run() loads the shared resources and the pack, then hands them to a
constructor that builds a game with no world open. The same constructor will
build an integrated server's game on its own thread. The struct literal moves
unchanged; Options derives Clone so a server can start from the client's.
No behaviour change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Resources now holds only what is loaded once and never changes (blocks,
items, recipes, biomes, mobs, collision, motion) and is Send + Sync behind
an Arc, so an integrated server thread can share it with the client.
What a resource pack reload replaces moves to PackResources (assets,
meshing context, item renderer, names, panorama, clouds, forcefield), built
by one loader used at startup and on reload. The client-only mutable state
(the HUD atlas with its font cache and skin slots, fog state, screen effects)
becomes the game's own fields.
No behaviour change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The threaded test added seed + 1 for seed = i64::MAX, which overflows in debug
builds (CI's verify).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Loading a world read every file name under a folder pack's data/ to find
functions, function tags, predicates and item modifiers. The built-in pack
holds about 9,900 other data files, so this walk took 190-340 ms on the
render thread at every world load. A folder pack now walks
data/<namespace>/<kind>/ for those four kinds only, in the same order as
before; zip packs keep filtering their index.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/modpacks.md: how any pack (ATM10 down to small Fabric packs) gets to
Horizon: import (.mrpack, CurseForge zip or instance, Prism, packwiz),
inventory and classification, dependency order, mechanical carry-over of
data and assets, one agent per mod for code with a per-mod harness against
a dev-only NeoForge/Fabric oracle, and a coverage report. Also the mod API
gaps against 0.1, the registries 0.2 design (template blocks through the
overlay root), licence handling, how "10x" is measured against NeoForge,
and the scope for 2026-10-23.
docs/modpacks/atm10-mods.csv: ATM10's 495 jars from the pack's own
crash_assistant modlist (2026-10-15), with Modrinth metadata looked up by
hash: class, API tier, licence, sides. Metadata only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fills an area with N workers, optionally flies the centre along +x, and
reports wall time, chunks generated, sections meshed, how often a column
near the player is still unlit, and what Streaming::update costs the
calling thread. Allocations can be counted with
HORIZON_BENCH_COUNT_ALLOCATIONS=1 at build time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The macOS release build has failed since 05:54 UTC on 2026-10-16: on the
self-hosted Mac mini runner the game drew its 1500 frames in about 2.5 s,
before the world had started generating (chunks 0, player null), so no
release has been published since v2026.10.16.03.
--world-frames N exits after N frames drawn with the player in the world.
It is otherwise a frame-limited run (classic menus, --timeout-seconds), and
--frames keeps its meaning for the harnesses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 128 stronghold ring positions each need a biome search over a 57x57
quart area. The first worker to ask computed all of them inside a OnceLock
while every other worker blocked on it: about 2 s on an M4 Pro with 12 of
13 workers idle, after every world load and every return to the Overworld.
Vanilla's generateRingPositions runs each search as its own task on the
background executor. Now the ring chunks and search randoms are planned
serially (cheap), and every thread that needs the rings takes searches
until none are left, then waits for the rest. Positions are unchanged:
each search only uses its own forked random.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The mesher groups each section's solid quads by the axis side they face
(Sodium's ModelQuadFacing) and records each group's nearest plane. When
the eye is behind every plane of a group, back-face culling would discard
all of it after vertex shading, so the renderer leaves the group out of
the section's draws. Seen from the village path, 29-36% fewer solid quads
reach the GPU at render distances 8-32; the picture is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BlockCollisionScan asked every cursor cell for its collision shape,
moved it into place and intersected it with Shapes.create(box), even
for air. Mob movement and every other Entity.collide spent most of
their time there. Now:
- CollisionBlock carries empty_shape, set by the registry for states
with no contextual rule and an empty cached shape (air, plants): the
scan skips them without asking. Providers that build CollisionBlock
themselves (the query-order fixture) leave it false, so their query
logs are unchanged.
- A shape whose bounds miss the box is skipped before moving it:
joinIsNotEmpty(AND) of the two is false anyway.
- Shapes.create(box) is built the first time a shape is not the unit
cube (finite bounds always make one), not for every scan.
- SceneCollisions hands the scan the resident chunk itself, so a cell
costs one section read instead of a chunk-map lookup.
Results are unchanged: running both paths side by side over 900,000
mob moves in a 600-mob village gave bit-identical movement.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PoiTypes.forState ran per palette entry of every section a villager
searched (and per block of the sections that hold one), with a string
match on the block name each time and a per-search hash map in front.
At world load, when every villager looks for a job site, bed and
meeting point at once, that was 16 % of the tick. The POI type of
every state is now worked out once, the first time a villager looks,
and read from a table after that.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The probe builds MobWorld itself; it now keeps the same SeenGrid over
its snapshot for the mob ticks, so the mob_ai harness checks pushes
through the grid against vanilla.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Entity.checkSupportingBlock runs findSupportingBlock after every move of
a mob on the ground. Horizon looked up and built the collision shape of
every cell of the cursor (27 to 48 per call, twice when it falls back
to where the mob came from). Like BlockCollisions' cursor, a block whose
shape stays inside its own cell (neither hasLargeCollisionShape nor a
moving piston) can only meet the area when its cell does, so only those
cells, and large shapes such as fences, are shaped now. The pick is a
total order, so visiting column by column finds the same block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LivingEntity.pushEntities compared each mob's box with every other mob
of the tick's snapshot: quadratic, about 1 ms per tick with 600 mobs.
The snapshot is now also kept in a grid of 4-block columns by each box's
minimum corner, updated as each mob finishes its tick, and the push
reads only the columns its box can reach. Candidates come back in
snapshot order, so the pushes, their order and their sums are the same:
over 800,000 pushes in a 600-mob village both ways gave the same
targets in the same order. A query wider than the grid (or not finite)
falls back to the full scan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drop this branch's own staging buffer (gfx/uploads.rs); the sprite uploads
go through the StagingBelt from #501. What stays is the animation mip-chain
cache holding every vanilla frame, with tests.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Renderer::render's buffer writes (globals, section origins and draws,
entity, cloud, celestial, border, debug-line and screen-effect vertices)
and the HUD's vertices now go through the frame's staging belt instead of
one Queue::write_buffer staging buffer each. The lightmap is uploaded only
when its pixels change instead of every frame.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every frame that crossed a tick re-uploaded the animated block sprites, and
that took about 1.8 ms on an M4 Pro (more than the world tick itself):
- The mip-chain cache held 8 MiB, which fits 256 of the 594 vanilla
animation frames (64-pixel layers, chains stored with spare capacity), so
the other 338 frames had their resize and mip chain rebuilt on the render
thread whenever their animation reached them. The budget is now 32 MiB and
chains are allocated at their exact size: all 594 frames fit in 13 MB.
- Each changed sprite mip level went through its own Queue::write_texture,
which allocates and frees a GPU staging buffer per call (a few hundred a
tick). They are now packed into one reused buffer, written once, and
copied with one submission per tick.
The texels are unchanged: a GPU test compares the batched copies with
write_texture over several flushes, and the existing cache tests compare
cached and generated chains.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Queue::write_texture creates a new staging buffer for every call, a kernel
call each on Metal. Animated sprites write every changed frame's mip
levels once per tick, so tick frames spent about 1.3 ms in buffer creation.
The copies now go through one wgpu StagingBelt recorded into the frame's
encoder; a GPU readback test checks they write the same texels as
write_texture on every mip level.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nobody could reproduce it, and no issue tracks it, yet every release since
2026-10-14 repeated it. Players noticed (rdbt on Discord).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>