A POST-PASS, not a sweep change. The sweep carries the 0.55 step-up,
CONTACT_SKIN and the terrain-cliff logic, and every existing contract was
written against it, so it is untouched. separate_movers runs after the whole
integration loop — which also makes the result independent of who stepped
first — and before rider pinning, which stays authoritative.
HORIZONTAL ONLY, resolving the least-penetration axis of x/z. Resolving
vertically is exactly how characters end up standing on each other's heads; an
overlapping pair is pushed apart on the ground plane and a stack unpicks
itself.
Three FIXED relaxation passes, deliberately not convergence-based: an
early-exit on "nothing moved" makes the result depend on iteration order, and
this has to be bit-reproducible. Broad phase is a uniform grid sized 2x the
widest half, with buckets as a sorted (cell_key, index) array rather than a
hash map — allocation-light AND ordered without a second sort. That replaced a
hash map of per-cell Vecs and took allocations from 617/tick to ~15.
Each shove is clamped by sweep_axis against the solid world. Without that, a
crowd pressed against a wall squeezes its outermost members straight through.
push_mass weights the split by the OTHER body's mass, so equals each give half
and a player at 4.0 shoulders through NPCs at 1.0. 0.0 — the Default — READS
AS 1.0, not as weightless: a literal zero would make every default-constructed
mover infinitely shovable and divide by zero when two met. Same discipline as
`hidden` over `visible`.
Projectiles are excluded, and that is CORRECTNESS not taste: collect_touches
reports a strike from the overlap itself, so separating projectiles would mean
a bullet could never touch anyone. Sensors, collide:false decor and attached
riders are skipped too.
50 packed movers 0.023 ms/tick
200 packed movers 0.123
12 villagers + 500 static 0.107
200 movers + 500 static 2.020
Packed crowds where everyone overlaps a neighbour — the honest worst case. The
200-among-500-statics figure is dominated by the per-shove static clamp; at the
realistic 12-50 NPCs it is 0.1-0.25 ms. The fix if 200+ becomes normal is
accumulating pushes and clamping once per mover per pass, deliberately not done
because it changes Gauss-Seidel to Jacobi and the numbers don't justify it.
THE GOLDEN HASHES DID NOT CHANGE, and that is genuine rather than lucky:
mover_scene's walkers start 1.7 apart with 0.4 halves and diverge, and its only
other mover is an attached rider, so no pair ever overlaps and the pass is
inert. Nine new tests prove separation works; the unchanged goldens prove it
does nothing where movers never meet.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Eight Kenney civilians (character-female-a..f, character-male-a..b) plus the
KayKit knight as a standout — nine kinds across eleven villagers. Picked
through AssetIndex::casts() by taking the rig with the MOST members rather
than by naming a joint count, so a library that later grows a better-populated
rig gets used without editing this. Townsfolk are civilians because a village
wants people, not nine fantasy heroes; the knight stays because he is the
figure the player already knows, and having him on the other rig is what makes
the multi-rig path real rather than theoretical.
CLIPS RESOLVED BY NAME, PER MODEL. The rigs name locomotion differently —
Kenney's 7-joint civilians use idle/walk/sprint, KayKit's 41-joint heroes use
Idle/Walking_A/Running_A. clip_index is case-insensitive, so one ordered
fallback list covers both. Borrowing an index across rigs would have animated
a spellcast or a death pose.
TEXTURE BINDING was the real bug this exposed: SkinnedBatch carried ONE texture
for all items, which silently renders one character in another's atlas. It now
carries a texture palette with a per-item index, clamped rather than indexed
blindly so a bad slot cannot panic mid-frame. (KayKit embeds its atlas and
ships a sidecar; Kenney characters reference a pack-shared colormap — both
arrive as bytes, so the distinction disappears at load, but the BINDING had to
become per-item.)
Cost went DOWN: 17,440 verts skinned per frame, 408 KB/frame, against 958 KB
for the eleven-knight village, because a civilian is ~1,300 verts to the
knight's 3,716. The shape of the cost is unchanged and still doesn't scale —
GPU skinning remains the right fix.
Found by looking, not by testing: height normalisation was INVERTED, and the
first capture showed villagers about half the height of their own front doors
(Kenney's "mini" characters are ~1 unit against the knight's ~1.8). That
normalisation is keyed off joint count, which is crude — a third rig would want
measured rest-pose bounds, and SkinnedModel exposes none today.
63 props, 58 colliders, 11 NPCs of 9 kinds, 19 draw items, 12,244 triangles,
88 shadow casters.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE FINDING THAT CHANGES THE PREMISE: "1 rigged model in 4,442" was measured
with the BROKEN GLB probe (the gLUF magic bug). With the probe fixed the
library holds 36 rigged models across three rigs:
41 joints — 9 KayKit heroes + undead, up to 95 clips
7 joints — 22 KENNEY civilians (male/female a-f, orc, human, archer, shop
employees, skaters, soldiers), 32 clips
6 joints — 5 Kenney platformer characters, 25 clips
So a village can be populated with 22 visually distinct civilians TODAY, with
no third-party pack at all. Every conclusion drawn from that probe before it
was fixed needs re-checking, not just this one.
KayKit: 9 characters fetched (Adventurers + Skeletons), pinned by commit +
sha256, 37 MB, gitignored. CC0 verified by READING LICENSE.txt at each pinned
commit, recorded in the script header and CREDITS.toml.
THE SHARED RIG HOLDS ACROSS PACKS, proven rather than assumed: hashing the
joint-name list of all nine files yields the SAME digest — 41 joints, same
names, same order — despite two separate repositories. Skeleton clips are a
strict superset (95 = the adventurers' 76 + 19 undead extras: awaken,
resurrect, spawn, taunt). So a clip authored for the knight plays on the
skeleton warrior and one animation path drives the cast. A test pins this,
including that both packs are present, so a version bump cannot silently break
it.
The texture trap that cost the Kenney fetch three attempts does NOT apply:
KayKit GLBs EMBED their atlas (image/png in a bufferView), verified by parsing
all nine.
tests/rigged.rs parses all 36 rigged models through makepad_game_render::skin
— the loader the app actually runs — and asserts the index's joint and clip
counts match it. Deliberate: the index's own probe was wrong for the entire
library once and survived because the fixture encoded the same error.
find_cast groups by JOINT COUNT rather than pack, because the valuable fact is
cross-pack interchangeability. Cast states are the INTERSECTION, not the union
— advertising a state one member cannot perform is worse than a shorter list.
Added the state words the skeletons needed (spawn/resurrect/taunt/use):
Skeletons_Awaken_Floor previously matched nothing, so "an undead that rises
from the ground" was unfindable.
One bug found in its own work: casts_to_json emitted a doubled closing brace —
malformed JSON that still looked fine in a log. Fixed with a structural test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Eleven villagers as Movers driven by the Npc block — no special-casing, so they
collide with houses, benches and the fence exactly as the player does. Between
tick 120 and 300 they redistribute along the street, TWO PAIR UP AND TRAVEL
TOGETHER (the Follow behaviour), one settles by a bench, and an east-bench
cluster disperses. The rigid crate pyramid topples in the same window, so the
physics demo is still live underneath.
Knight split into a shared rig + per-villager pose: model, atlas and clip
indices load once; Villager holds only pose buffers, walk phase, tint and
build. follow() reads velocity back off the entity AFTER the sweep, so facing
comes from actual travel and the walk cycle advances with distance covered —
a villager stopped against a bench stops its legs instead of moonwalking. The
old hardcoded triangle-wave patrol is gone.
Per-villager tint (one vec4, one multiply in the vertex stage), because one rig
serves the whole village and without it every passer-by is the same knight in
the same colours — the identical-clones failure the prop variety work had just
fixed.
Scene faults fixed: the fence ran along z=17 while the yard sits at z 14..26,
crossing the green and enclosing nothing — now two legs meeting at a corner.
The yard is dressed with stock crates and barrels so it reads as a working yard
rather than a physics harness. The stray teal/orange lozenge is fixed AT THE
SOURCE: "rock stone" used Spread::Mixed, which round-robins across families,
and the neighbouring family is cliff_blockCave_rock — a cave-mouth tile that
reads as a small teal-roofed building. Variants keeps it inside
nature-kit/rock_largeA..F. Camera pulled 56 -> 44 units; a third of the frame
was bare lawn.
63 props, 58 colliders, 11 NPCs, 19 draw items, 12,244 triangles, 88 shadow
casters.
KNOWN COST, left documented at the call site rather than buried: CPU skinning
is ~41k verts and ~958 KB uploaded EVERY FRAME for eleven villagers. Fine here,
wrong for a town or a Quest. The bone palette is already computed, so the GPU
swap is this one loop plus a shader.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE WALK-THROUGH BUG WAS NOT THE COLLIDERS. Knight::tick wrote a triangle wave
straight into self.pos, and find(|e| e.tag == "knight") returned None — no
mover, no sweep, no Character block. He passed through benches, houses and
trees alike because there was nothing to collide WITH. The collider maths was
right all along, which is exactly why the house-wall and tree-trunk tests
passed while the user kept reporting walk-through three times over.
