`locate` found an entry's word by multiplication (the MAGIC table) but
still divided `64 / bits` at run time for the shift, on every block read
and write of a section stored in 5, 6 or 7 bits. A division by a value the
compiler does not know costs tens of cycles on x86: `perf annotate` of a
10-bot server on the ThinkPad put 12% of `PalettedContainer::get`, the
server thread's top function, on that `div`. The table now also holds the
bits a word's entries fill, so the shift is `i * bits - word * filled`.
A test checks `locate` against the division for every width and entry.
ThinkPad (i5-7200U), 4M random reads, 5-7 bit palettes: 28.5-36.9 ->
24.6-32.4 ns over 2,000 sections, 6.7-7.1 -> 5.4-5.8 ns within one;
4 and 8 bits (shifts, unchanged) the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The batched loop and the one-at-a-time reference were two copies of the
same 100 lines; a const generic picks the path, so the test compares the
code that ships with itself minus the batching. Drops the ignored
microkernel timer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Evaluate four same-cell Y samples together while preserving coordinate
setup, gradient caching and each sample's floating-point operation order.
Safe arrays let LLVM vectorize the interpolation on the portable target;
cell boundaries and tails keep scalar evaluation.
Add bitwise differential coverage against the original volume loop.
Two paired chunk-generation batches reproduce a 4.6–4.7% Nether wall-time
reduction, with smaller or unresolved effects in the other dimensions.
Co-Authored-By: Codex gpt-6.1-sol <noreply@openai.com>
The file is generated from the pinned version lock, and test_gen_version_consts
checks JAVA_VERSION against it; deleting constants by hand breaks the next
regeneration and that test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
93 pub items (78 functions and methods, 14 constants, 1 struct) in horizon-client,
horizon-world, horizon-core, horizon-friends and horizon-server-files had
no caller anywhere in the repository: not in the crate, the other crates,
the examples, the tests or the binaries. rustc can't flag them because
horizon-client and the others are libraries, where a pub item counts as
used. The imports and the private helpers only they used go with them
(title.rs date_string, goals.rs flag_index).
Also drops the #[allow(dead_code)]/#[allow(unused)] items that only hid
dead code: chat_debug_fill, merchant_title and Goals::locked_by had no
caller, and four stubs (_keep, _unused, _rule_kind, const _) existed only
to keep an import used. LightTables::from_parts is used by its tests, so
it is #[cfg(test)] instead of allowed.
Protocol id tables and entity data serializer ids stay: they mirror
vanilla's registries in full. The launcher's three unused items are left
for a launcher PR.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
horizon_core::no_ai reads and writes the noAi line of the game's
horizon-extras.txt, so the launcher and the game share one switch. While
it is on the game never opens its control socket, horizon-ctl refuses every
command with exit 4 and kind denied, horizon-ctl mcp and horizon-launcher
mcp refuse to start (and refuse calls if it is turned on mid-session), and
horizon-launcher no-ai [on|off] shows or sets it. The game's extras keep
the key when they rewrite the file. Both e2e harnesses check it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the integration run on the desktop (main 34c3c1cf + the 14 contributor PRs: workspace tests 1,869/0, launcher 113, parity --changed 320/0).
🤖 Agent: merged after verify passed and the integration run on the desktop (main 34c3c1cf + the 14 contributor PRs: workspace tests 1,869/0, launcher 113, parity --changed 320/0).
🤖 Agent: merged after verify passed and the integration run on the desktop (main 34c3c1cf + the 14 contributor PRs: workspace tests 1,869/0, launcher 113, parity --changed 320/0).
Vanilla 26.3 Mth.floor(double) is (int) Math.floor(v), so a value below
i32::MIN floors to i32::MIN. The truncate-and-step version wrapped to
i32::MAX there (block_collision_scan parity failed at the chunk of
x = -2147483648.8: Rust 134217727, vanilla -134217728). Step down with
saturating_sub; the fast path is unchanged, and the test now checks
the ends of the i32 range, infinities and f64::MIN/MAX against
f64::floor as i32.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tick_chunk_random asks for the lowest randomly ticking section from a height
up, instead of asking about every section: a chunk costs one question per
section that ticks, not 24. PalettedContainer::has tests whole words for the
few matching palette indices (a field-wise zero test) instead of reading
every entry, so a section read again after an edit costs a few hundred word
operations.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Vanilla random-ticks a section while its count of randomly ticking blocks
and fluids is above zero (LevelChunkSection.isRandomlyTicking), and asks as
it reaches each section. Horizon ticked every section whose palette listed
such a state, even after the last one was gone, and decided before the tick's
scheduled ticks, which drew positions vanilla doesn't and shifted its
position sequence.
