LCull

Performance

• New pre-folded plane test on the entity path. The six frustum planes are re-derived once per frustum rebuild from Frustum#matrix (Gribb-Hartmann row combinations plus JOML's own invsqrt normalisation, and no JOML member is ever touched), and the box radius is folded into each plane's w, collapsing testAab and its 18 per-entity min/max sign choices into a branchless dot product. -26% (25.38 to 18.69 ns/entity) on the 10k-entity JMH scene, camera rotating, 3 forks. A per-plane early exit was measured and rejected at 21.5 against 17.0 ns in isolation, because which plane rejects is effectively random, so every early exit is a mispredict. (1.19.3+) REWORK
• Cull decision order now mirrors vanilla's own, running players, then the Entity#shouldRender distance test, then the "never cull me" opt-out, then near-camera Displays, then the margin, then the frustum. Hoisting the distance check above the opt-out dropped the opt-out from 18.1% to 1.2% of the cull path. (all versions) REWORK
• Frustum#isVisible uses testAab instead of intersectAab(...) == -2 for -8%, and it also serves the block entities, particles and chunk sections that share that method. (1.19.3+)
• offsetToFullyIncludeCameraCube is now bounded. It starts at vanilla's own 4.0 step so near-camera behaviour is unchanged, then grows that step x1.6 per iteration, caps total travel at cubeSize * 8 and caps iterations at 12, where vanilla steps a flat 4.0 forever. That is about 6 plane tests against the 16 vanilla needs at 70 degrees FOV and the 185 it needs at 10 degrees. An earlier revision decayed the step instead (x0.75, floor 1.0) to land precisely near the camera, which was backwards for this loop whose whole purpose is to over-include near the camera, and it cost ~47 intersectAab calls against vanilla's 16. (all versions) REWORK
• Frustum copies rebuild their planes too. Vanilla's Frustum(Frustum) constructor never calls calculateFrustum and copies intersection, matrix and the camera fields straight across, so without a second inject at the copy-constructor TAIL the pre-folded planes were never built on any copy. That matters because copies are exactly what the camera-cube walk and the particle engine receive, LevelRenderer#offsetFrustum being new Frustum(frustum) followed by offsetToFullyIncludeCameraCube(8). Costs about 200 ns per rebuild, or 0.022 ns per entity over a 10k pass. (all versions) CRITICAL
• The camera-cube walk now runs the pre-folded "fully inside" test instead of intersectAab, so the widened volume feeding LevelRenderer#applyFrustum no longer allocates a fresh AABB per plane test. (1.19.3+)
• Two candidates were measured and rejected. A bounding sphere before the AABB test over-includes, so it rarely rejects and only adds work. A quantised sqrt-free margin came out +14% worse because the threshold chain's unpredictable exit mispredicts more than the square root costs. (all versions)
• testAab is kept in plain isVisible(AABB) even though the pre-folded test is cheaper, because that method also serves block entities, particles and anything else that asks the frustum a question, and making those decisions slightly more permissive is not LCull's call. (1.19.3+)

Bug Fixes

• Forge refmap fix. lcull.refmap.json is now wired into the processed mixin config, because production Forge remaps member names to SRG and without a "refmap" key in the config every @Inject failed on an obfuscated target. (1.19.2, 1.20.1 Forge) CRITICAL
• Forge loaderVersion was hardcoded [47,), which cannot load on 1.19, so a 1.19.2 artifact could never be enabled. It is now major-based, [43,) for 1.19.x and [47,) for 1.20.x. (1.19.2, 1.20.1 Forge)
• The generated manifest now honours deps.minecraft_min instead of always using the Stonecutter version as the lower bound, so the 1.19.2 jars declare [1.19,1.19.2]. (1.19.3+)
• Fixed 26.3 API drift. 26.3 added a trailing float partialTicks to shouldRender and moved it into getBoundingBoxForCulling(T, float), so the injector carried the old signature. The >=26.3 handler stores partial ticks in a @Unique field consumed by the box read. (26.3)
• Fixed 1.21.2 API drift. EntityRenderer#affectedByCulling(T) only exists on 1.21.2+, so below that the "never cull me" opt-out now reads the public Entity#noCulling field. Without the gate, the ender dragon, lightning bolts and noCulling displays were culled. (1.21.2-) CRITICAL
• Fixed 1.19.2 API drift. Display only exists on 1.19.4+, and so does Entity#level(), where lcull$level falls back to getCommandSenderWorld() below that. (1.19.2, 1.19.3)
• MClientChunkCache kept both redirects on 1.19.2, where replaceBiomes does not exist. Only the replaceWithPacketData redirect is kept there now, still require = 0. (1.19.2)
• Fixed the Forge 1.20.1 launch. The one hand-written per-version source is the pack.mcmeta declaring format 15 under versions/1.20.1-forge, and Loader.FabricLike.excludedResources filters it out of Fabric jars while NeoForgeLike and Forge filter the foreign manifests, so no jar ships two pack formats or a foreign mods.toml.
• The pre-folded plane tables are now refreshed through Frustum#set(Frustum) on 26.1 and newer, where the copy constructor delegates to it, and through the copy constructor below that. Only the copy constructor was hooked before, which left the persistent CameraRenderState#cullFrustum object holding the planes derived from the identity matrices it was constructed with, so any mod asking that object for visibility would have been answered from a degenerate frustum. Vanilla never asks it, so nothing was broken in practice, but the state was reachable. Hooking both would have been worse, since the copy constructor reaches set on 26.1 and every copy would then be rebuilt twice per frame. (26.1+)
• Removed the WARNING! LCULL COULD PRODUCE NEGATIVE PERFORMANCE! startup line, which printed on every launch and referred to a measurement that in fact produced the shipped margin numbers.