Dumping the bench's real collider before changing anything confirmed it:
graveyard-kit/bench yields 1.36 x 0.9 x 0.78 at ground level — a perfectly good
obstacle that nothing was ever tested against.
The Knight is now a BodyKind::Mover (hidden, so the mesh stays his appearance).
tick sets a HEADING; desired_velocity feeds entity.vel; his rendered position
is read back AFTER the sweep. Intent goes in, physics decides where he ends up.
Two things that fell out: his half-extents would have been 1.4 m wide and 3.6 m
tall, because spawn takes FULL size; and his patrol line ran straight through
the bench row, which — now that he genuinely collides — would jam him against
the first bench forever, so he walks the pavement between road edge and
furniture. Walking AROUND obstacles is NPC behaviour, not layout.
prop_collision.rs loads the real bench GLB, reproduces compose_village's
scaling, and walks a Knight-sized mover into it. spawn() now routes through
push_entity rather than entities.push, so the sorted-id invariant is asserted
rather than assumed.
VARIETY WIRED: 5 house designs instead of one model five times, 4 distinct
pines, a real lamp post instead of a CACTUS, two real benches instead of a
coaster-train carriage and a park entrance. Two genuine bugs in the variety
layer, both making find()'s correct answer worse:
- dominant_pack SUMMED 60 hits, so mass beat quality: nature-kit's incidental
"tall" matches out-summed racing-kit's three lightPosts, and a lamp query
returned a cactus. Only hits within 25% of the top score count now
- Spread::Mixed wanders on multi-word queries — "park bench wooden" let bench,
coaster-train-wooden and park-entrance each pass on one word
RANKING: whole_query_bonus tested only the ENTIRE query, so "fence" scored the
real fence 28 while "wooden fence" scored it 8 — tied with everything and
decided alphabetically, which is how asking for a fence returned arena/wall.
STATIC SHADOWS 15 -> 69 CASTERS: base_y was the MAXIMUM static top, and the
per-prop colliders are static entities, so the receiver plane sat at roof
height and every prop projected onto a plane above itself. Each prop now uses
its own lowest point — also correct on a slope.
Fog 0.004 -> 0.0015 (24% -> ~10% wash at the treeline), set on the demo rather
than SkyConfig::default() which gamemaker also reads. Fence spacing derived
from the panel's own scaled width. Crate stack is a 3-2-1 pyramid, not a
six-high chimney. Aspect guard so a short wide model can't explode sideways
into a coloured slab — any library picked by description eventually returns
something oddly proportioned.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
blocks/npc.rs + sim/sense.rs. Utility scoring re-run every ~0.45s per NPC,
staggered so a crowd doesn't re-plan on one tick. Candidates scored additively:
visit a POI (tag appeal x distance falloff x novelty x jitter), loiter near
someone (gated on sociability), go home (grows with time away), wander (the
fallback, so a world with no POIs still moves).
Three things do the legibility work. POIs, so NPCs walk to THINGS rather than
coordinates. Per-NPC seeded personality (haste/patience/sociability/curiosity/
homebody) so identical config still yields unlike villagers. And a day clock
with a per-NPC phase offset — benches read as afternoon, doors as evening —
which is what stops ten villagers doing the same thing in unison. Activities
are deliberately only four (Idle/Travel/Dwell/Follow); routines come from
sequencing them, not from twenty verbs.
Sensing reads THE SAME SOLID FILTER the mover sweep uses, so perception and
collision cannot disagree. obstacle_ahead sweeps the NPC's own box rather than
casting a ray, because a ray through a doorway reports "clear" for a body twice
its width. Blocked -> jump if the top is in reach with landing room, else
sidestep toward the side with clearance (blended with the goal so it curves
rather than turning 90 degrees), else a stuck timer abandons the goal.
Reading the existing tests caught a bug in the new logic: "low obstacle -> walk
over it" is wrong, because the 0.55 step-up is a TERRAIN contract and
sweep_axis blocks against static boxes at any height. That branch is gone —
it was exactly the perception/physics disagreement this module exists to avoid.
Two bugs the tests caught:
- MUTUAL SOCIAL LOCK: two sociable NPCs each chose to loiter near the other,
permanently. One moved exactly 0.0 units in 90 seconds. A social cooldown
stops Follow being re-picked immediately
- VILLAGE DRIFT: an unbiased random walk has no centre, and a trace showed a
villager 43 units out with every POI inside 18. Wander steps past a 26-unit
leash now aim home
Cost against a 16.6 ms budget: 50 NPCs 0.007 ms/tick, 200 NPCs 0.046 ms/tick
(full sim step — a pre_step-only figure would be a lie, since without
step_world the NPCs never move and re-decide more often).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE RANKING WAS NEVER WRONG. "suburban house building" already returned 21
distinct houses at equal score and "pine tree" six distinct pines. The API had
no way to say "give me five DIFFERENT ones", so callers took hit #1 and placed
it five times — with 4,753 models installed, a scene used about six.
find_many(query, VarietyParams{count, spread, seed, filters}) never returns the
same model twice. Spread::Mixed spreads across variant families before
repeating a shape, which handles both real cases with one rule: five houses
come back as building-type-p/r/s/t/u, eight trees as pine/oak/palm/fat/cone/
detailed. palette(query, seed) returns a matched set from ONE pack.
Three things only visible by looking at output, not by reasoning:
- VARIETY MUST STAY ON-TOPIC. Round-robin across families returned one house
then two driveways and two fences (city-kit-suburban themes all of them
"house"). A relevance band was the obvious fix and was WRONG: an exact
one-word hit ("tree") outscores a compound sibling ("tree_blocks") merely for
being shorter, so banding cut real variety while keeping the drift. What
separates them is whether the family NAMES the thing asked for — applied only
when it leaves something, since functional queries name no shared noun
- VARIETY MUST NOT BECOME INCOHERENCE. Maximal spread gave five houses from
five packs — the junk-drawer failure reached from the opposite direction. The
dominant pack is exhausted before crossing; a test asserts a street uses
exactly one pack
- RE-SKINS AREN'T KINDS. tree_blocks/_dark/_fall is one tree in three palettes;
counting them as three kinds returned the same silhouette six times. Colour
and season tokens are stripped from the family key
Palette grouping needed a coarser key of its own: family_of produced 167 groups
of one id each — a listing, not a palette. Bucketing on tile role or first
meaningful token gives 23 usable groups.
Selection is seeded, so multiplayer replicates a scene as (query, seed) and a
re-run looks identical.
Also fixed: "boulder" returned tower-defense-kit/weapon-ammo-boulder — catapult
ammunition — because that filename says the word while landscape rocks reached
it only via a synonym. A confidently wrong top hit matters more than a miss
here, because a composer places it several times.
The perf test now takes MIN-of-N instead of an average: it shares a machine
with 23 other tests, and the same query measured 2.1 ms alone and 44 ms under
the full parallel suite — a 20x swing with no code change. The fastest run is
the one that actually got the CPU. Same protocol the box3d benchmarks use.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three complaints from looking at the running app, all one problem: the world
didn't behave like a place.
COLLIDERS COME FROM EACH PROP'S OWN PRIMITIVES, not its AABB. Kenney authors a
house as walls + roof + door frame and a tree as trunk + canopy, so
StaticModel::parts records per-primitive bounds during the existing vertex
bake (they were being merged away). collider_parts() drops boxes under 10% of
the model's span, merges near-coincident ones, caps at 8 — low-res by design.
Policy falls out of the decomposition: buildings/fences/rocks take every
qualifying box; trees keep only parts both narrow and low, so the trunk blocks
and you walk under branches; lamps and decals take none. A prop whose parts all
filter out gets a synthesised box (trees a narrow post), because silently
reverting to walk-through scenery is the bug being fixed — a real catch, since
that fallback first shipped for Solid only and colliders dropped 39 -> 20 when
single-mesh pines found no trunk.
`hidden` rather than `visible`, deliberately: Entity derives Default, so the
field defaulting to false must be the UNUSUAL case. A `visible` flag would make
every default-constructed entity invisible — the same trap as the zero-seed rng
and the zero-gravity bodies this codebase has already been bitten by twice.
Proven by test, not by eye: a walker stops at a house wall but passes through
its DOORWAY (this fails with a single AABB), a trunk blocks while its canopy
doesn't, hidden colliders still block. One test initially "failed" because 120
ticks at 4 u/s travels exactly 8 units — it was measuring the tick budget, not
the collider.
STATIC PROPS NOW CAST. rebuild_static_shadows only walked entities, and props
are ModelInstances whose colliders are hidden, so trees and houses cast
nothing. Placed models feed the same baked layer, caster points sampled from
the model mesh (extremes always kept, then strided to ~48 — a stride alone
misses roof ridges) so a pine's shadow tapers. Cached on (render_rev,
bake_generation, models_rev), merged into one geometry, one draw.
THE SCENE IS COMPOSED: five suburban houses set back from a road all FACING it
(uniform facing is the point — random yaw reads as debris), lamps on one verge,
benches on the other, a fence line, three separated tree stands rather than
uniform sprinkling, and the physics demo gathered into a builder's yard. Props
scale to a target height from their own bounds, since a fixed multiplier gives
a 12-unit bench beside a 2-unit house. Exhaust only emits above 3 u/s (a parked
car under its own smoke column read as a bug).
