- 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
- Use the radial three-dimensional bound whenever rotateTowardsSlope is set, including placements whose configured rotation is disabled
- Group the support read margin helper with the other static factories
- Fell back to per-group grid sampling when a height sample nests inside another on the same thread
- Covered nested sampling with a bit-exact parity test
- Add PlacementReach: centered half-extent reach from loaded object keys, scale and shrink chains, enumerated rotation kernels, translate, warp and vacuum
- Size the object pass write radius and read radius from PlacementReach instead of full object width
- Size floating reach from scaled and shrunk object spans
- Count cave profile vacuum modes and upper-dimension effective rotation in the object reach
- Keep the transition-band footprint check on its existing footprint
- Add runner-backed and pack-wide reach tests
- Classify mixed cells without fade, boost or protection through a branch-light column loop
- Hoist the per-row bilinear terms shared by every column of a cell row
- Mark open cells from the pass bitset when every pass column is active
- Read interpolator corner bounds through a per-thread chunk block and skip groups whose cell corners are all zero
- Prefilled chunk context caches inline on generation threads with heights ahead of biomes
- Kept the burst-pool prefill only for the platform main thread, with heights and biomes in one task
- Painted surface fluid from the chunk context cache instead of re-sampling the fluid stream
- Decide whole 8x8 adaptive cells as solid or open from lattice corner bounds with an ulp guard
- Evaluate only mixed cells per column with the original bilinear arithmetic
- Share the row kernel between serial and parallel adaptive classification
- Count solid, open and mixed cells and exact rows, printed by the generation probe
- Pin carve write sequences for eight carver configurations with golden hashes
- Removed the unused decoration, derivative and identity streams and Engine.getBiomeID
- Skipped the shore-height noise when a column sits outside the proven shore band
- Dropped interpolator groups whose bounds are zero in every base-capable biome
- Stop writing cave ceiling markers, which nothing reads
- Write cave floor markers only when the dimension has cave spawners
- Reuse the zone floor subterrain cell for zone decoration
- Print measured-window stage totals from EngineMetrics as IRIS_GENPROBE_STAGES
- Time each context prefill task into its prefill timer
- Add gated generation counters for generator evaluations, corner starcasts and hydrology lookups
- Add -PprobeHydrologyPreplan to plan the probe area's hydrology tiles before warmup
- Add -PprobeHistoryRouter to generate through a generation-history router and registry contract
- Add -PprobeProductionBuffers to generate into production-equivalent block and biome hunks
- Print the mantle radius as IRIS_GENPROBE_MANTLE
- Place ellipsoid deposit clumps from the footprint cells instead of rebuilding the clump object
- Skip ore cave-biome lookups when no pack biome scales ore deposits
- Reject terrain-limited clumps before biome work when scaling cannot apply
- Compute the chunk terrain limit once per chunk
- Draw the exposure discard roll only when the discard chance is positive
- History coordinate scopes read the router and routing flag lock-free
- Subterrain planner shares bounded plan and chunk candidate caches
- Cave blend roll uses a stateless hash identical to the parallel RNG
- Custom biome resource keys are cached per dimension and biome identity
- Mod targets run through bounded Gradle workers with separate caches and daemon slots
- Target logs retain verification failures and bounded output summaries
- Nested tests use bounded forks and JVM thread counts
- Lifecycle tests control deadlines through injected clocks
- Timeout overrides retain original values for restoration
- Timeout restoration preserves operator changes and symbolic links
- Tests cover timeout persistence and failed configuration writes
- Runtime targets use separate Gradle caches and shared module outputs
- Modded build inputs include test fixtures and runtime cache settings
- Compact connected roots extend bounded gaps into supporting terrain
- Unsupported cliff placements leave no partial trees
- Root and canopy roles follow object scaling and rotation
- Canopies preserve solid trunk cells across chunk replay and saved continuations
- Branch wood replaces leaf clusters within each tree
- Regression tests cover slopes and overlapping tree placements
- Gameplay checks compare grounded spindle geometry across both packs
- Added DeferredPlacementRecords to gate tracking on Identifier and TileWrapper slices
- Cleared deferred custom and tile records at positions native structure placement changed
- Implemented block change tracking in the Bukkit and modded structure policies
- Covered invalidation through the custom block materialization pass
- Covered modded wrapper change tracking and protected cells
- Passed disabled change sets in existing wrapper tests
- Skip deferred custom block records whose live block no longer matches the provider carrier type
- Leave air-carrier furniture and unresolvable carriers to the provider unchanged
- Resolve each carrier once per chunk pass and log skipped records once per chunk at debug level
- Skip tile records quietly when the live block no longer matches the tile material
- Apply the same carrier check to the modded custom block pass
- Add Bukkit, CraftEngine and modded regression tests for flattened, replaced, intact and furniture records
- Truncated or malformed generation manifests and epoch metadata report as invalid history again after the strict reader change
- The datapack metadata test expects the range and supported_formats to start at the 1.21.1 format
- Merge feature/modded-versions into the modded fix integration
- Take loader pins from gradle/mod-targets.json for every target
- Pin the 26.3 NeoForge target to 26.3.0.83
- Build Fabric, Forge and NeoForge per target listed in gradle/mod-targets.json (1.21.1, 26.3)
- Add 1.21.1 adapter source overlays under adapters/*/src/versions/1.21.1
- Add verifyModdedMatrix and per-target Java runtime verification of loader jars
- Re-inject persistent dimensions from the levels-ready hook before the watchdog starts