Removed

• The per-entity cull cache is gone. ICache (duck) and MEntity (mixin on net.minecraft.world.entity.Entity) are dropped from lcull.mixins.json5. The cache stored a last-decision boolean plus an entity-position and a frustum signature per entity, and the validation costs more than the test it guarded, since a frustum signature is 24 float compares plus camX/Y/Z against a cached test of 18 multiplies, 6 adds and 6 compares, and a miss is every frame the camera moved at all. It only paid when the camera was perfectly still, in which case the whole cull path is already a rounding error in the frame. It also put a field on Entity, the hottest class in the game. (all versions) REWORK
• The section-visibility mirror is gone. MLevelRenderer and ILevelRenderer short-circuited the cull path on one hash lookup of a copied LevelRenderer#visibleSections, but vanilla already gates entities on section visibility itself through extractVisibleEntities and isSectionCompiledAndVisible, so it duplicated an existing check at 16-block granularity, and the section path only existed on 1.21 through 26.1, silently degrading to a no-op elsewhere. (all versions)
• LevelRendererAccessor removed earlier for the same reason, a hook surface with no real consumer.
• The lcull.ct and accesstransformer.ct templates were removed as unused. LCull still ships no access widener and no access transformer, since all mixin targets are already reachable and @Shadow covers private fields.

Changed

• New supported versions 1.19.2 (Fabric + Forge) and 26.3 (Fabric + NeoForge), 20 variants in total, with stonecutter.properties.toml, settings.gradle.kts, the per-version Java toolchain (Java 17 below 1.20.2, 21 for 1.21.x, 25 for 26.x), the JOML versus Mojang-math split and the version-gated mixin signatures. (1.19.2, 26.3) NEW
• Speed-scaled cull margin, 1.0 at rest and +0.2 * |deltaMovement| above 0.5 blocks/s, hard-capped at 2.5 so a fast projectile cannot exempt itself by moving fast. This costs +7.4% more drawn entities than vanilla at rest and up to +26.5% for fast movers, and every survivor costs a render state, a model build and a draw call, which dwarfs the ~20 ns test that rejected its neighbour. It is the deliberate trade for no mid-stride pop-in, and lowering it remains the single highest-value FPS lever available. (all versions)
• Renderers reporting a NaN or zero-size culling box now fall back to a fixed box of radius 2.0 around the entity, matching vanilla's own fallback, instead of testing a degenerate AABB. (all versions)
• The leash escape is preserved on 1.21+, where vanilla keeps a leashed entity visible while its holder is so the rope does not visually detach, and both of vanilla's tests are kept because the union is not redundant, two boxes on opposite sides of a plane being able to straddle it while neither intersects. Below 1.21 the check is a stub, because net.minecraft.world.entity.Leashable does not exist there and vanilla had no equivalent escape. (1.21+)
• Near-camera Displays within 5 blocks defer to vanilla, since their transformed boxes cannot be trusted, and use a flat 1.0 margin beyond that. (1.19.4+)
• MFrustum splits on the 1.19.3 JOML transition. 1.19.2 predates JOML, so it shadows vanilla's private cubeInFrustum and cubeCompletelyInFrustum plane walkers, which allocate up to 48 Vector4f per call, and keeps vanilla's frustum assembly. The pre-fold is deliberately not attempted on 1.19.2, because although its planes are readable, no 1.19.2 class is on the benchmark classpath, so the plane sourcing and sign convention would ship unverified. (1.19.2)
• Mixin priority raised from 600 to 2100, so LCull's decision runs before default-priority mixins from other mods. (all versions)
• offsetToFullyIncludeCameraCube is now an @Inject at HEAD with cir.setReturnValue and cancellable, rather than an overwrite, which would have had to restate vanilla's whole frustum assembly and need a separate signature per version family. calculateFrustum is likewise not overwritten. (all versions)
• neoforged.moddev 2.0.141 to 2.0.147, with NeoForge for 26.3 pinned to 26.3.0.16-beta.
• neoforged.moddev.legacyforge now builds both Forge variants, pinned to 1.19.2-43.2.0 and 1.20.1-47.4.10.
• Fabric below 1.21.11 and NeoForge below 1.21.11 get a compileOnly JSpecify 1.0.0 dependency, and Forge gets it unconditionally plus the mixin annotation processor.
• Full LGPL-3.0-only metadata, with the license key emitted into fabric.mod.json and both mods.toml variants, the full LICENSE copied into every jar root, and the header block maintained by the root licenseHeaders task, which now auto-runs before each variant's compileJava. SPDX is deliberately LGPL-3.0-only, version 3 exactly and no "or later".
• README gained the platform badges, the project icon and the AI-assistance note.
• LICENSE reformatted to the FSF's verbatim LGPL-3.0 text. The shipped terms were correct, the formatting was not the FSF's own.