44 props, 39 colliders, 7 draw items, 15.8k triangles.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Third attempt at this, and the first two failed in instructive ways.
A Kenney GLB's material URI is `Textures/colormap.png` RELATIVE TO THE GLB,
and the GLBs land in the pack root. So the atlas has to end up at
`<pack>/Textures/colormap.png` — no more, no less.
- Attempt 1 extracted no PNGs at all ("GLB is self-contained" — false here).
- Attempt 2 flattened every PNG into the pack root, so the file was present
but at a path nothing resolves; indistinguishable from missing, and it also
dragged in ~200 MB of Preview/Sample/thumbnail images nothing loads.
- This one keeps the `Textures/` tail and drops everything above it, because
archives nest it under a per-pack folder (`FBX format/`, etc.) that must not
survive. Preview/Sample images are filtered out.
The `.extracted` marker added earlier is what makes each attempt verifiable
rather than hopeful: a pack extracted by an older, wrong version is refetched
instead of being reported cached.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two bugs, and the second hid the first.
The extraction moved only .glb/.gltf/.bin on the reasoning "GLB is
self-contained" — false for Kenney, whose materials reference an external
Textures/colormap.png shared across the pack. PNGs are extracted now. They are
tiny: 212 atlases, ~42 KB total.
The resume check then counted MODELS only, so a pack whose atlas had never
been extracted was cheerfully reported "already cached" and never refetched —
which is why a full re-run fixed nothing and 48 of 52 packs rendered
untextured. Caching is now gated on a `.extracted` marker written only after a
complete extraction, so a partial or superseded extraction can never pass as
done.
(The PNG fix existed briefly in a concurrent branch of work and was clobbered
by another edit to the same script; the marker is what makes it verifiable
rather than hopeful.)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE BLOCKER WAS UPSTREAM OF THE RENDERER: zero of the 4,442 models had a
texture on disk. download_assets.sh extracted only .glb and deleted the zip,
reasoning "GLB is self-contained" — false for Kenney, where every material
points at an external Textures/colormap.png shared across the pack. PNGs are
extracted alongside now, and the resume check REQUIRES a texture, because a
pack of GLBs with no atlas renders white, which is worse than missing.
(The atlases are tiny: 212 textures, 42 KB.)
Static path (model.rs) reuses skin.rs's container/JSON/accessor code rather
than growing a second parser. A static mesh is a skinned one minus joints,
plus one difference: a prop never animates, so each node's world transform is
BAKED into its vertices at load and the model becomes one buffer. Dropping
that bake is exactly how a prop silently renders at the origin, so there is a
test for it. All 4,442 models parse: 1.31M triangles total, 294 average —
comfortably Quest-sized.
Kenney ships TWO conventions, and the second only turned up by looking at a
failure: most packs UV-map into colormap.png, but nature-kit and friends carry
no texture at all and colour each primitive with a material baseColorFactor.
Rather than branch, that factor is baked into the packed vertex's colour lane
and multiplied in the shader (albedo * v_tint) — atlas models carry white,
untextured models get a white 1x1. One shader, both conventions. A model that
DECLARES an atlas but cannot find it stays a hard error; that case really is
broken.
Batching sorts instances by model so equal geometry+texture land adjacent and
accumulate into one draw item: the demo runs 36 instances in 5 draw items,
9,887 triangles. Copies of a prop are free; cost is per distinct model.
The demo picks props BY DESCRIPTION through the asset index (find("pine
tree")), not by hardcoded paths, so it exercises the same path a generated
game takes — and it walks the ranked hits taking the first that loads, so a
pack with a missing atlas yields to the next candidate instead of leaving a
hole. Pillar ring and cone removed; they read as a test harness.
Honest read of the captures: before, coloured cylinders and spheres on a slab
— unmistakably a tech demo. After, a woodland treeline at mixed scale and
species, a suburban house with windows and a teal roof, wooden fences,
textured crates, correctly lit and shadowed. Still imperfect: "boulder"
resolves to nature-kit/cliff_blockCave_rock, a cave-mouth block that reads as
a small building scattered about — a SEARCH-QUALITY gap for the alias owner,
not a render bug.
Washed-out look diagnosed (not fixed, out of scope): it is FOG, not the bake
or the textures. SkyConfig::default()'s density mixes every surface toward the
pale horizon (0.75,0.87,0.96) over a 34-unit camera distance — the far
treeline desaturates toward sky colour while near crates keep their brown. Fix
is either a lower default density or making fog colour follow the sun's
horizon tint so it reads as haze rather than a grey wash.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
THE BUG: GLB_MAGIC was 0x4655_4C67, which spells "gLUF" — not "glTF"
(0x4654_6C67). The magic check therefore rejected EVERY REAL GLB, probe()
returned defaults, and the entire 4,442-model library indexed with
rigged:false, animated:false, size:None. Size filters silently matched
nothing; no model was ever detected as rigged. Any claim made from that
metadata — including "Kenney has essentially no rigging" — was measuring a
no-op, not the catalogue.
It stayed invisible because THE TEST FIXTURE WROTE THE SAME WRONG MAGIC, so
the test and the bug agreed with each other. Fixing the constant broke that
test, which is exactly how a fixture should behave once it stops encoding the
defect. A second bug sat behind it: bounds() searched for "max" only AFTER
"min", but Kenney's exporter writes max first, so bounds would have failed
even with the magic fixed. Both fixed, both with regression tests.
Consequence: the previously-reported 120 ms index build was timing a no-op.
Real probing is ~1.8 s for 5,309 models, now cut to the declared JSON chunk
and parallelised across <=8 threads (std-only, order preserved,
deterministic). The proper fix is caching probes by path+mtime — NOT done, and
the perf bound is now 12 s with a comment saying why rather than a tight
number the test cannot control under contention.
KIT INVENTORY — 23 kits, 2,064 tiles, grouped so a query returns a coherent
visually-matching set instead of one tile from each of five kits. Tile size is
the MEDIAN horizontal extent (kits ship occasional double-width pieces, and a
mean lands between grid pitches — a value no tile uses). Highlights:
city-kit-roads 72 tiles @1.00, coaster-kit 183 @4.00, tower-defense-kit 160
@1.00, marble-kit 162 @1.20, platformer-kit 153 @1.00, modular-buildings 108
@1.00, racing-kit 112 @1.05. The most useful single fact: modular-dungeon,
-cave and -space kits have IDENTICAL role histograms — one layout algorithm
drives all three and the kit choice is pure theming.
Honest failure: city-kit-commercial (41) and city-kit-industrial (25) yield
ZERO roles — their files are building-a..building-z, whole buildings with no
role vocabulary. Grid-placeable but not composable; arguably not kits.
Adjacency ships as DATA (ROLE_ADJACENCY) for the composition layer and is
deliberately coarse: Kenney filenames say what a piece IS, never which edges
are open, so anything finer would be invented. Also added: role/kit/clips/
joints on entries, kits()/kit_tiles() grouping, a find_kit agent tool (<2 KB
so the AI can discover a coherent set before composing), and composition-intent
vocabulary.
Inert per the Kenney-only scope cut: clip extraction, Quaternius source
support, .gltf support — tested and harmless. 64 fetched Quaternius models
were deleted after verifying they parse (46 joints/13 clips).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Procedural LAYOUT + authored TILES: the AI decides where things go, Kenney's
artwork decides how it looks. Beats both random prop scatter and purely
procedural geometry, and it is how low-poly games are actually made.
ADJACENCY HOLDS BY CONSTRUCTION, not by rules. Rather than pairwise rules
between named roles (fragile and quadratic), layout marks occupied cells, each
cell reads its target mask off its occupied NEIGHBOURS, and a tile is chosen
matching that mask at some rotation. Both sides of every shared edge derive
from the same grid, so only rotation arithmetic can be wrong — and that is
what the tests pin. Junction type is never specified by a caller: two crossing
paths yield a crossroad, one teeing in yields a T, purely from neighbour count.
The interface deliberately keys on a 4-bit N/E/S/W `mask`, not on `role`, so
these algorithms don't depend on the asset index's filename taxonomy — if a
kit classifies `road-split` oddly, setting the mask keeps everything working.
Incomplete kits fall back to a superset tile: a crossroad standing in for a
missing tee leaves a stub opening onto nothing, which reads as unfinished road
rather than a hole in the world.
Generators: road_network (polylines), road_from_spline (the authored-tile
counterpart to the existing ribbon mesh — a kart track wants the ribbon, a city
street wants tiles), town (street grid, buildings on lots that front and face
a street, props at junctions), dungeon (BSP rooms + corridors, connectivity
guaranteed by the spanning tree and PROVED by flood fill over 12 seeds), plus
place_tile as the escape hatch.
track from closed spline 13 us 120 tiles
road network (13 paths) 32 us 397
town 24x24 71 us 547
town 60x60 822 us 2710
dungeon 48x48 137 us 1180
dungeon 96x96 932 us 3616
Town road histogram: 1248 straight, 121 cross, 44 tee, 4 corner, 0 dead ends —
correct for a closed grid. Zero mismatched edges on both large levels.
