native/build.sh builds MoltenVK v1.4.2 from its pinned, checksummed tag (its dependencies at the
commits its ExternalRevisions pins), thinned to arm64, id @rpath/libMoltenVK.dylib, signed ad hoc;
its licence goes in native/LICENSE-MoltenVK. Every render3d program links it through its rpath -
the package's lib/ while developing, Contents/Frameworks in a bundle, where ludic bundle puts and
signs it - and vk_mac.ll also looks for @rpath/libMoltenVK.dylib (after an SDK loader, so the
validation layer still stacks in development). The suite's SDK stand-in (vk_env) is gone: the
render checks draw on the MoltenVK the package carries. gpu_is_gl() removed; nothing calls it.
ludic-dev test 306/306 with no Vulkan SDK in the environment.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- gvk_sampler built a key string on every call it was reached (a texture read two ways in turn
misses its per-texture cache each time); samplers are found by their seven numbers instead
- gvk_tex_mips_into allocated a VkImageBlit per level, every frame (the exposure measure's
chain); it comes from the scratch ring, as does gvk_tex_grow_mips's copy
- gvk_program_new made a new program - modules, pipelines - on every call, and every actor asks
for one; a variant is now made once and shared (a program is immutable), and one that cannot
be made stays 0 for every later asker
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
e634177 started a layer at 256 instances and grew it in layer_add and the stream gather, but a
game writes l.inst directly (Maroon Lake's track prints, trees and rocks), past what had been
grown: "index out of range: 2048, len 2048" in play. layer_new takes its cap up front as before;
layer_reserve(l, n) is exported for a caller that writes l.inst itself once layers start small
again, opt-in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
17.10 found almost every animal turn-back was for a climb, where the navmesh (40-45 degrees) is stricter than the step's 1.2 per metre. So the frame-sized test was refusing a few centimetres of steep ground: a pebble a walker steps across. It also refused sooner the faster the frame rate. wl_climb takes the ground half a metre ahead along the way the step actually goes. climb_test holds it at 60 and 240 Hz: a pebble is crossed and a wall is not. Both tests fail under the old rule. wildlife 24/24.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Found with malloc_history over 1200 frames of play without the interface (the first Ludic function
on each allocating stack, live bytes at two marks): sh_loc built `name + "[0]"` per program per pass
per frame - 24.6 MB of 33 in the window; its callers pass both names as literals now. tick_set fills
the frame loop's one Tick instead of tick_new making one a frame. A ridden boat or horse said
VEHICLE_MOVED as a new fact every frame; the metres are added up in VehiclesState and taken once
(vehicle_moved_take / _kind / _x / _z). Play's growth without the interface: 92 -> 14.7 MB a minute
(maroon-lake tests/leakcheck.sh); what is left is phys_push (Physics' fix on lang/nav) and the HUD's
strings. Packages 407, ludic-dev test 305 pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- model_release(model) frees each primitive's mesh, and each texture once no other model uses it:
the glTF texture cache counts the primitives using each texture (a path loaded again, and a LOD
chain borrowing its LOD0's material, each take a reference); the last one frees the texture and
forgets it along with any remembered material naming it, so a later load is a fresh one
- a scatter layer's instance and sort arrays start at 256 and double as layer_add or a stream
fills them, to the layer's cap, instead of the whole cap up front (0.4 GB in Maroon Lake)
- gvk_tex_read's copy struct comes from the scratch ring
- examples/rendering/release.ludic holds it (RELEASE OK on Vulkan and OpenGL), in the suite
smooth renders pixel-identical before and after (max |d| 0). ludic-dev test 306/306.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gvk_pipeline_fast cached only pipelines it made, so one the driver refused was attempted again on
every draw, each attempt building its key string and create structs anew; the refusal is cached
too (a Vulkan program's variants are fixed when it is made, so it would fail the same way).
