- 26.3 custom biomes write natural_mob_spawns as an overlay modifier with its argument, matching vanilla biomes
- NeoForge biome modifiers can add spawns to Iris biomes again
- The 26.3 custom biome codec test checks the overlay modifier
- PregenChunkRetention holds every chunk in a FULL request's dependency square until all in-area regions reaching it drained; drained regions keep only the border strips undrained neighbours need
- Removes the unload/reload churn (about 1.5 reloads per generated chunk at r1000: lattice passes, next region's border, future-region and ring protos)
- Capped at one chunk per MB of heap; over the cap the drained strips farthest from the active region are released; heap pressure and close release everything
- Regions still save and unload at drain; the early-unload test now pins that plus border retention
- Limit terrain cleanup to affected columns and batch cave boundary reads.
- Reuse offline generation workers and native air states.
- Gather history writes and defer unpublished append synchronization until atomic publication.
- Cover terrain parity, worker lifecycle, history durability, and publication failures.
- Packages regrouped by domain: generation, world, pack, studio, structure, platform, command, localization, diagnostics
- Forked VolmLib utilities dropped for the shared library: noise, stream, interpolation, hunk, matter
- Engine-coupled stream decorators moved to generation.stream.GenerationStreams
- Hunk section splits now run on the Iris burst pool through HunkBurst
- Artifact entry lists, purity allowlists and shrink rules follow the new paths
Surface hydrology policy split, native terrain height cache, watched-edit
hotload cancellation, world-save requestSave with tracked ownership flush,
background-task drain that follows saves admitted mid-drain, peaceful spawn
guard, cached-completion terminal state, IO drains off the owner thread.
Studio pregen at scale: detached historical runtimes keep the configured
noise cache instead of the Studio-wide cache; retired runtimes are no longer
pinned by the volumetric fragment filter's per-thread scratch; pregen mantle
cleanup and native region flushes run on per-run threads instead of the
server thread and the shared Iris IO pool (the flush on the IO pool starved
tectonic plate loads and wedged every Paper worker in Mantle.getChunks);
background-task cancellation claims tasks atomically.
- Move art.arcane.iris.testsupport into core/src/testFixtures and apply java-test-fixtures to core
- Consume core test fixtures from the bukkit plugin, nms, probe and the three loader adapters
- Delete the duplicate ProjectPaths copies in the bukkit-plugin and modded-common test trees
- Point ModdedMessageContributorTest at the shared ProjectPaths
- Guard CommandPackCompatTest, NativeTransitionColumnTest and StubPlatformStateTest with PlatformLeakGuard
- Release the probe runtime state after DimensionStackSuperflatProbeTest opens a workspace
- Keep core main classes ahead of the core jar on the test classpath so the bytecode gates still scan a directory
- Add the Await test helper that polls a condition every ten milliseconds under a bounded timeout
- Replace every Thread.sleep in the fourteen polling test classes with a condition wait, a latch wait or a bounded quiet window
- Keep the crash-consistency fsync interception in GenerationHistoryAdmissionTest and SavedBiomeStoreBatchTest on real barriers
- Removed the core processResources skip that fired on test-only invocations, which packaged a core jar with no generation kernel catalog and made all 10 probe and bukkit-plugin tests that reach GenerationKernelRegistry fail on every ./gradlew test
- Set maxParallelForks on :core:test to half the available processors, overridable with -PtestForks, and left the small modules single-forked so they do not re-pay the Minecraft bootstrap per fork or steal worker leases from core
- Resolved B.registries() per bound platform instead of latching the first one for the JVM lifetime, keeping the previous unbound behaviour, so a test that binds its own platform no longer reads a foreign platform's block states
- Resolved IrisObject.States per bound platform through accessors instead of static finals fixed at class-init, removing the NoClassDefFoundError and null material-key failures that appeared whenever class partitioning changed
- Bound an explicit platform in EngineMantleCoverageTest and IrisDimensionStackActuatorMetadataTest so they no longer depend on another class leaking a binding first
- Gated BukkitSpiConformanceTest registry probes on a live RegistryAccess and restored the IrisPlatforms binding it leaked, which was silently swallowing the stderr two SettingsHotloadWatchTest assertions read
- Initialised org.bukkit.Registry and Biome inside the RegistryAccess mock scope in ObjectStudioGeometryTest so Mockito no longer instruments Registry mid-clinit
- Replaced the SettingsHotloadWatchTest fixed 1200ms diagnostic window with a 10ms condition poll bounded at 4s
- Replaced the IrisCaveCarver3DNearParityTest wall-clock speedup benchmark with a deterministic density-sample count, cutting 28 carves to 2
- Cut HydrologyPlannerTest cold-owner parity from 12 tile plans to 4 and shared the immutable baseline plan across the early-owner tests
- Built one prototype project or staged history per class in the jigsaw studio and world replacement tests and copied it per test, replacing repeated fsync-bound fixture construction
- Replaced Mockito block-state fixtures with concrete implementations in the object IO, deposit modifier and nearest-block-index tests, removing about 2.5 million interceptions
- Replaced Thread.sleep with bounded condition polling in HydrologyTileCacheTest
- Deleted 19 test classes and 60 further test methods that asserted on production .java source text rather than behaviour, and removed the 19 now-unread iris.*Source system properties that fed them
- Terrain and cave scans skip repeated work
- River and structure lookups reuse cached results
- Chunk history saves share disk writes
- History lookups continue during disk writes
- Failed saves retain their original errors
- Deposit failures stop incomplete chunks
- Tests check output and save safety
- Native structure POI changes persist dirty chunk sections before deferred updates
- Chunk load reconciles final POI states on the owning thread
- Matching POI records retain claimed tickets
- Native flag detection keeps unsupported server behavior intact
- Regression tests cover persistence and failed repair retries
- Cold startup and hydrology changes integrate with terrain history
- Flat authoring worlds skip native features under the history write guards
- Native decoration checks retain recorded structure and terrain ownership
- Automatic build revisions replace manual source seals
- Flat authoring studios skip terrain and native structure preparation
- Normal Studio entry prepares the exact hydrology column asynchronously
- Hydrology prefetch admits the complete frontier before stale queued work
- Concurrent hydrology loads avoid cache monitor contention
- Cave density tasks drain failures before releasing chunk state
- Terrain and interpolation queries reuse exact bounded samples
- Pregeneration counts requested chunks directly from its bounds
- World-local pack epochs preserve generated chunks across restart-only updates
- Kernel catalogs and source seals retain executable generation behavior across jar changes
- Crash-safe ownership, boundary, registry, semantic, and mantle records survive activation changes
- New terrain blends lower surfaces and inverted ceilings against frozen historical boundaries
- Cave, object, structure, biome, and hydrology work respects activation ownership
- Find commands merge recorded facts with eligible active predictions for biomes, regions, structures, objects, and rivers
- Bukkit, Fabric, Forge, and NeoForge route generation through retained runtime scopes
- Shutdown ordering preserves class-loader resources until world and service cleanup finishes
- Tests cover activation chains, failure recovery, retained registries, bounded lookup, and transition geometry
Vanilla structure and feature placement now runs through the bounded
worldgen access with a protected-position predicate, so a jigsaw piece
or an imported feature can read a static object's blocks but never
overwrite them, and never schedules a tick or block entity there.
Dimensions without static objects keep the plain level as before.
