Two things, both found by pulling on the handcart's thread.
THE LADDER. I said three models had parts floating a few millimetres. One did.
The other two readings were my own measurement error and worth writing down,
because it is the same error as the vertex-sampling ones: I measured each part
against the model's BOUNDING BOX minimum, and a part deliberately bedded into
the ground drags that datum below the floor, so every correctly placed part
beside it appears to float by exactly the bedding depth. The forge's base
stones settle 6.8 mm and the well's drystone kerb 9.3 mm, both on purpose, and
the bellows and the well frame sit exactly on the floor their own source
declares. Nothing wrong with either.
The post mill's tail ladder was real: 14.5 mm clear. It already carried two
analytic corrections for the same corner, both correct about what they
describe, and neither accounting for the chamfer — the corner they compute the
drop of has been cut off, so the residual was very nearly the chamfer itself.
Rather than add a third derivation to a spot that has now been got wrong three
times, the assembled ladder is measured and put on the floor. A bounding box
cannot be wrong about this the way arithmetic can, and it stays right when the
chamfer or the rail section changes.
`restsOn` now covers the mill's trestle and ladder too, and its doc comment
carries the bedding caveat, so the next person to reach for it knows the datum
is the bounding box and knows what that costs.
THE ANIMATED COLUMN. `catalog-table.ts` decided it with `built.action`, which
is the VIEWER's hand-registered button and exists for five models. So the
generated table published the grindstone, the stone well, the post mill, the
cauldron, the forge and the handcart as static — six of the eleven models that
carry actions, including three of the most animated things in the kit.
The obvious replacement is worse. `core/kit.ts` wraps update as
`update: (dt) => options.update?.(dt, runtime)`, so it is a function on every
model whether or not that model does anything with it, and switching to it
marked all thirty-seven animated. One flag too narrow, its replacement too
broad.
So the generator now runs each model — a second of frames, then its action if
it has one — and asks whether anything moved. That catches what moves by
itself. What moves ON DEMAND it cannot catch, because the motion is behind an
interface the catalogue never exposed, so `Entry` now carries `actionNames`
from the model's own actions. Eleven models, and both documents list them.
Worth keeping: this table is generated precisely so it cannot go stale, it says
so in its own prose, and it was wrong for months. Generated is not the same as
correct — it only moves the mistake into the generator, where nobody looks.
Also in the npm README, which is the page npmjs.com renders: `vibe3d init` is
now the first line rather than unmentioned, `@types/three` is called out (three
ships no declarations, so without it the installed source does not typecheck),
the `@/` alias and `allowImportingTsExtensions` have a Project setup block, and
the one relative link is gone — npm resolves those against the repository, and
this one pointed at a directory the package deliberately does not ship.
`homepage` now points at the live site instead of a repository that is private.
Reported as "the wheels are in the wrong place, it is floating", and it was
three separate faults in one wheel.
The one that showed: the part declares `origin` at the axle so `setRoll` turns
the wheels in place, and a part with an origin has its geometry written
RELATIVE to that point — the anchor supplies the rest. The geometry was ALSO
translated by the axle height and position, so both were applied twice. The
wheels sat 304 mm off the ground at twice the axle height, and the cart stood
on its two shaft tips.
The one hiding behind it: the felloe segments were turned into place and then
moved out to the rim. That looks equivalent to moving out and then turning and
it is not — `rotateZ(a)` sends local +X to (cos a, sin a) and local +Y to
(-sin a, cos a), so a piece placed at (cos a, sin a) has its ARC pointing
radially. Every rim segment was a stubby block sticking 89 mm past its own
felloe, straight through the iron tyre meant to bind it. `cart-wheel` has had
the correct order all along, with a comment saying why; this model did not
follow it.
And the hub: a lathe is built around Y, the wheel is drawn in the XY plane, so
the hub stood upright inside the wheel like a little drum instead of running
along the axle. Invisible behind the spokes, which is how it lasted.
Then a fourth, found while checking the fix. The ground meets a pole's SURFACE,
not its axis: the shafts ran their axis to y = 0, so the square-cut end of an
inclined pole buried its low corner 13.7 mm, and the cart rested on two buried
tips with its wheels still fractionally clear. The axis now stops one tip
radius short and wheels and shafts share one floor.
`restsOn` is the check that would have caught all of it. The support check asks
whether every piece is connected down to the ground, and a cart standing on its
shaft tips answers yes: bed hangs off shafts, wheels hang off bed, nothing
floats. It reported 0 failing cases throughout. Connectivity cannot tell a
wheel from a decoration, and the renderer cannot either — the contact shadow is
cast from the model's own lowest point, so the shadow looked right the whole
time.
