akgl_Partitioner is a record of function pointers and an initializer, the same
shape as the render and physics backends. `move` is its own slot rather than
remove-then-insert, and that is the whole design: a uniform grid's move is a
comparison and a return when the proxy has not left the cells it was in, which
is the steady state for a walking actor and the permanent state for a static
one. Spelling it as remove-and-insert would turn the incremental grid into the
rebuild-every-frame tree that PERFORMANCE.md already measured and rejected.
The grid is a dense 128x128 array of cell heads over the world, cells keyed on
tile size, with two intrusive chains per entry -- one through the cell so a query
can walk it, one through the proxy so removal unlinks a whole span without
searching. Everything is an int16_t index rather than a pointer, so the
structure is relocatable and clearing it is a memset. A proxy covering more
cells than AKGL_COLLISION_GRID_MAX_SPAN spills onto one chain every query walks,
which is what makes the cell pool's ceiling provable rather than hopeful.
tests/partition.c compares every query against a linear scan over the same
proxies and asserts containment in one direction, not equality. That asymmetry
is the contract: over-reporting costs a narrowphase call, under-reporting is a
wall an actor walks through. The suite is a table over partitioners so the same
assertions run against every implementation, which is what the second one
landing later will be held to.
Three deliberate breaks, and two of them were not caught the first time:
- Removing the min-cell rule so a shared-cell pair reports repeatedly: caught.
- Dropping the unlink in `move`: **not caught**, initially. Stale entries do not
produce wrong query answers, because the bounds re-check filters them out --
they leak the cell pool until the grid stops working, on a timescale no unit
test reaches by accident. Counting live pool entries across 200 moves is the
only thing that sees it, and now does: 505 entries against an expected 5.
- Swapping floorf for a truncating cast: **not caught, and correctly so.**
Everything below zero is clamped to cell 0 either way, so today the two are
indistinguishable. The comment claiming truncation is a defect was overstating
it and now says what is actually true: the clamp is the only thing hiding it,
and a negative world origin or a relaxed clamp would make it real with no test
standing in front of it.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A proxy is what the broad phase will index: an actor's shape, where it is, and
the bounds that follow from those. It is a sixth heap layer, two per actor --
one for the actor itself and headroom for static geometry a game registers that
is not tile-aligned. Solid *tiles* deliberately get none: the broad phase will
read those out of the tilemap's own array, because one proxy per tile is tens of
megabytes to index data that is already a grid.
The pool follows the existing convention rather than improving on it.
akgl_heap_next_collision_proxy finds a free slot and does not claim it;
akgl_collision_proxy_initialize claims it. That is TODO.md item 8's asymmetry,
and it is kept on purpose: an acquire abandoned before initialization leaks
nothing, because the slot still reads as free, and a sixth pool with its own
rule would be worse than one defect with five instances. Fixing item 8 means
fixing all five together with every call site audited, and that is its own
change.
The proxy holds a *copy* of the shape rather than a pointer. An actor's own
logic may rewrite its shape mid-step -- a character crouching, a projectile
arming -- and a broad phase whose bounds came from a shape that has since
changed is a bug that only appears when two things happen in the same frame.
akgl_heap_release_actor now releases the actor's proxy. Without it the proxy
holds a borrowed `owner` into a slot that has just been zeroed, so the broad
phase keeps a registration whose owner reads as a free actor, and the next
contact against it is a collision with nothing. Verified by removing the release
and watching test_proxy_dies_with_its_actor go red.
The tests pin the convention as well as the behaviour: that two acquires with no
initialize between them return the same slot is asserted, not merely tolerated,
so that changing the convention has to change a test that says why it exists.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A shape is a convex volume positioned relative to an actor's origin. It lives on
akgl_Character, so every goblin sharing a character shares one shape definition
the way they already share speeds and the state-to-sprite map, and an actor may
override it with akgl_Actor::shape_override.
The flag is not redundant with an empty shape. "This actor deliberately has no
collider" and "this actor has not been given one yet" are different states, and
without somewhere to record the difference akgl_actor_set_character cannot tell
whether it is allowed to overwrite what it finds.
include/akgl/collision.h is a leaf: it names actors and tilemaps as incomplete
struct pointers rather than including their headers, because both of those need
akgl_CollisionShape by value and a cycle forms otherwise. The `headers` suite
compiles it as the first include of a translation unit, so that property is
enforced rather than merely intended -- which is why the tilemap types got struct
tags two commits ago.
The masks are asymmetric and the defaults are the load-bearing part: layermask
ACTOR, collidemask STATIC. An actor given a shape and nothing else collides with
map geometry and with no other actor. "Everything with a hitbox shoves everything
else" would be a surprising default for a town full of scenery and a painful one
to discover after the fact; opting in to actor-versus-actor is one bit, opting
out of it would have been a hunt through every NPC a game spawns.
Every shape gets a z depth, because the narrowphase behind this is three
dimensional and answers with the *minimum* separating translation. A thin
extrusion makes z the cheapest axis, and an actor pushed along z goes nowhere a
player can see while remaining inside the floor -- a collision system reporting
success while doing nothing. The ratio of 2 is the smallest for which the z
overlap of any two shapes the setters build provably exceeds any planar
penetration they can reach.
The test for that asserts the inequality across a spread of sizes rather than
asserting the constant is 2, so a change to the constant fails here rather than
in a game.
Two things about that test are worth recording, because the first version had
both:
- It could not fail. TEST_ASSERT expands to FAIL_BREAK, which reports by
`break`ing, and the assertion was inside a nested `for` -- so the break left
the loop rather than the ATTEMPT block and the failure evaporated. This is the
hazard AGENTS.md documents for CATCH; it applies to the whole family. Verified
by setting the ratio to 0.5 and watching the suite still pass, then fixing it
and watching it report the 512x512 case by name.
- tests/collision_arena.c had the same bug in a loop added in the previous
commit, and it is fixed here too.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