Two bugs caught by its own tests: indexing one kit with another kit's
placement indices (now impossible — layers merge by kit id, invariant
documented), and a superset-fallback that allocated a Vec per cell and tripled
generation time. The allocation-free count-then-pick rewrite is faster than
before the fallback existed: dungeon 96x96 went 1952 us -> 932 us.
Seed-deterministic via GenRng, never the world rng, so a town replicates as
(kit, seed, params). Not done: walls/doors around dungeon rooms (floor-only
today), multi-cell buildings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A generated game handed box3d an infinite extent and took the whole process
down — found by the eval harness actually generating games, not by reading
code.
NaN was already absorbed (Rust's `max` returns the non-NaN operand), but
INFINITY survives it and poisons every plane normal to NaN. box3d's face query
then never beats its -f32::MAX starting separation, leaves max_face_index at
the -1 sentinel, and convex_manifold.rs casts that sentinel through u8 into
255 and uses it to index a 6-element array.
The port is FAITHFUL to upstream C here — convex_manifold.c does the same
(uint8_t)maxFaceIndex cast; C reads garbage where Rust panics — so box3d is
not the place to diverge. The boundary is: never hand the solver a value it
cannot reason about. sane_extent() clamps non-finite and out-of-range
dimensions (and density) to a workable range.
Regression test covers +inf, -inf, NaN, 0 and negative extents: none may panic
and all must leave a finite pose after 30 ticks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tools/arcade_eval drives the REAL agent through a 14-prompt suite, evaluates
each result headlessly through ScriptHost, and scores it — then the context is
improved from the aggregate failure modes and re-measured.
Dominant functional failure, now quantified exactly: game.terrain without
smooth:true spawns one static box PER CELL. Every high-entity case matched
cells^2 arithmetic precisely (96^2=9216, 64^2=4096, 48^2=2304, 40^2=1600) —
one misunderstanding causing the entity explosions, the slowdowns and an
engine crash. Racing went PANIC/2382 entities -> PASS/22.
Engine bugs found by generating games rather than by reading code:
- `input` was NIL in every on_tick (the host's NIL marker went through
unresolved), so every generated game reading input was broken
- wrong-typed options coerced silently: size:[1,2,3] -> vec3(0,0,0),
color:"#ff0000" -> grey
- `loop:` as an option key hangs the VM until the instruction limit fires —
and dispatch.rs ADVERTISED `loop` in the model-facing API text, so the
engine was telling the model to write the thing that hangs it. Removing it
from the doc string turned two eval failures into passes
- box3d panic (still live, fixed separately): convex_manifold.rs launders a
-1 "no face" sentinel through `as u8` into 255 and indexes a 6-element array
THE FINDING THAT MATTERS: every generated game looks catastrophically bare —
and so does our own hand-written 72-line model-answer fixture when rendered
through the same path (an empty green field, two dark rectangles, two white
boxes). That decisive test rules out generation failure AND documentation gap.
game.model, game.material, game.tree, game.scatter and find_model appear ZERO
times in splashgame.md and in the model-facing api_text(): the 4,400-model
Kenney library, the generated trees/rocks/scatter, and the material presets
all exist as crates with no script binding. The AI is faithfully reproducing a
bare aesthetic because bare primitives are the only vocabulary the engine
exposes. Prompt tuning cannot fix that; binding the libraries to verbs can.
Also visible in the reference capture, engine-side rather than generation:
flat lighting with very low sky/ground contrast, shadows too weak to ground
objects, no AO on primitives, and a default camera that frames poorly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
decodable is no longer gated on format — acb315614 took the in-house Vorbis
decoder to sample-exact on every shipped file, mono and stereo. The
`undecodable` reporting path stays for a future format we might index before
we can play it; the test now asserts the CURRENT catalogue is clean rather
than asserting ogg is broken.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The bug was NOT residue type 2 — that lead was a reasonable inference from
"stereo-only, transient-heavy", and it was wrong. The cause was overlap-add
placement of early long blocks.
A block's window is centred on `center` and reaches n/2 either side. A file
opening [256, 256, 2048, ...] puts the first long block's centre at 832, so it
starts at -192 — before sample zero. Those leading samples lie outside the
stream and must be DROPPED. The code used center.saturating_sub(n/2), clamping
the start to 0, which slid the whole block 192 samples later. Every sample was
corrupted until the centres grew past n/2, then decoding was perfect again.
That shape is exactly why it read as a residue fault: a wrong head with a
correct body looks like "specific blocks have wrong amplitude", and
correlation averaged it to 0.82. Mono appeared flawless only because no mono
file in this corpus happens to open with an early long block — a corpus
accident, not a decoder property.
mono 47/47 exact, mean 1.00000 -> 186 files, mean 1.00000, min 1.00000
stereo 68/107 exact, mean 0.826 -> 370 files, mean 1.00000, min 1.00000
corpus 115/160 exact -> 556/556, zero decode errors
The 73 "frame-count mismatches" are afconvert trimming further than the
container specifies; afinfo's valid-frame counts match OUR output exactly and
every file still correlates at 1.0000.
The fix is extracted into a shared overlap_add because decode and debug_raw
each had their own copy — a diagnostic that can disagree with the decoder it
diagnoses is worse than no diagnostic.
New test is fixtured on a file that opens [256, 256, 2048, ...] and asserts
PER-SAMPLE agreement, not just correlation: correlation alone hid this at 0.82.
Decode cost 5.16 ms/file; 11.5 MB compressed expands to 143.3 MB of f32 PCM.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two root causes, both found by building an oracle rather than guessing.
1. Amplitude ~75x low: the IMDCT applied a 2/n normalisation the encoder's
forward transform had already carried. Because 2/n varies with block size
it produced DIFFERENT errors on 256- vs 2048-sample blocks — exactly the
reported symptom. Removing it gives fit scale 1.0000.
2. Leading trim, the real remaining defect. The first audio packet produces NO
output (its window only primes the overlap-add) but we emitted from the
first block's centre, injecting half a priming window of garbage and
shifting everything early. And Vorbis carries encoder delay in the GRANULE
POSITION, which varies per file — afinfo confirms 128 / 1103 / 960 frames
on three samples — while our Ogg reader kept only last_granule and
discarded per-page granules, making it unrecoverable. Added per-page
granule tracking: the first page reporting a granule pins priming as
centre - granule, and valid audio starts at priming + blocksize_0/2. That
reproduces afinfo's numbers exactly on all three.
A premise in the brief was also wrong and worth recording: our output length
was already correct. afinfo reports valid frames matching OUR output — it is
afconvert that trims a further 128. The reference WAV was short, not us.
mono 47 files mean corr 1.00000 (min 1.00000) 47/47 exact
stereo 107 files mean corr 0.826 68/107 exact
Decode cost 5.13 ms/file average; 11.5 MB compressed expands to 143.3 MB of
f32 PCM, which is why the sample bank's LRU cap matters.
Honest remaining defect: ~39 stereo files decode wrongly and it is NOT
alignment — a full lag sweep peaks at 0.40-0.89 with fit scales 0.40-1.87, so
specific blocks have wrong amplitude. Mono being 47/47 rules out floor,
residue 0/1, MDCT, windowing and priming; coupling matches the spec's
square-polar mapping including reverse order; floor 0 is rejected rather than
mis-decoded; and both channels are identical in the failing files, so it is
not a swap. The failing set is transient-heavy impact/footstep sounds, so the
lead is residue type 2 partition counting on short blocks.
reference_decode.rs is no longer #[ignore]d: 3 real tests asserting mono
correlation > 0.999 and length == granule, plus a 3000-mutation fuzz that must
never panic, all skipping cleanly without fixtures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
apps/arcade/src/lib.rs (plus the [lib] section committed alongside it in
3cef1eb29, which referenced a file that wasn't tracked yet — HEAD did not
build without this).
The point is stated in the module doc: tools/arcade_eval must send the same
system prompt and the same tool policy the app sends, or it measures a
fiction. Exporting the modules is what keeps the harness and the app from
drifting; the binary keeps its own mod declarations because app_main! owns
the process entry point.
Carries the game_script changes the harness needs alongside it (input.rs and
the dispatch/host/value edits made while wiring the headless evaluation path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
libs/converse gains a default-on `tts` feature so gamemaker and route are
unaffected; Arcade takes default-features = false and re-enables it through
its existing `voice` feature. With tts off, SpeechOutput still exists and its
worker drains the queue silently, so no Arcade source changes were needed —
which also avoided colliding with concurrent edits to those files. No
capability is lost: speech is fully reachable under --features voice.
Arcade before 25.11 MB
Arcade after (default) 15.01 MB (-10.10 MB, -40%)
Arcade --features voice 26.18 MB
hello_world 13.46 MB
Isolated by A/B in a throwaway worktree before making the real change. Of the
10.1 MB, 2.78 MB is us_lexicon.bin embedded via include_bytes!; __const alone
was 11.2 MB, larger than __text.
Also: Kokoro construction is now lazy, deferring ~327 MB until first
utterance. That already closed the RSS gap this pass was chasing — Arcade's
max RSS measures 190 MB against hello_world's 222 MB on the same method, i.e.
