct comparator, same canonical order
- geometry source reuses the last encoded voxel
- equivalence test against the previous encoding
- ColumnExtentListeningHunk reports writes through a primitive listener instead of boxing ListeningHunk
- Palette layer generators memoized per seed slot instead of one shared recent entry
- Paint chunks z-major so neighbouring columns share block-buffer cache lines
- Fill the solid run below layers, river paint and cave ceilings with ores or rock in one loop
- Resolve chunk constants, air and bedrock once per chunk; drop the unread shared lastBedrock write
- IrisOreBands replaces IrisOreGeneratorBounds: per-Y candidate generators in list order
- Resolve a constant (FLAT) ore chance once instead of sampling it per block
- Column parity test against the per-block algorithm over randomized columns
- close reads each window slot before the atomic swap instead of swapping all 5,929 slots
- terrain-ready flags live with the terrain access and exist only on content writers
- Over the plate cap, the least recently used idle plates are saved one at a time on the IO pool while submissions continue
- Submissions wait only past the cap plus a small overshoot
- Replaces trimming and unloading every idle plate, which evicted the active working set and read it back
- eager terrain window covers the replayed object sources plus one chunk of footprint (radius 13 instead of 19 for overworld)
- content writers generate a chunk's missing terrain components before handing it out, so reads past the window stay identical
- getEagerInputRadius replaces isInputGenerationLazy
- sCave copies the chunk's cavern and hydrology terrain view once (journal originals win) instead of per-voxel monitor, journal and slice lookups
- hydrology lookups answer null without work in chunks with no cave cells
- decorateNaturalCaves reuses the sCave snapshot through the chunk context
- Each thread keeps its recent final hasColumnsAround answers next to its chunk slots, so the object and hydrology gates skip the nine-chunk scan on repeat
- frames live in a 256-slot table indexed by thread id, ThreadLocal only on slot collisions
- the single ThreadLocal probed long collision chains in the workers' thread-local maps
- Index planned tiles by occupied chunk; empty chunks skip composition, tile residency and cave-cell lookups
- Plan the whole pregen area nearest first with several roots, past the tile budget when plans persist
- Restore persisted tiles without a planning root permit
- Never memoize answers the server thread gave before its hydrology plan landed
- Semantic journal and saved-biome appends write without forcing; a shared flusher forces touched files within 2s outside every store lock
- GenerationHistory.sync forces pending history at world save, engine shutdown, pregen completion and before activation promotion
- A failed deferred force poisons the store the same way an uncertain append did
- Cache the active epoch pack root instead of regex-validating it on every stage open
- Tests pin append-without-force, sync barriers, failure poisoning and the save, shutdown and pregen hooks
- Ellipsoid footprint dry run replays the build draws without building
- Terrain-scoped clumps above every column limit are dropped before building
- Surface biome filters no chunk column passes skip the deposit
- parent biome lookups read the biome stream directly instead of building the boundary field
- fields are only built where a parent with floating children needs edge distances
- Bukkit noise generation keeps one biome per column until a partial column write
- drops the full-resolution biome array allocation and fill per chunk
- Cache vanilla withProperty variants per state
- Cache rotated states per rotation and quantized spin
- Hoist the stilt dirt state and rotate only target-layer stilt blocks
- Store source plans as packed per-destination cell runs over a value table
- Weigh shared payloads once and give tiles and markers finite weights
- Release a plan once every touched destination consumed it, keeping its index
- Rebuild only on repeated consumption; debug stats for builds and waits