What the geometry cannot know, the author does. Naming the parts that carry the
weight turns "it rests on its wheels" into arithmetic, at 1 mm, which is tight
enough to have caught the 13.7 mm as well as the 304 mm. Declared on the cart,
the trestle table, the bench and the market stall.
`vibe3d add @medieval-kit` resolves the namespace through `models.json` and
stops dead without it:
Registry @medieval-kit is not configured in models.json
The registry README printed the two install commands FIRST and the
`models.json` entry underneath, introduced with the word "first". A reader
copies the top block, runs it, gets that error, and leaves. `vibe3d init`
seeds only `@scifi-kit`, so this is every new reader, not an edge case.
Config block first now, commands after it.
The cover card said `bunx vibe3d add @medieval-kit/wooden-chest`. Both address
forms are valid, but the bare namespace resolves through the registry's
`defaultItem` and installs the whole kit, which is what someone looking at a
card of the whole kit is being offered. A card has room for a command and not
for the `models.json` entry that command needs, so the shorter promise is the
honest one — and naming one arbitrary model implied a precision the card cannot
deliver.
Verified against the CLI rather than inferred from it:
add @medieval-kit/registry -> Registry @medieval-kit has no item named registry
add @medieval-kit -> Would install @medieval-kit · 46 files
add @medieval-kit/grindstone -> Would install @medieval-kit/grindstone · 10 files
The card a shared link unfurls to was a picture of the models with no title on
it, and the models on it were too small to tell apart. Both are fixed, but the
part worth writing down is how.
The first attempt was an HTML page composed in the site's own fonts and
screenshotted. It looked right and it was wrong: the model count on it becomes
a picture of a number the moment the shutter closes, and `check-docs.ts`
cannot read a picture. That is the exact failure that check exists for — the
READMEs claimed 27 models when there were 37 — and a card is the worse place
to have it, because nobody reading one can see the kit to know it is stale.
So `bun run cover:build` counts the catalogue and draws the numbers it counted.
`site:build` runs it, which means the published card cannot be older than the
models it shows, and the day a model is added the card says 38 without anyone
remembering to make it.
Drawing type needed type. `scripts/text.ts` reads the outlines out of a `.ttf`
and fills them with a scanline pass — four sub-scanlines a row with exact
horizontal coverage, which is more code than sampling a 4x4 grid and visibly
better at the 16px the two mono lines are set in. It is deliberately partial:
format 4 cmap, simple and composite outlines, advances, no kerning or shaping.
What it cannot draw it throws on, because a headline missing a letter is not
something you notice before it is published. Archivo and IBM Plex Mono come
with it under the OFL, the same two faces the site is set in.
The renderer moved to `scripts/raster.ts` so the card and the contact sheet
draw from the same code and cannot disagree about what a barrel looks like;
`render.ts` is now the command line over it, unchanged in what it does. Its
one new flag is `--ground`, a flat background instead of the per-cell
gradient: forty gradients side by side read as banding, and on the page's own
ink the models float on the card as one field instead of sitting on it as a
grey slab.
Ten columns by four, because 1.91:1 and 37 models want 40 cells and at that
shape the cells come out square, which is what lets a barrel look like a barrel
at 116 pixels. The last row is centred — three empty cells left flush is a
notch in the corner, and a short row centred is how a contact sheet ends.
Byte-identical across runs, so it does not churn the diff on every build.
A link to the site showed a URL and a blank square. It now carries a title, a
description, a canonical URL, an icon and a full set of OpenGraph and Twitter
tags, so a page nobody has opened yet arrives as something you can look at.
Two of those tags have to be ABSOLUTE and it is not a style question: a
scraper fetching the page has no base to resolve `/cover.png` against and
drops it silently. `build-site.ts` now copies the images into the output as
well, because an `og:image` pointing at this host has to have the file ON this
host — a card that 404s shows nothing, and nothing about the page itself looks
wrong while it happens.
The cover is a real 1200x630 render rather than a crop of a square one, which
took a change to the renderer worth having anyway. `newFrame` and
`frameCamera` both took a width and a height already; only `renderOne` was
insisting they be equal. And with a non-square frame the old camera fit showed
its flaw: it fitted the bounding SPHERE, which is one number and wrong for
anything not roughly round. The kit scene is wide and flat, so its sphere is
far taller than the scene is, and fitting that into the vertical field left the
models in a band across the middle with empty air above and below — most of a
1.91:1 card. It now fits the BOX, iteratively, because perspective is not
linear in distance: place the camera, project the eight corners, move by how
far the worst one lands past the edge. Two passes is inside a pixel.