- Wait for queued chunks and the final save before pregen completes
- Add the 1.21.1 data version and datapack supported_formats range
- Cascade regions whose sea or shore pool loses every biome
- Use Gson calls available on the 1.21.1 runtime
- /locate biome on non-stacked Iris dimensions searches the visible biomes chunks are filled with instead of falling back to the vanilla structure-biome search
- BiomeSearchFilter rejects columns whose region, land, sea, shore, cave, carving, subterrain or image-map biomes cannot hold a target, so terrain height is computed only for columns that can match
- Hydrology planning state is consulted only when a target comes from a river policy, so a locate no longer queues hydrology tile plans across its whole radius
- Admitted samples resolve inside their generation-history coordinate scope, so saved-biome reads for generated chunks no longer fail the command
- HintedLocator surface hints and IrisBiome visible key enumeration are shared with the locate filter
- Bukkit and modded biome policies share the same filtered search
- Modded create, enable and replace-overworld prepare worlds in the background
- Worlds still being created refuse duplicate create, delete, disable and update
- /iris create, /iris world create, /iris world enable and /iris world replace-overworld acknowledge at once, build the world in the background, and report success or the existing failure messages to the sender when it finishes
- Pack validation, the pack dimension check, generation history creation, pack data loading and the engine build run off the server thread; the server thread only registers the world, creates the level and publishes it
- A second create of a world that is still being created, and delete, disable or update of that world, are refused until it is live
- Server shutdown cancels in-flight creations, waits for an engine that is still building, and rolls back its level and registry entry
- Startup restore and world creation share one dimension injection orchestrator, and abandoning an injection also closes an engine that was already built
- Added the creating-in-background message to every locale file
- Refused /iris replace for dimensions that no level stem defines, including mod levels created while the server starts
- Skipped hand-edited dimensionReplacements entries for such dimensions at startup without moving their data or writing a level stem override
- Derived replaceable dimensions from the recorded worldgen catalog and the applied replacement state
- Recorded a replacement as applied only when the loaded level generates from Iris with the configured pack and seed
- Kept matching Iris generation history without recording it as an applied replacement
- Dropped applied records whose loaded level uses another chunk generator
- Stopped Iris engines bound to replaced levels that a mod swapped out during startup
- Listed installed packs and pack:dimensionKey forms when the replace type is unknown
- Added the refusal and installed-pack messages to every locale
- Modded /iris replace stages vanilla, Iris and mod dimension replacements with Paper syntax
- /iris world takeover removed; replacements live in modded.json dimensionReplacements
- Added /iris replace with override and overwrite aliases on Fabric, Forge, and NeoForge using the Paper target, type, and seed arguments
- Resolved main, nether, end, level-name aliases, iris: and bare names, and any loaded mod dimension as replacement targets
- Kept the saved seed when seed is omitted and stored explicit seeds per replaced dimension
- Applied replaced vanilla and mod dimensions through the forced datapack level stem with the configured seed bound at engine creation
- Applied staged Iris world replacements at startup by moving their terrain aside and repointing the dimension registry
- Refused a second staging while a replacement is pending and checked pack environments for vanilla slots
- Added the modded-only off form to cancel a staged replacement or return a replaced dimension to its own generator
- Reset the Overworld spawn and moved returning players out of unsafe positions after an Overworld replacement
- Renamed the takeover reconciler, model, config key, state file, and recovery folder to replacement names
- Removed /iris world takeover and its help entry and localized messages
- Saved Iris dimensions are prepared in parallel during world load (server-about-to-start on NeoForge and Forge, server-starting on Fabric), before the server watchdog runs
- Server start only creates the restored levels, builds their engines in parallel, and publishes them, so its wait no longer grows one world at a time
- Each restored dimension opens its generation history once instead of twice and resolves its active pack once instead of three times
- An epoch snapshot whose content matches its validated source pack reuses that validation instead of running the full pack validator for every world
- Level creation reads the configured pack and structure biome keys from data warmed during preparation
- Immutable generation packs feed the prepared hydrology cache identity from the data's memoized fingerprint instead of hashing the pack twice per engine build
- Restore logging reports each dimension's preparation and engine time and how long the server thread waited
- Themed regions, those at least five times rarer than the dimension's median region, no longer receive adopted biomes while a common region accepts the biome's role
- Themed regions still take a mod biome when no common region has a pool for its role or every common template is full
- Category affinity now chooses among common regions, so a rare biome is no longer locked inside a region that can be absent for thousands of blocks
- Each pack biome now serves as the template for at most five adopted land biomes and three adopted sea, shore, or cave biomes
- A mod biome whose best template is full falls back to the next best template under the cap
- A mod biome that finds every compatible template full is skipped with a reason naming the cap instead of being dropped
- Converted patch decorators also lose the share of surface columns that earlier surface patches in the declaring biomes already occupy, estimated from those patches' expected coverage in the catalog
- Canopy and surface occupancy are combined into one vegetation occlusion estimate per decorated feature