RandomTickSections now reads each section's blocks (PalettedContainer::has),
again only for the sections that changed, and tick_chunk_random asks for
each section when it reaches it; the game reads a chunk's sections again only
after a block changed. New harness random_tick_sections compares the
sections and positions with vanilla's real section counts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
floor_i32 is now Java's Mth.floor (truncate, step down for negative
fractions): f64::floor is a libm call on baseline x86-64, made for every
Perlin sample. Perlin's gradient table is stored as floats (every
component is -1, 0 or 1, exact either way), so the corner dot products
skip three conversions each. About 10% fewer instructions generating
terrain; values unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M3pNrSHA2ifNrH7SAMHgSW
PerlinNoise::add_volume worked out each Y coordinate's cell, offset and
fade once per X/Z column. Every column walks the same Y coordinates, so
they are now computed once per volume. The arithmetic is unchanged, so
the output is bit for bit the same.
ClimateIndex kept its tree as nodes with a Vec of child ids each. The
built tree is now renumbered breadth first into flat arrays, so a
branch's children are one contiguous run. The search visits the same
nodes in the same order, so ties resolve exactly as before.
terrain_bench (seed 12345, 8x8 chunks, 1 thread, medians of 3 runs):
terrain 14.5 -> 12.4 ms/chunk, biomes 1.20 -> 1.06 ms/chunk.
Checked against vanilla 26.3 in-process: 2304 chunks on seeds 12345, -1
and 987654321987 identical, and the biome-source and biome-generation
parity suites (2.1M lines) match.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M3pNrSHA2ifNrH7SAMHgSW
release.yml builds every game release with HORIZON_VERSION, the tag without its v
(docs/versioning.md), as launcher.yml already does for the launcher: no version-bump commits.
- horizon_core::release::{RELEASE, VERSION}: the built-in version, or Cargo.toml's for a build from
source. (horizon_core::version stays Minecraft's, generated by gen_version_consts.py.)
- horizon-server and horizon-prepare --version, the server's /api/v1/status, horizon-ctl's MCP
serverInfo and the "Horizon/<version>" user agents use it; they all said 0.1.0.
- The client's crash reports and telemetry say v0.2.3 from the built-in version wherever the game
is installed; releases from before 0.2 (no version built in) still read the launcher's folder name.
- The launcher's benchmark telemetry used Cargo.toml's version instead of the release's.
- release.yml checks that horizon-prepare and horizon-server --version say the release's version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/gen_version_consts.py writes what the pinned client jar's version.json
states (name, id, protocol, data version, pack formats, Java, stable, build
time) into horizon_core::version, and every Rust site reads it from there:
PROTOCOL_VERSION, GAME_VERSION, DATA_VERSION, the save's and the datafixer's
version name, /version, the crash report, the legacy ping, the dedicated
server's start line, the title screens, the default resource folder, the
resource pack format, and the launcher's and horizon-ctl's fallbacks. CI
checks the file against the pinned jar.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify and CI parity passed, and a combined check with rdbt's #796 on current main (fmt, clippy, 359 tests, 14 world-gen harnesses). YUNG/Moog terrain adaptation around structures (Better Fortresses netherrack inside bridges 39–41 % → 2.5–7 %), its carry, and YUNG modifiers.
🤖 Agent: merged, thanks Online5277! Terrain LOD lands off by default. With LODs off it matches main (0 differing pixels in 6 RD16 views, same FPS/GPU ms on the HD 620, same time to full render distance); CI verify and parity pass. We merged main in three times and adapted it to #905/#939 and the packed texture coordinates. The follow-ups from the review (edge band, integrated-GPU cost, lighter far meshes) we're taking on ourselves.
YUNG's API and MoogsStructureLib add their own kernel to the Beardifier:
a size and falloff of its own, carving or burying above and below a piece's
floor, around rigid pieces and junctions. Better Strongholds buries its
stronghold, Better Fortresses carves the netherrack over its bridges and
Moog's mega fortress carves a band around its corridors. A structure or a
pool element holding `horizon:enhanced_terrain_adaptation` now gets that
kernel (horizon_core's TerrainKernel), summed after vanilla's density as
the libraries' mixins do; YUNG's also grows the start's box by its radius.
