Isekai v2.1.0

Three features that the documentation promised but the implementation never
delivered are now real: biome zones vary with the world seed, count_scale
thins structures, and TransitionRule blend/gap are evaluated.

No Java API breaks. Everything added is a default method or a new static,
so variants written against 1.x / 2.0.0 compile and run unchanged.

Changed — datapack behaviour

• isekai_api:noise_threshold and isekai_api:edge_jitter now vary with the
  world seed. Their JSON seed field is unchanged in shape but changed in
  meaning: it was an absolute noise seed, and is now combined with the world
  seed via BiomeZone.deriveSeed. Two worlds made from the same datapack with
  different seeds get different patterns; the same seed reproduces. Worlds
  created before 2.1.0 will generate a different pattern in newly explored
  chunks — finish or retire such a world before updating.
• Combinators (and / or / not) propagate the world seed to their children.
  Purely geometric zones (always, y_above, y_below, y_between,
  within_distance, beyond_distance) are unaffected by seed, as before.
• A world's terrain shape is not seeded by this change. Density functions
  built only from isekai_api:distance / repeat / add / max and constants
  stay identical in every world; seed variation there comes from including a
  vanilla minecraft:noise term, which the world seed already drives.

Added

• structure_strategy accepts isekai_api:count_scale. A factor below 1.0
  deterministically thins vanilla structures from (world seed, chunk x, chunk z, structure id). It vetoes candidates at Structure.findValidGenerationPoint
  rather than touching RandomSpreadStructurePlacement.spacing: that placement
  object is shared by every dimension referencing the structure set, and
  ChunkGeneratorStructureState caches the candidate grid at world load, so
  rescaling spacing would leak across dimensions and break seed reproducibility.
  The consequence of the safer route is that structures can be made rarer but
  not more common — a factor above 1.0 is a validation error, not a silent
  round-down.
• TransitionRule.Blend and Gap are evaluated at layer boundaries.
  Blend(b) spreads the seam across a b-block band whose per-position choice
  between the two descriptors ramps from 0 to 1, turning a flat seam into a
  grainy gradient. Gap(g) leaves the top g blocks of the lower layer
  unclaimed. Blend affects a descriptor's positional output — block_overrides
  (surface_top, default_block) and structure_predicates — not terrain
  shape, and not ore/mob/feature remap (those select a layer by biome, not Y).
• BiomeZone.withWorldSeed(long) and BiomeZone.deriveSeed(long, long).
  withWorldSeed defaults to returning this, so third-party variants keep
  working; a third-party combinator that does not override it will not
  propagate the seed to nested built-in noise zones.
• IsekaiRemap.getDescriptorAt(dimension, x, y, z) as a default method
  delegating to the existing Y-only form.
• A server-start warning when a rule biome source never received a world seed.

Changed — isekai_api:pool behaviour

pool used to carve, floor and fill unconditionally, so a basin that straddled a step
in the ground ended up with its water surface above the neighbouring ground and drained
over the lip. It now follows LakeFeature's shape: it tests the shell for solidity
before writing anything and returns without placing a single block when the site
cannot hold the fluid, it lines the whole shell with rim_block rather than only
the floor, and its outline is a union of several random ellipsoids instead of a
mathematical disc. The ice-capping pass is deliberately not ported — it calls
getBiome() up to 15 blocks out, which throws during worldgen.

No codec change: fluid / rim_block / xz_radius / depth are untouched and 1.x
datapacks decode unchanged. Placement gets rarer on uneven ground — measured at
roughly 2–6x fewer successes on rolling terrain, and a wider xz_radius costs more
than a deeper depth — so raise the rarity_filter chance rather than the radius if a
world wants the old frequency back.

Fixed

• exclusions.features did not hold. The REMOVE phase deleted the listed
  features and the ore-remap ADD phase put them straight back: it re-injected
  every snapshot feature whose placement exposed a HeightRangePlacement —
  which is not only ores — and did so as an anonymous Holder.direct, with no
  ResourceKey for any later pass to match. A world that excluded
  minecraft:spring_water still generated springs, and nothing in the datapack
  could remove them. Measured on a generated Sky World save: 437 water sources
  and 18,247 lava sources across 9,409 chunks, all from excluded features.
  REMOVE and ADD now derive their target set from one function
  (impl/RemapTargets.select), so the two phases cannot drift apart again.
  Carvers and structures were checked for the same shape and have no equivalent
  path. additions.* is deliberately unchanged: naming the same id in both
  additions and exclusions still lets additions win, since that is an
  explicit-versus-explicit conflict rather than an unrequested re-injection.
• IsekaiValidator rejected every non-identity structure_strategy while its
  own comment said it warned.
• The layered-worldshape reader documented a file-level "transition" field in
  its javadoc; LayeredFile has no such component and the key was silently
  discarded. The layered_overworld example wrote it, put blend on the
  topmost layer (which has nothing above it), and used a structure_strategy
  the validator rejects.
• README.md reported 3 SurfaceAnchor and 7 RemapStrategy variants; there
  are 4 and 8. Its two BiomeZone counts disagreed with each other (9 and 11);
  there are 11.
• examples/2_placement/moon_world/worldshape.json used the deprecated
  isekai: prefix.
• tools/check_since.py read return foo(...) inside a method body as a
  declaration.