BELOW the baseline app, so the premised memory problem no longer exists and
was not invented into one. The lean-isolate prelude was likewise dropped after
instrumentation showed Arcade allocates no script isolate on the demo path at
all; it should be re-measured against a loaded script game before anyone
spends effort there.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Seed-deterministic, baked at spawn, emitted straight into the 24-byte packed
vertex format. Two devices given the same seed produce byte-identical
geometry — which is what lets a forest replicate as (preset, seed, position)
tuples instead of mesh data.
- L-system plants: expansion + 3D turtle + skeleton, 8 species (oak, pine,
palm, bush, fern, cactus, dead, grass)
- Surface nets (not marching cubes — fewer, better-shaped triangles at
low-poly) for rocks, boulders, mushrooms, clouds, blobs
- Spline tracks with width, banking, curbs and rails, returning centreline
frames that carry lap distance — so spawn points and checkpoints derive
from the track instead of a second hand-written list
- Poisson-disk scatter with flatness/height rules
- Texture generator with CPU mip chains (backends never generate them)
- LRU cache keyed by FNV-1a over the full recipe with floats hashed by exact
bits; -0.0 and 0.0 normalise together (identical geometry), 4.0 and
4.000001 stay distinct. Meshes hand out as Rc so eviction cannot pull
geometry out from under a frame mid-draw
- DrawGameFoliage: growth and wind as an OPT-IN shader variant, a sibling of
DrawGameSkinned rather than a flag inside the shared shaders — wind costs
~20 vertex ALU and the cube shader draws most of the world. Growth and flex
weights share one packed nibble pair, so both animations cost zero extra
vertex bytes
A realistic forest — 150x150 m, 582 trees, 3 species, 6 seeds — generates in
1.70 ms with 470 KB resident: 6 generations and 576 cache hits.
Three bugs the unit tests had passed, found by writing an ASCII silhouette
probe because captures were out of scope: every species came out ~4x its
requested height (the test compared two sizes RELATIVELY, so a uniform
overshoot sailed through — now the finished skeleton is measured and rescaled,
and the test asserts absolute height for all 8 species at three sizes); palm
emitted no foliage at all because its L-system contained no leaf symbol; and
cactus sprawled sideways like a shrub. Pine also dropped from 4 iterations to
3 — 15032 -> 2504 triangles, 728 -> 68 us — because 15k triangles for one
background tree is indefensible.
Honest caveat: that is a silhouette judgement, not a rendered one. Shading,
leaf-card orientation and the wind/growth animation are visually unverified.
Script verbs are not wired yet — the crate is a library only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two ways HEAD failed to build for a fresh checkout, both from work landing
across concurrent streams:
- 623ee745e renamed skin_to_pbr -> skin_to_packed, but arcade_view.rs was
being edited by another stream at the time, so its call site fix stayed
uncommitted while the rename landed
- Cargo.toml listed tools/arcade_eval as a workspace member while the crate
itself was untracked, so loading the workspace failed outright
Both verified: the crate builds, and a stash-everything check now leaves a
working tree that compiles.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The compiler is where hostile shaders get stopped, because this engine lowers
the script shader language to Metal/HLSL/GLSL/WGSL rather than passing source
to a driver. Two genuinely exploitable holes found and closed:
- Loop bounds were arbitrary EXPRESSION STRINGS written straight into the
emitted shader, so `for i in 0..some_uniform` compiled to an unbounded GPU
loop and `loop{}` emitted a bare `while(true)`. A hang triggers a driver
device reset that kills the app — ugly anywhere, worst in a headset. Rather
than reject (which would break legitimate code), the bound is now EMITTED: a
provably-small integer literal compiles unchanged, anything else gets a hard
65536-iteration cap. literal_bound() is deliberately conservative — a
uniform, arithmetic, a call, hex or a negative all count as unprovable
- compile_fn INLINES at every call site, so a branching call graph expands
exponentially with depth (recur_block stops self-recursion, but not f1
calling f2 twice calling f3 twice). MAX_EMITTED_BYTES (1 MB) bounds what was
an unbounded compile-time DoS
Recursion was already safe (recur_block errors); nothing added there.
Honest remaining gap: the cap bounds ITERATIONS, not cost per iteration — a
shader doing 65536 heavy texture samples is legal and slow. Bounding real GPU
time needs a cost model or driver watchdog; neither exists here and this does
not claim otherwise.
Verified empirically, not just by compiling: all 24 loop{} constructs in the
built-in shaders are guarded (48 emitted lines, 4 unique guard names per
shader confirming no shadowing), zero for( loops exist in any built-in, and
the capture shows rendering intact — including text, which is exactly where
those guarded loops live.
MEASURED FIRST, then declined to build: 34 shaders compile at Arcade boot in
~15 ms total on Metal (first 4.47 ms cold, rest 0.10-0.74 ms). A persistent
shader cache is NOT worth its invalidation-and-staleness surface to save 15 ms
of one-time boot, so it wasn't built. Vulkan/GLES on Quest may differ — that
needs on-device measurement before anyone builds speculatively. Benchmark
retained behind MAKEPAD_SHADER_BENCH (it was logging every boot);
MAKEPAD_SHADER_DUMP=<dir> writes generated source, which is what an AI has to
debug and was otherwise invisible.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
52 packs, 4442 GLB models, 136 MB on disk — fetched sequentially with resume
(a hash-valid pack is skipped, so an interrupted run costs nothing) via
kenney.nl's content-hashed URLs with per-zip sha256. MIRROR.toml records
every pack's canonical URL, sha256, size and file count, so a mirror is
reproducible; --mirror=/ARCADE_ASSET_MIRROR redirects the base URL and fetch()
verifies the digest identically whatever host served the bytes — a mirror we
control is never trusted more than upstream. --packs= keeps a fresh clone from
being forced to pull everything.
Aliases restructured to survive the scale: per-pack theme rows (55) so every
model in a pack inherits its setting, filename-token parsing with variant-
marker stripping as the workhorse, and ~240 hand-curated query-time synonyms —
the layer whose curation compounds across the whole catalogue. 82-query suite
reports misses instead of being tuned green; the list is down to 2, both
defensible (a floor IS somewhere to stand; a bell IS a metal clang).
Three ranking bugs root-caused, not patched:
- No stemming, so "smashing" never reached the alias "smash" and "glass
smashing" returned glass PIPES. Added a conservative stemmer probed at
synonym strength (only ever adds matches), which refuses to mangle
glass/grass/class and routes "trees" to "tree", not "tre"
- An overreaching alias: `spaceship` sat on four spaceEngine SOUND families.
An engine hum is not a spaceship. Removed; "spaceship engine" still resolves
- Kind confusion on ties: spacecraft models tied with spaceTrash sounds and
lost the alphabetical tie-break. Added kind-aware tie-breaking driven by
query intent — deliberately a TIE-BREAK, not a score bonus, so it cannot
drag a weak model above a strong sound (laser gun / explosion / coins scores
verified unchanged)
Repo-policy violation fixed: all three asset .gitignore files were deny-lists
covering only .glb/.png/.jpg, leaving 302 .gltf files from 3d-road-tiles fully
committable. Converted to allow-lists — 4,744 asset files are now unstageable
by accident.
Scale at 4,999 entries: build 120 ms, search ~0.2 ms, 2.1 MB heap, and the
prompt summary still 479 chars — flat as the catalogue grows, which is what
keeps it affordable in every AI turn.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two bugs, one visible symptom — the Knight strolling through the crate stack.
1. Rigid bodies were absent from the mover sweep set and from raycast /
camera-boom queries. When box3d dynamics landed, movers kept sweeping only
against Static|Kinematic, so a character (and a bullet, and the chase
camera) passed straight through every crate. Rigid poses are read back from
box3d at the end of the previous tick, so at snapshot time a rigid is as
settled as a kinematic and belongs in exactly the same set.
2. Fixing (1) exposed a deeper one. Clamping left the two boxes EXACTLY flush
(|d| == sum of halves), where float error decides the next axis' overlap
test either way — and a "yes" sent the falling mover UP onto the crate,
straight through the documented 0.55 step-up limit. It then walked along
the crate top. CONTACT_SKIN (1e-3) makes resting contact stop a hair short
of flush, so contact is unambiguous instead of borderline. Verified: a
walker into a 1.0-tall crate now stops at its face (x 1.20, y 0.50) rather
than climbing to y 1.50.
mover_is_blocked_by_a_rigid_body pins the behaviour. The mover golden hash is
re-baselined once, deliberately — the skin shifts every clamped position by
1e-3, and the movement it now describes is correct rather than merely
different. The reason is recorded at the assertion so a future change to that
hash needs the same justification.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
step_world cloned two things per tick purely to dodge a borrow: the whole
Terrain (heights AND colors) and every static/kinematic Entity at 208 bytes
each. Terrain dominated — a 257^2 field is 1.3 MB/tick, 79 MB/s of memcpy at
60 Hz, on a world containing seven entities.
- Terrain: copy -> borrow, splitting the struct borrow the way the bottom of
the same function already did
- Statics: 208-byte Entity -> 48-byte Solid. This one MUST stay a copy —
movers sweep against kinematic poses from BEFORE this tick's integration
and that ordering is load-bearing — but it only ever needed
id/kind/pos/half/vel
- owner_pose: skip building the table when nothing is attached (most worlds)
scene ms/tick B/tick
demo 0.002 -> 0.003 15,140 -> 4,796 (-68%)
racing-ish (129) 0.007 -> 0.002 362,316 -> 8,576 (-98%)
terrain 257 0.019 -> 0.001 1,323,964 -> 896 (-99.93%)
large (500 static) 0.063 -> 0.056 591,386 -> 82,382 (-86%)
stress (2000 static) 0.583 -> 0.457 2,353,936 -> 327,812 (-86%)
Result-neutrality proven, not assumed: new mover_golden.rs covers what
rigid_dynamics.rs doesn't reach (terrain cliffs/floors, sweeps, platform
carry, attach pin, projectile lifetimes, auto-face) and its golden hash is
identical before and after — verified by stashing only the source changes
and re-running, not by re-baselining. Also includes a test pinning the
pre-integration snapshot ordering, so a future "obvious" simplification that
reads live positions gets caught.