Vanilla data never has the key, so vanilla worlds are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keeps #905's single command buffer and #939's occlusion culling: distant
tiles draw in phase 1, so phase 2 stays exactly the deferred sections the
culling test reads back, and the flat path's origins include the tiles in
main's relocated upload block.
Adapts to main's APIs: block sprites and models now live in RenderAssets
(WorkerContext::render), BlockFacts behind Resources::facts(), and
terrain vertices hold the UV as 16-bit fixed point (#df58a474). Merged
distant faces repeat their sprite once per block, which the [0, 2) section
UVs cannot hold, so tile vertices now store UVs from the face's corner in
1/32-block steps (TerrainVertex::tiled, faces split at 1024 blocks) and the
instance's origin.w carries the UV step: 1/32768 for sections (decodes
exactly as before), 1/32 for tiles.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: Merged. Bit-exact against the per-corner oracle (unit test over a dense grid, 7 seeds, wrapping extremes), CI parity passed, and the surface pass is about 10% faster in chunkgen_bench. Thanks rdbt!
palette_ids (behind PalettedContainer::pack, every saved and sent section,
and from_values, every generated one) scanned the palette for each of the
4096 values while it had up to 16 entries, then hashed every value with
SipHash. A run of one value is now matched once, and past 16 entries a
thread-local table indexed by value (block states and biomes are below
65536) replaces the hash map; only the palette's entries are set, and they
are cleared before returning. The output is unchanged: palette in
first-occurrence order and the same ids.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BlockSection::get and set check the section index and the state id once,
then read and write through new PackedStorage::read/write, which skip the
storage-level bounds and value checks and the Result that get/set keep for
external callers. A section index is below 4,096 and a palette id always
fits the width bits_for_palette chose; both methods debug_assert that
contract. Used for the hot read, the hot write and the growth re-encode.
Benchmark (ignored tests, best of alternating-order rounds, fast profile):
BlockSection::set 13.15 -> 9.12 ns/write (1.44x), get 2.33 -> 1.45 ns/read
(1.61x). See docs/performance/paletted-container-access-results.md.
BlockSection::set resolved the old cell through get and then probed the
section palette three times per non-trivial write: keying the old state by
id, testing the incoming state and finding it. The slot at the written index
already is the old state's id, so read it straight from storage and resolve
only the incoming state, with a single map probe. Palette reuse, free slots
and growth are unchanged, so palette, counts, free list and NBT are identical.
PalettedContainer::unpack and unpack_u16 matched the palette variant inside
the per-cell loop; dispatch Single/List/Global once before it.
Measured against copies of the pre-optimization code (ignored tests):
BlockSection::set 12.73 -> 11.37 ns/write (1.12x), unpack list 1.956 ->
1.860 us/call (1.05x), global 2.255 -> 2.172 (1.04x). See
docs/performance/paletted-container-access-results.md.
set_matches_baseline_under_random_edits drives both implementations over the
same 50k-edit streams for 1/2/16/17/24/64 distinct states, with periodic
whole-section collapses, and asserts equal results, bits, palette, counts,
free list and NBT.
getFiddledDistance restarts a six-step LinearCongruentialGenerator mix for
each of the eight quart corners. Its first two steps depend only on the X
and Y corner offsets, so quart_at_block now mixes the two X states and the
four XY states once, then runs the remaining steps across the eight corners.
next(m, salt + offset) == next(m, salt) + offset, so a coordinate folds into
its step with one addition.
Bit-identical to BiomeManager.getBiome and getFiddledDistance; a reference
oracle test pins the equivalence over a dense grid and wrapping extremes.
Block collision scans (vanilla's BlockCollisions) are most of every mob's
move, and moves were the largest share of the world tick with 600 mobs.