That change improves every render, not just the card: on the contact sheet
every model is visibly larger in its cell and nothing crops.
`check-docs` now watches `viewer.html` too. Its description quotes the model
count and the triangle total, which is the same sentence that went stale three
times in the READMEs before that check existed — and a stale number in a card
is worse than one in a README, because nobody reading the card can see the kit
to know it is wrong. The description says "37" rather than "thirty-seven" for
exactly that reason: a word is not a number the check can find.
Also repointing the npm README's two images from `raw.githubusercontent.com`
to the live site. They were pointing into a private repository and would have
404'd on npmjs.com; now they do not depend on the repository being readable at
all.
Five attempts and it was never good. The last one is better than the cabbage
that prompted the complaint, and "better than the version you called ugly" is
not a standard worth shipping against — every other model here would have to
carry it.
The constraint underneath is real and worth recording, because it is not a
tuning problem and no sixth attempt was going to find it. The kit's budget is
2500 triangles a model. The oak wanted a bole you can see, limbs radiating out
of it, and a dome of foliage over the lot; at 2500 that is roughly 128 for the
trunk, 880 for the boughs and 1400 for the crown, and 1400 buys about forty
masses. Forty masses hung on twelve limbs is twelve clusters, and twelve
clusters do not make a dome — from a third of the compass you see the gaps
between them. A turntable says so plainly where a single three-quarter view
does not, which is the other thing to keep: I judged this model on one angle
for four rounds.
Everything else in this kit is something a person made, with flat faces and
straight edges, and lowpoly is not a compromise for those — it is close to how
they are actually built. A tree is the one subject where the budget genuinely
bites, and the honest options were to except it from the budget, to abandon
realism for a deliberately stylised shape, or to drop it. Dropped.
Out with it goes the `leaf` slot and its palette key, since the oak was the
only thing using either. Back to thirteen slots. `bark` stays, because the log
pile uses it; `limestone` stays, because the trough and the grindstone do.
`scripts/probe-tree.ts` goes too — it measured one model against one pair of
photographs and there is no longer a model.
37 models, 30 390 triangles, thirteen material slots, all checks pass, 0
floating cases, GLB round-trip clean, Conformant, documentation in step.
"Very ugly" was the report and it was right. Put beside its reference the
diagnosis is not subtle: the crown read as crumpled paper, the colour was flat,
and the trunk had been swallowed whole.
The correction that mattered was to a rule I had invented myself. I decided
early that "a crown is one piece" and made the probe fail anything else, so
every version since had been tuned until the flood fill returned 1 — which is
the instruction "be a single smooth blob", and that is exactly what it kept
producing. The support rule requires the crown to be connected to the TREE, not
to itself: a clump that swallows a branch end is held up by that branch
whatever its neighbours do. Letting the masses separate is what buys the gaps,
and the gaps are the silhouette. The probe now reports the piece count as
information instead of grading it.
Four other things, each of which cost a render to see:
- The clumps were three rings on five sides — a coarse bipyramid — and
sixty-five of them creased every facet against its neighbour. Five on eight
fixed the surface and made a cabbage. Four on six is the middle, and the
middle is where it belongs.
- `roughenGeometry` came off the crown entirely. Ragged edges read as foliage
where smooth ones read as stone: true of a SILHOUETTE, false of a surface.
All it did was crumple the skin. Raggedness has to come from the
arrangement.
- The tone between clumps varied by +/-0.09, so every mass was the same green
and the crown was one object with facet shading on it. Widening it to
+/-0.19 put seven whole clumps through the floor into black, because the
tint jitter, the mottle and the baked occlusion all pull the same way
underneath. It is asymmetric now, -0.03 to +0.23, which is also truer: a
lit mass is far brighter than the base green while a shaded one is filled
in by the sky.
- Every clump sat on the outer half of its limb, making a crown that is a
RING — full at the rim, hollow in the middle, and from a third of the
compass you look straight through the hole. I had that right once and
removed it while cutting triangles, which is the cheapest regression to
make and the hardest to notice: nothing failed, the model just got worse.
The bole went from 0.037 of the height to 0.046. The reference's own number
came off a photograph of a tree seen whole from a distance; at this size, under
a crown, 0.037 disappears and what is left is a cabbage on a stick.