stream_evict made a new chunk list per eviction and never freed the evicted chunks' records; it
compacts the list in place and frees each evicted chunk with its data.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
terrain_generate, terrain_use_dem and terrain_use_ortho each overwrote the previous height
texture (256 MB of RGBA32F at 4096^2), the heights read back (64 MB), the DEM texture and the
orthophoto's texture and pixels. terrain_reload unloads first, so the world swap was already
clean; any other caller building a map over a live one now lets the old go, and the read-back
array is reused, being the same size for every map.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
phys_pose, phys_ray, phys_push and phys_walker_pose each made a new record per call, some several times a frame, and Ludic never gives one back. They now fill the caller's PhysPose / PhysHit / PhysPush / PhysWalk and return whether they found anything. phys_overlap returns a count; phys_overlap_id(i) reads each id back. The package's own uses (phys_resolve, phys_row, phys_walker_move) read the values buffer directly. Tests take small allocating helpers. physics 20, character 15, vehicles 8.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- every texture upload malloc'd a CPU copy of its pixels and never freed it (236 MB over the
valley's boot, and again for every model loaded later): the pixels are converted straight
into the mapped staging buffer
- the per-level and per-layer views gvk_view_of made outlived their texture, and on MoltenVK a
view keeps its Metal texture alive: a released texture takes its views with it, and the cache
is keyed by numbers instead of a string built on every call
- the Vulkan structs filled for a draw, pass, barrier, descriptor set, buffer, allocation or
upload (about sixty call sites) come from a reused 1 MB scratch ring (gvk_tmp)
- the descriptor-set cache and the retired buffers are emptied in place, not replaced; the grass
cull's dispatch arguments are made once
- macOS drains an autorelease pool each frame (Vk.frame_pool, lvk_frame_pool in vk_mac.ll)
- R3D_VK_PROF reports Vulkan objects made and destroyed by kind, and every cache's length
- examples/rendering/smooth presents through render3d, so it runs on Vulkan too
The full valley on headless Vulkan loads to 2.18 GB and holds (it passed 8 GB while loading
before). ludic-dev test 305/305, selfhost-test 33/33.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rest of the leak Physics found: a drain that carried facts gave its list away and made a new
one, every queue every frame something happened. The queue keeps two lists and hands one out while
the other fills; a drained list is good until the next drain of that queue (nothing in the game or
the packages keeps one past it). With ab34f82's empty drain and in-place clear, a queue allocates
nothing in steady state. queue_test holds the reuse; every package (403) and lab/unit (90) pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every per-entity store (@S_ components, @H_ flags, alive, kind, freelist, owners) is a heap block
L_grow doubles from 1024 as L_alloc hands out a slot past it, the new slots zeroed; each site loads
the store's base where it indexes it (ecs_base, its registers %ecsb* so a raw function's t0 labels
cannot collide). Prop.has bounds against @L_cap, Pool.capacity answers it, a mod's registered
stores grow with the rest, every main grows the stores once before anything reads them. A snapshot
records its slot count first and a load grows to it before reading back. The overflow stop of
1c7ce84 is gone with the wall. ludic-dev test 305 passed, selfhost-test 33 passed; 1000 / 5000 /
100000 entities spawn and count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nav_us() gives the microseconds spent in nav_path across the loaded meshes. The shim times each path with the OS's monotonic clock (clock_gettime, or QueryPerformanceCounter from KERNEL32), so the DLL still needs no C++ runtime. Both libraries are rebuilt, and 13 tests pass on the Mac and the PC.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The count is wildlife_turned(), the 75-degree fallback an animal takes when the way was not enough. It is a measure only and is not saved.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A step round what is in the way, a target planned again after twenty seconds without progress, and a target given up are each pushed as NPC_F_STUCK (how, x, z), so the game can see where the way failed a person. The lake test expects one when a walker turns back from the water; the way test expects none when the world gives the way.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It sets a target straight away from the threat (40 m when it flees, 20 m when it is wary) and heads for it with wl_head, so the way port takes it round water and cliffs. Where the world gives no way, it still runs straight away as before. way_test now has a flee round the wall's end.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