Leak check: 36,000 ticks (10 simulated minutes) of a busy world with
projectiles spawning and expiring — RSS flat at 3.8 MB, +0.4% drift.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
libs/game/audio (77 tests): WAV decoder (8/16/24/32-bit int, f32/f64,
total on malformed input), Ogg container with packet reassembly, sample
bank with on-demand decode + linear resample + LRU eviction and pinning so
a playing voice can't be evicted, and a 24-voice mixer with equal-power
pan, playback-rate pitch, click-free fades, priority+age voice stealing and
a limiter. Generation-tagged voice handles mean a stale handle cannot
retune a reused slot. The limiter test caught a real bug: peak was being
measured per-voice, but what clips is the SUM.
Emission is the point — sounds come from the engine observing gameplay,
not from script calls: Material/MaterialPair (order-independent, the softer
material names the sound), an impact curve mapping closing speed to
gain/pitch, and an AudioDirector with repeat-avoiding variant selection,
per-category volumes, per-pair cooldowns and a per-frame cap. A 200-contact
frame yields <=6 sounds and the cooldown map is proven not to leak.
Blocks now emit their own audio: car engine tracking revs, skid on lateral
slip, suspension thud on landing; character footsteps timed off the WALK
CYCLE rather than a timer, so feet and sound stay together when slowing;
jump/land scaled by fall speed; plane engine by throttle; lap and win
stings. RNG isolation proven: heavy audio work interleaved with world-rng
draws leaves the drawn sequence bit-identical to a silent run.
KNOWN GAP, reported rather than hidden: every Kenney audio pack is Ogg
Vorbis only (471 files, zero WAV), and the from-scratch Vorbis decoder is
NOT correct yet — setup header parses exactly, channels/rate/frame count
and envelope shape are right, but floor magnitudes come out ~75x low and
best correlation against an afconvert reference is 0.67. Two real bugs were
found and fixed en route (type-1 residue filled one codeword instead of the
partition; MDCT post-twiddle carried the pre-twiddle's 1/4 term). The
reference test is committed as #[ignore] with its measurements in the
message so it stays runnable. Ogg stays flagged unplayable in the asset
index; WAV and the --transcode path work today.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Quest is vertex-bandwidth bound, so this is the measured headline:
cube instance 176 B -> 128 B (-27%)
skinned character vertex 64 B -> 24 B (-62%, re-uploaded EVERY frame)
shadow mesh vertex 64 B -> 24 B (-62%)
The Knight went 238 KB/frame -> 89 KB/frame: CPU skinning re-uploads the
whole buffer each frame, making it the largest recurring saving available.
Instance sizes are read from the compiled shader (RenderStats::
instance_floats), not counted by hand. The instance win was pure
duplication: sun_color/sun_sky/sun_ground/fog_color are identical for every
instance in a batch — 12 floats per cube — and moved to uniforms.
fog_density stayed per-instance because shadows switch it off individually.
Unblocked by adding geom.GameMeshVertex in draw/geometry_gen.rs and making
the existing pack_pair_f16/pack_unorm8x4 public, rather than writing a
second f16 rounding implementation that could drift from the first.
Three constraints found, worth keeping:
- Vertex attributes here are f32-ONLY. Compression means bit-packing into
f32 lanes; unpack2f16/unpack4u8 are builtins on every backend
- Pod vertex structs need flat f32 fields, not Vec3f — std140 pads a vec3
to 16 B and the repr(C) size assertion fails at runtime
- In the shader language `let` is immutable and helpers can't be forward-
referenced, so the octahedral decode uses branchless step(0,v)*2-1: the
sign() builtin returns 0 at 0, which would collapse the fold on
axis-aligned normals
Tape BYTE_IDENTICAL; captures verified after each conversion (shadows
unchanged by packing, Knight correct with packed normals/UVs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
75 models (5 KenneyNL starter kits, .glb, pinned commits + sha256, 2.8MB)
and 556 sounds (7 packs via kenney.nl content-hashed URLs, each zip
sha256-verified, 13MB). Nothing large enters git: dirs are gitignored, only
CREDITS.toml and .gitignore files are tracked.
The index is the point — an AI cannot use a library it can't name:
- id is kenney/racing/vehicle-truck-yellow, anchored to where the file
LIVES, not its category, so retuning the category tree never invalidates
a saved game
- Filename tokens are the floor; the value is two hand-curated alias tables
(76 model rows, 116 audio FAMILY rows — Kenney's footstep_wood_000..004
collapse to one family, so 556 files stay maintainable) spanning
synonyms, kid vocabulary and misspellings (vehical, hosue, motercycle),
function over identity ("something to hide behind"), colour/size/
material, and theme, plus ~190 query-time synonym expansions
- AssetKind model/sound/music so a 30-second track can't be returned as a
hit sound; GLB probe reads skins -> rigged, animations -> animated
- FIND_MODEL tool descriptor (provider-neutral plain data) + compact
results; library_summary() is 469 chars for 632 entries and provably
doesn't grow with the catalogue; resolve_or_explain() rejects
hallucinated ids with near-misses; local_spawn() gives the local
librarian a best match plus a confidence blending strength with margin
HONEST GAP: every Kenney audio pack is Ogg Vorbis only — no WAV exists
upstream — and this tree has no vorbis decoder. Sounds are indexed and
searchable but NOT playable: entries carry decodable:false, the agent JSON
emits playable:false so a game cannot fire a silent sound, and
--transcode converts via ffmpeg when present. A real decoder is the fix.
Three bugs found by testing, all fixed: sci-fi-sounds.zip ships a directory
with no owner-write bit so that pack alone silently extracted 0 of 73 files;
prepositions matched phrase aliases ("...at the roadworks" hit "something to
shoot at"), so function words must be dropped, not down-weighted; and exact
names lost to incidental aliases (a coin SOUND outranked the coin MODEL).
Miss list left visible at 2/52 rather than tuned away — both are defensible
answers against over-narrow expectations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Arcade had no chat UI and no cx.audio_output at all, so M7's positional
audio was silent there and there was no way to talk to the AI.
- synth.rs is a port of gamemaker's 24-voice synth (a shared libs/game/audio
crate was out of this task's scope; deduplicating the two copies is a
mechanical follow-up). Gamemaker's audio path is untouched, so no tape
run was needed. One behavioural addition: STEREO — gamemaker's synth
writes the same sample to every channel, so a positional sound had
nowhere to go. Voices now carry pan, and centre keeps full volume in both
channels rather than equal-power, so every existing 2D sound is exactly
as loud as before
- audio.rs drains the AudioRequest queue and resolves SfxAt against a
listener built from THIS device's camera — Local tier by construction,
nothing reaches the wire. A sound past its range queues no voice at all,
so a busy world doesn't burn its 24 slots on things nobody can hear. The
demo world clanks when crates land, so the positional path is audible out
of the box rather than merely implemented
- chat.rs: PortalList with User/Assistant/System bubbles — engine trouble
gets its own colour because the player didn't say it and the AI didn't
either. main.rs is now a Splitter: chat + input + status left, game right
- Voice degrades per tier: a `voice` feature gates the mic (local-llm
implies it — a judge needs a mic to judge), the text box exists in every
tier. Caught in headless boot: naming ptt_use_escape in script logged an
[E] every startup when the feature is off, because VoiceWave is a stub
View there; push-to-talk moved to feature-gated Rust
- authoring.rs submits the agent's edit to the intent log as a transaction
against the generation the turn started from, then writes the merged head
back to disk — otherwise the next mtime poll would re-propose the agent's
stale text as new. Eval errors return via CoeditResponse::EvalError and
land in chat as System messages naming the last version that worked.
Tests assert a typed request reaches the log, and that the LOCAL agent
gets rebased when it loses a race to a remote one — no privileged path
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CPU light baker (bake.rs): per-static AO (5 face samples x 8 Fibonacci
rays), a sun-visibility term, and a trilinear probe lattice for moving
objects — all folded into instance colours the renderer already sends, so
zero extra bandwidth and zero GPU cost. Demo world, release: AO 15us,
sun 34us, probes 61us. The split is deliberate — AO is the expensive half
and is sun-independent, so a day/night cycle only pays the 34us. A ray
starting above the heightfield peak and heading up skips the terrain march
entirely: that early-out took the probe pass from 5.4ms to 61us.