Per cell and per hit, the scan did work vanilla's JIT makes cheap:
- the cursor divided its index by the box width and height for every cell;
it now counts x, then y, then z, as Cursor3D does (same cells, same order,
same budget);
- the unit cube test cloned the shared unit cube to compare identities
(`VoxelShape::is_block` compares with it in place), and built a
normalised box to test it against the entity box (`intersects_corners`);
- every non-empty block looked its state up in the contextual rule map
(scaffolding, liquids, powder snow); a flag in the scan table now says
which states have one;
- a moved shape decomposed its occupancy into boxes again; the moved
shape's boxes are its own boxes moved, the same values.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PalettedContainer::pack hashed all 4096 values with SipHash to build the
palette. It now shares from_values' scan, which only switches to a map
past 16 distinct values. Same palette order, same output.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The trading screen runs MerchantMenu on the menu engine with the villager's
merchant::Session: level-ups add offers while it is open, prices follow the
TRADE reputation event a tick later, orbs drop at the villager with its own
random, the yes/no voice and regeneration as vanilla; special prices stay
after closing. Offers use the world's RandomSequences (now saved as
random_sequences.dat), the villager's position, its namespaced type (desert,
snow... trades) and exploration maps found through /locate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Beardifier, density pow/log, Marsaglia's Gaussian, dripstone, ores, fancy
trunks, icebergs, end spikes, strongholds, note pitches, sculk charge
particles, the warden's and brains' weighted shuffles.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rust's float functions use the platform libm, and vanilla's own Math
results differ between x86-64 (Intel stubs, close to glibc) and AArch64
(fdlibm, with sin/cos evaluated using fused multiply-adds). jmath picks
the same implementation HotSpot uses on the machine, so options, the
world border's alpha and RandomPos match vanilla on an Apple silicon Mac
as well as on Linux.
The vanilla protocol client no longer records the game-time sync sent
every 20 ticks: whether one arrives before the first chunk batch depends
on when the client joined.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep the integration branch's StringSplitter port and HUD overlays: the tab
list now emits hud_layout::GuiDraw calls and wraps with its split_lines, and
the vignette's border warning comes from WorldBorder::warning_strength
through hud_overlays::border_warning.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
monster_brain runs BreezeAi and CreakingAi as vanilla's Brain does:
the sensors with their random first delays, memories that expire, the
core, idle and fight activities, and each behaviour (Shoot, LongJump,
ShootWhenStuck, Slide, SlideToTargetSink, StartAttacking, MeleeAttack,
RunOne...) started and ticked in vanilla's order, with durations drawn
from the level random. It reuses the villager brain's memory slots.
Breeze.tick draws its particle count and whirl clock before the entity
tick; its whirl and idle sounds are local (client) sounds; it is immune
to fall damage and plays its landing sound. Entity.lookAt and the long
jump's atan2 (an fdlibm port) are bit-exact. A frozen creaking keeps
thinking while its navigation and controls wait, stopInPlace zeroes its
motion, and its attack sound follows the hit. Mobs update their ground
flag on every move, as a server's authoritative entities do.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The client border now wraps horizon-core's vanilla-checked BorderGeometry
(previous size for partial ticks, status, lerp speed, absolute limit) and
border_render ports WorldBorderRenderer: extract's fade and tint, the cached
wall buffer, the scrolling texture transform and which walls draw. The
vignette takes vanilla's int colour and the warning strength from the
border. test_world_border_render_parity runs vanilla's renderer against
examples/world_border_render_fixture: 1,500 cases, 16,859 frames match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Zombies (and husks, drowned, zombie villagers, zombified piglins) now
register ZombieAttackTurtleEggGoal first, as vanilla does, so their goal
selector draws the same random numbers. LivingEntity.pushEntities runs in
aiStep against the player and other mobs, with pushes to other mobs applied
at once as Entity.push does (replacing the post-tick push_apart). Spider
target goals check the light before their random interval, LookAtPlayerGoal
uses the non-combat targeting conditions, raiders register the
PatrollingMonster/Raider goals before their own, and the witch runs its
aiStep (heal cooldown, drinking, particle roll) before super.aiStep with
NearestHealableRaiderTargetGoal's nextBoolean.
The mob AI oracle's player is now a fully loaded, hurtable client: its
connection starts with a 60-tick load timer and 26.3 moved the hurt cooldown
to damageCooldownTime, both of which only ServerPlayer.tick counts down, so
whether mobs could hurt the player depended on the scenario order. The probe
reports the player's hurt sound, and an rng-trace feature records the call
sites of Horizon's mob random draws next to the oracle's trace.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Formats the overnight parity work, fixes the lints it introduced, and
updates the LAN bot test and mob AI probe for fields added on the other
side of the merges (the bot test also set a field that never existed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>