v2.0.0

The rule-adaptation layer becomes an open extension point, terrain shape gets a
copy-free hook mechanism, and the library ships as a proper Gradle dependency.

Datapacks are unaffected. Every 1.x datapack loads unchanged: the isekai:
dispatch prefix is accepted as a deprecated alias (logs one warning per id), and all
existing payload schemas are byte-for-byte compatible. The breaking changes are on the
Java API only (see below) — this is a major bump because the SPI interfaces changed
shape, not because datapacks broke.

Breaking (Java API)

The five in-house dispatch interfaces — SpatialPredicate, RemapStrategy, BiomeZone,
SurfaceAnchor, TransitionRule — are now registry-backed extension points instead of
sealed codec unions:

• sealed is removed from all five. Exhaustive switch statements over their variants no
  longer compile — handle them through their behavior methods or a default branch.
• String typeId() is removed from all five. To recover a variant's registered id, ask
  the registry: registry.getKey(variant.codec()). The registry keys live in
  api/registry/IsekaiRegistries.
• New abstract methods carry each interface's evaluation contract:
  SpatialPredicate.test(EvaluationContext), RemapStrategy.remap(VerticalRange, RemapContext),
  SurfaceAnchor.resolveY(PlacementContext, BlockPos). Custom Java implementations must
  supply them.

Datapack authors and consumers who only use the built-in variants from JSON are unaffected.

Added

• Open SPI. Third parties register their own variant of any of the five interfaces from
  their own mod id by adding a MapCodec to the matching registry in
  api/registry/IsekaiRegistries — no fork, no PR. See
  docs/DATAPACK_REFERENCE.md → "Extending the SPI".
• api/predicate/EvaluationContext and api/remap/RemapContext — the world-access and
  math seams that variant test / remap implementations receive.
• isekai_api:hooked_overworld preset + hook/final_density density function. A new
  dimension references the preset and overrides one small hook file to change terrain shape —
  no 2500-line noise_settings copy. A whole floating-island world is two files.
  (README)
• Terrain-relative ore remap — isekai_api:column_local (RemapStrategy),
  isekai_api:world_floor (SurfaceAnchor) and isekai_api:column_relative (placement
  modifier). Every other remap strategy produces one absolute Y band per feature, which can
  only ever be right for terrain at one altitude; column_local normalizes a feature's
  vanilla Y to a depth and resolves it against each column's own surface and underside as the
  feature is placed. Floating islands, orbiting planets and sky continents get the same
  internal ore layout whatever height they sit at. RemapStrategy gained a
  remapToColumn(VerticalRange, RemapContext) default method returning
  Optional.empty(), so every existing strategy — built-in or third-party — keeps the
  absolute behaviour unchanged.
• isekai_api:vanilla_overworld_surface — a delegate SurfaceRules.RuleSource that
  reproduces the vanilla overworld surface rule, so overworld-replacement datapacks skip the
  30 KB surface-rule copy.
• Distribution as a Gradle dependency: sources and javadoc jars, and a maven repo.
  See docs/COMPATIBILITY.md.

Changed

• Dispatch prefix unified to isekai_api:. All "type" values normalise to the
  isekai_api namespace; docs and examples use it throughout.
• Inline apply_worldshape / apply_worldshape_structures modifier forms are superseded by
  the _ref forms.

Fixed

• isekai_api:worldshape_surface_top self-gates to the surface layer. Prepended bare into a
  surface_rule sequence it replaces only the topmost surface block; the blocks beneath keep the
  underlying surface rule's output. No minecraft:stone_depth wrapper is required, and wrapping it
  in one is harmless.

Deprecated

• The bare isekai: dispatch prefix — accepted with a one-time warning, removed in the
  next major. Rewrite "type": "isekai:*" to "type": "isekai_api:*".
• Inline modifier forms — use the _ref forms. Same one-major grace.