Silhouette shadows (shadow_mesh.rs) replace the flat oriented quad: caster
points -> projection along the sun -> 2D convex hull -> fan triangulation,
which has no self-overlap and therefore cannot double-darken in an alpha
blend (the reason naive projected geometry bands). Draped over terrain
(vertices drop to ground height, long edges subdivide), soft rim from a
penumbra ring that widens with height, statics cached against (world edit,
sun position) — every shadow in a frame is ONE geometry, ONE draw call.
Z-fighting handled structurally: offset along the RECEIVER's normal with a
slope-scaled term (world-up slides the shadow on a slope), depth test on,
depth write off.
Instance stream 176B -> 128B (-27%), measured from the compiled shader:
sun_color/sun_sky/sun_ground/fog_color were 12 floats of identical data on
every cube and moved to uniforms. fog_density stays per-instance because
shadows switch it off individually.
RNG isolation is structural: GameWorld has no bake field and the ray set is
fixed, so there is no RNG here to share with the sim. Tape BYTE_IDENTICAL.
Deleted an unwired SDF-blob path and the dead project_box_shadow call site
rather than leaving two shadow implementations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Desktop entry runs it under studio like the other apps; the quest entry
uses the same RunQuest wrapper as makepad-example-xr, which M3 verified
packages an arm64-v8a APK with the passthrough/handtracking manifest bits.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Committed together: both streams landed in libs/game/script, so splitting
them would produce two commits that don't compile.
M6 — packaging and sharing
- libs/zip_file gains a writer (store + deflate); real `unzip -t` validates
our archives in an interop test. Packing is deterministic (fixed
timestamps, sorted entries), so a package can be addressed by its own
sha256 — which is what makes the registry's digest check mean anything
- libs/game/pkg: .arcade format (game.splash + manifest.toml + assets),
total manifest parsing (attacker bytes always yield a Manifest or an
error, never a panic; non-finite numbers refused rather than defaulted),
registry client that verifies sha256 INSIDE download so tampered bytes
never reach the extractor
- Hardened extraction: absolute paths, drive letters (C:x is absolute on
Windows), UNC, backslashes, .., NUL/control chars, symlink members (via
mode bits), duplicate names (the ambiguity IS the attack), declared-size
caps checked before decompressing plus a post-decompress check, entry/
total/archive caps, and a post-join re-check that the resolved parent is
still inside the destination — which catches a pre-existing symlink the
name test cannot see. 4000-round mutation fuzz with a canary file beside
the destination; a 320 MB deflate bomb under 1 MB on the wire is refused
- Capability stripping rebinds fs/run/net to FRESH EMPTY OBJECTS rather
than shadowing known verbs, so there is no hole the day someone adds one.
Applied before the game handle is registered. Vacuity guard: an unstripped
isolate genuinely reads a file, so the sandbox tests can't pass for
unrelated reasons. Browser-installed games load Trust::Downloaded
M7 — pretty pass
- GameSun adopts draw::SceneSun (axis-converted: SceneSun is map-space
y-south/z-up, games are y-up). Shaders compute hemisphere ambient +
direct instead of each hardcoding its own split; defaults collapse the
new formula to the old constants exactly, so unifying did not restyle
existing games. write_into is the single write path — "one sun" is
compiler-enforced
- Projected shadow geometry: the caster's silhouette along the sun, fitted
in the sun's own (u,v) frame, so it stretches as the sun swings. Nearest
N casters get projection, the rest blobs; one instance in the existing
alpha batch, no extra pass. 0.6us for 24 casters
- Two pre-existing shadow bugs found via capture: the pipeline blends
premultiplied, so unpremultiplied dark RGB ADDED light instead of
removing it; and shadows were fogged, mixing them toward the bright
horizon so a distant shadow came out lighter than the ground it darkened
- Particles are structurally isolated from the sim: GameWorld has no
particle field and step_world has no particle code — the renderer owns
simulation and its own RNG. particles_never_advance_the_world_rng
interleaves particle verbs with real rand() draws over 32 rounds and
asserts both the RNG state and the drawn stream are identical
- game.sfx_at with listener-relative gain/pan and a near-field ease so a
sound at your feet doesn't flip channels; 2D verbs unchanged
- apps/arcade/BUDGETS.md: measured particle/sim costs, Quest columns marked
as estimates (the real particle limit is fill rate, not CPU)
Tape probe BYTE_IDENTICAL. Not done: arcade has no audio backend, so
positional sound is implemented and tested but not audible there yet.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
libs/game/coedit (zero deps, so merge logic is testable without a socket or
a VM). Not a CRDT: transactions are host-serialized into an append-only
generation history, and a conflict is answered by handing the author the new
base so THEY re-derive their intent.
- Transaction carries the author's whole intended file, not a patch: the
diff against its declared base is derived host-side, so a stale or
malformed patch can never be applied — and an AI writes whole files anyway
- diff3 over lines with LCS anchors. Merge::Conflict deliberately carries NO
merged text: a half-merged game file that still parses is worse than an
honest rejection. Conflict -> Rebase{new base + per-generation summary of
what landed underneath}. An edit already present in the tip is refused as
NoChange rather than appended as an empty generation
- Leases are advisory as designed: a test asserts a submit SUCCEEDS while
another author holds the lease. They shape who chooses to edit; they never
gate the log. TTL expiry means a crashed author cannot lock a region
- Wire: coedit is reliable-channel ONLY — a test signs a valid submission,
sends it by datagram, and asserts it is ignored, so no datagram can
rewrite the game. Every response is addressed, never broadcast
- Arcade bridge routes in exactly one place; local agent and remote authors
share the queue and the rules, and a test asserts the local agent gets
rebased identically when it loses a race. Remote players map to
AuthorId(player+1) so a client holding player id 0 cannot impersonate the
host's agent
- 200-round deterministic fuzz: 4 authors submit against deliberately stale
bases, rebase, resubmit — asserting linear append-only numbering, no
generation claiming a base from the future, and that replaying accepted
diffs from generation 0 reproduces the head exactly
Two real bugs found building it: validate used `?` on the base lookup, so an
unknown base returned "no refusal" instead of UnknownBase; and
MAX_COEDIT_SOURCE (512 KiB) exceeded MAX_FRAME_BYTES (256 KiB), so the host
would have accepted a source it could never hand back inside a Rebase,
stranding the next author on an answer that never arrives. Both now 192 KiB.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
game_view.rs 4525 -> 1878. Deleted in one marker-verified cut: the 84-arm
game_dispatch chain, all arg/option/value helpers, spawn_entity/spawn_terrain
/spawn_car/spawn_plane/spawn_character/spawn_block_body/set_brain, the
GAME_API table, suggest_verb/edit_distance, and the duplicated CallbackTable.
register_game_handle now binds game_script's verb table; unknown verbs keep
the same hard-fail plus did-you-mean. Kept: run_tick and everything parity
depends on — tape input, gamepad poll, camera mailbox, perf channels, agent
RPC, save/log flush.
Rollback deliberately still gamemaker's hand-written WorldSnapshot, not
GameWorld::clone(). Clone is strictly safer and is what Arcade uses, but
eval_body is the parity-critical path and the snapshot is what the tape was
established against — swapping it deserves its own tape run, not a rider on
a 2700-line deletion.
Three pre-existing divergences fixed in game_script (sfx/beep/jingle predated
the 31-verb port and had drifted): beep's `to` defaulted to 0.0 instead of
freq, so every beep swept to silence; beep never read its wave option at all
(allowed but unused, and AudioRequest::Beep had no wave field); gain and
jingle ms were off. Audio drains host-side in run_tick and after a successful
eval so a startup jingle isn't held a frame; a failed eval discards its queue.
Regression the migration exposed: block verbs spawn through the shared box
path, which re-validated keys against the BOX allow-list, so game.car(...)
logged 8 bogus "unknown option" warnings per eval straight into the channel
the agent reads. Added spawn_entity_unchecked for the block path; a real typo
on game.box still warns. Racing fixture: 8 warnings -> 0.
Tape probe BYTE_IDENTICAL (re-run after the audio and warn changes too).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage as presentation, never simulation. Stage{mode, origin, yaw, scale} is
applied as the scene draw list's view_transform uniform — one value per
frame, so sky/terrain/cubes/skinned characters all move together, cached
static slabs never invalidate (they hold stage-independent world transforms)
and per-instance cost is zero. The view matrix cannot carry this: in XR the
platform overwrites camera_view with the runtime's eye matrices every frame
(openxr_opengl.rs:20-25), discarding anything the app wrote.
- MrDiorama: world scaled onto an anchored slab, sky/fog/terrain-horizon
suppressed (the room IS the environment), shadow-catcher quad so it looks
planted. VrFullScale: 1:1, environment intact. Flat: unchanged
- stage_invariants.rs proves the design point: two worlds run 120 ticks, one
switching flat->MR->VR->MR(new anchor)->flat mid-run, hashes bit-identical
at every switch, with an assert_ne against a fresh world so the equality
isn't vacuous. The stage has no API by which it could reach the world
- Stereo needed NO work, and rendering twice would have been wrong: XR uses
single-pass GL_OVR_multiview2, so draw_pass.camera_view compiles to an
indexed [VIEW_ID] lookup and the GPU rasterizes both eyes from one
encoding. The existing draw path is already stereo-correct
- XR input maps controllers/pinch onto the same per-player InputState the
net layer sends — an XR player is just another player to the sim
- Quest APK packages (arm64-v8a, passthrough/handtracking/anchor/colocation
manifest bits) with no cargo_makepad changes
- Settings panel: provider/model pickers, masked key entry, pair button
showing LAN URL + confirm code, active-tier status line
- Authoring inbox: keyless clients' Intent::Authoring queued for the host's
agent, bounded and refusing with a reason — every entry eventually costs a
paid call and it is filled by peers who hold the lobby key but are not
trusted terminals
Pre-existing startup crash fixed: run_tick held a RefMut across
self.world.borrow() (`let _ = w` drops the reborrow, not the RefMut it came
from), aborting the demo path. Reproduced at HEAD with changes stashed.
Not wired, needs a device: passthrough is requested via
StageMode::wants_passthrough() but not handed to xr_passthrough (arcade has
no XR root widget yet); env-depth occlusion has a hook and no consumer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ports the 31 verbs game_script lacked — part, move_part, beam, attach,
detach, speed_mult, raycast, overlap_sphere, ground_peak, held, pressed,
axis, player_input, the cam_* readers and set_cam_* writers, save, load,
tone/tone_set/tone_stop, format, api, reset — verbatim, including the
details that bite: part's defaults and half-floor, move_part's only-given-
keys rule plus its leaves-the-static-slab redraw, attach's vec3-or-options
overload (vec3 parses first; the options path defaults to (0,1,0), not the
previous offset) with velocity zeroed, every documented clamp, raycast's
terrain-reports-as-minus-one convention, and save's strings-before-numeric
ordering (the numeric cast NaNs strings).
Audio stays host-installed: AudioRequest gained Tone/ToneSet/ToneStop/
StopAllTones with a ToneWave mirroring synth::Wave's parse fallbacks. tone()
must return an id synchronously, which a drained queue cannot do, so ids are
minted on Ctx and the host maps them to its own voices — script only ever
holds an opaque handle, so this is observationally identical. save/load
needed no hook at all: save_data lives on GameWorld and flushing was always
the host's job.
the_verb_surface_matches_gamemakers pins the count at 102. (Gamemaker's 98
match arms are 102 names — four are || aliases.) This unblocks migrating
gamemaker off its duplicate binding layer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
from_tcp_stream wrapped the socket in a BufReader that was dropped on
return, so body bytes read ahead into it vanished and handle_post blocked
forever on bytes that no longer existed — one wedged thread per request
that sent headers and body in the same TCP segment. Browsers split the two,
which is why nothing noticed. The function now owns its buffer, reads to
\r\n\r\n, and returns the remainder alongside the headers for handle_post
to consume first. The websocket upgrade path had the identical exposure (a
frame pipelined with the upgrade was silently dropped) and consumes the
same prefix now; EOF mid-head returns instead of spinning to the 4096-line
guard. New tests cover headers+body in ONE write (the case that hung, with
the connection held open afterwards so a regression blocks rather than
passing on EOF), the split case, and a plain GET.
game_net: Intent::Authoring{text} + MAX_AUTHORING_TEXT so a keyless client
in a hosted room can route a creation request to the host's agent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- libs/game/script (new): the table-driven game.* binding layer game.md
called for — 71 verbs in a HashMap built once per isolate, replacing
gamemaker's 84-arm linear chain. spawn_entity/spawn_terrain ported
verbatim so fixture terrain matches bit-for-bit; generation-tagged
callback slots; streaming eval + hot reload. Rollback snapshot is
GameWorld::clone() — M1a made the world Clone, and a clone cannot forget
a field, which is the exact bug class that put next_id in M0r's fix list.
(Gamemaker still runs its own copy; migrating it is a follow-up.)
- Capability tiers: Voice (VAD + Whisper + local judge) -> VoiceUnfiltered
(push-to-talk, every utterance costs a call) -> Chatbox (typing is the
gate). Text box in every tier, mic only above Chatbox. The chain sits
behind the local-llm feature so Quest/mobile never link a backend they
don't have
- Librarian: the local model is an optional override (None = no opinion),
never a gate — a flaky or absent model can only sharpen a decision.
Beneath it, deterministic word-overlap matching for load-by-description,
restart, and manifest-clamped knob writes. Creative requests deliberately
do NOT match an existing game (tested). Locally-answerable utterances are
dropped before they reach the cloud
- /pair: self-contained page (no external URLs, asserted), 4-digit confirm
code so a room of headsets can't take the wrong key, 0600 config-dir
storage documented as NOT a keystore — Android/iOS must move to the
platform one before shipping. Key never enters a log, package, or error
- Racing fixture evals through the new dispatch (28 entities) and renders;
hot-reload rollback verified live — a bad verb reports with a suggestion
and keeps the last good world
Found (platform, unfixed — out of this task's scope):
HttpServerHeaders::from_tcp_stream buffers past the headers into a
BufReader, so a body arriving in the same TCP segment is swallowed and
handle_post blocks forever on a body that's already gone. Browsers split
the two, which is why nothing noticed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A room of devices now plays one game: host simulates, clients send input
and render replicated truth.
- libs/game/sim/player.rs: Players roster on GameWorld, slot 0 is always
this device. Player 0's input stays in the world's original held/pressed/
pad/cam_yaw fields and is mirrored into the roster, which is what keeps
single-player numerics bit-identical (tape gate confirms). Ids are never
reused, so a stale reference resolves to None, never to somebody else
- The camera-movement knot resolved: world.player_move(p) rotates that
player's axes by THEIR cam_yaw, carried in their input packet. Player 0's
branch is the original expression character-for-character (the f32 cos
widened to f64 kept deliberately — tidying it would move the numbers)
- libs/game/session/replication.rs: Shared = pos/vel/size/kind/tag;
Derived = facing/anim/scale/glow/blob shadows, recomputed client-side and
costing zero wire bytes; Local = camera/audio/effects. Statics never enter
the per-tick stream
- Protocol (additive, version unchanged): EntityDesc + Descriptors message
splits rare reliable construction data from volatile unreliable state —
without it a joiner sees poses for entities it cannot build. EntityDesc
carries pos because statics never appear in the state stream (the
late-joiner test caught ground arriving at the origin)
- Script: game.players/player_name/player_entity/player_input/bot/on_join/
on_leave; blocks gained owner: PlayerId so a car reads its own driver
- Arcade: ARCADE_HOST=1 / ARCADE_JOIN=<addr>; clients skip world
construction and don't simulate
- Racing wire volume, 6 players x 60Hz x 200 entities: 2400 pps,
20.9 Mbit/s up (audit projected 74 Mbit for the XR stack). Asserted in
racing_scenario_wire_volume_fits_a_living_room
Two more not-a-playable-default bugs, same class as M1b's rng-at-zero:
Entity::default() leaves gravity_scale 0 (weightless wheels) and ground
without friction gives no traction — only the DSL path filled these in.
255 tests green; tape probe BYTE_IDENTICAL; xr/arcade/gamemaker build.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The thesis proof: examples/gamemaker/resources/fixtures/racing.splash is a
complete playable racing game in 72 lines (12-corner oval from a waypoint
loop, 4 cars, gates, standings, restart) with no physics, no AI and no lap
bookkeeping in script. Blocks run engine-side at 60Hz: Blocks::pre_step
(intent -> motion) before step_world, post_step after; Blocks is Clone and
snapshots beside GameWorld so a failed eval rolls both back together.
- game.car (4 suspension raycasts on a box3d rigid chassis), game.character
(drives the existing mover sweep + owns idle/walk/run blending),
game.plane; game.drive/autodrive/speed
- Brains: game.wander/chase/patrol/caught — the fixture's hand-rolled AI,
absorbed engine-side
- Race kit: spawnpoint, checkpoint, place, race, standings, lap/rank/
finished, score/score_of (Shared-tier data, ready for replication)
raycast_vehicle audit (defects documented in car.rs, still live in xr):
libm sin/cos in steering (unreplicable), a wrong side-impulse denominator
(iaj.dot(iaj) where it should be (I^-1 aJ).aJ), and an unguarded division.
Kept the structure, replaced Bullet's friction solver with an arcade force
model: suspension acts at the contact point, grip and drive through the
centre of mass, steering as yaw torque — no lateral force can generate
roll, so it is stable by construction rather than by roll_influence fudge.
Engine bug fixed at the source: GameWorld::new() left rng at 0 and
xorshift64* is a fixed point at zero, so rand() returned 0 forever for any
world built through the sim API. reset_content seeds on every eval, which
hid it from gamemaker entirely; found by a wander brain that never left home.
Tape probe byte-identical; racing fixture evals clean and drives (AI follow
the line, gates bank in order).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Admin