Twelve limbs instead of nine, spread more evenly, with cheaper geometry paying
for them. And the turntable is now the thing I judge it by rather than a single
three-quarter view — one good angle hid two that were full of daylight.
Honest state: better than it was and not yet good. Three of six angles read as
a tree; the others are still bunched. 38 models, 32 798 triangles, all checks
pass, 0 floating cases.
Reported for the third time, and the first two answers were wrong in the same
way: I treated "it should not be visible" as a number to reduce rather than a
rule to satisfy. The bench's through-tenons went 42 mm proud, then 3 mm, and
3 mm is still a raised patch with a lit edge on it — visible in the render,
and the thing being complained about. The market stall then reproduced it
exactly: four post heads through the awning, like tacks.
Both were pushed proud DELIBERATELY, to keep two faces out of one plane and
away from the z-fight check. That trade is false, and this is the sentence
worth keeping: a piece that must not show ends INSIDE the solid it enters. A
buried face cannot be coplanar with a surface and cannot be seen. Level with
the surface is the only genuinely bad option, and it was the one I kept
treating as the alternative.
- The bench's tenon is now housed an eighth of the seat's thickness below
the top. That is a real joint rather than a compromise: a stub tenon into
the underside of a board is how a bench is made when the top is thick
enough to take it, which at 0.135 of the height it is. What is lost is the
panel of end grain; what is gained is a seat.
- The stall's posts stop ON the cloth's mid surface, with half its thickness
over them, and the cloth went from 0.2 of a board to 0.42 so there is
something to bury them in. The heads now sit 25 mm under the top surface.
And the brief is a check now, because a rule I keep is worth less than a rule
the build keeps. `nothingAbove` names the part nothing else may rise above —
the seat, the counter, the awning — and it reports the offender and the
millimetres. Mutation-tested: pushing the tenon back to 3 mm proud fails it
with "PROUD: legs by 3.0mm".
38 models, all checks pass, 0 floating cases, documentation in step.
Two corrections, and the first was a bug I had just written.
`vercel.json` served `/viewer.js` and `/viewer.css` with a year of immutable
caching, and those names carry no content hash — they are pinned in the vite
config. Immutable caching on a fixed name means a returning visitor is handed
the bundle they already have, for a year, no matter how many times the site is
rebuilt. The rule and the filenames were written a few minutes apart and
contradicted each other.
So the bundles are hashed now, which makes the caching correct rather than
wrong, and the cost is exactly the thing the rule creates: nothing downstream
may name them literally. `build-artifact.ts` finds them by pattern. So does
`build-site.ts` — whose own first version listed `viewer.js` by name in its
size report, and fell over on the very build that introduced the hash. The
mistake and its consequence landed in the same commit.
The cache rules now live in a `_headers` file written into the output, where
the host looks for them: a year on the hashed bundles, and revalidation on the
HTML, which must NOT be cached that way — it is what points at the current
hash, and a stale copy of it pins a visitor to a build that no longer exists.
Second: the README had grown a hosting comparison — which providers are free
with a private repository, what GitHub Pages costs. None of that belongs in a
README. It is a public-facing document about a model registry, read by people
deciding whether to install it, and my deployment arithmetic is neither their
business nor true for their situation. What stays is the part that is about
this project: the command, what the output is, and that nothing is fetched at
run time. `vercel.json` goes with it, since it was config for a host this
project does not use.
`bun run site:build` and the output is a site: the viewer at a domain root,
with an `index.html` a host can answer `/` with. The viewer build already
emitted everything else — `viewer.js` and `viewer.css` loaded by absolute path,
which is correct at a root and wrong nowhere — but it emitted no index, so a
static host serving the directory answered `/` with a 404 and the site would
have looked broken while every file on it was fine.
The landing page is the VIEWER rather than the demo app, which is the one
decision in the script and the reason it is a script: someone arriving at the
address should be able to turn a model over and press play, not read a page
about a model they cannot see.
Verified the way a host would see it rather than through the dev server: built,
served from the directory as a root, and every one of the page's own requests
came back 200 — `/`, `/viewer.js`, `/viewer.css` — with a clean console and all
38 models in the rail. Nothing is fetched at run time, either; the registry
address in the sidebar is a label, not a request, so there is no API to stand
up.
`vercel.json` carries the build command, the output directory and a year of
immutable caching on the two hashed assets. Cloudflare Pages takes the same
two settings in its dashboard.
GitHub Pages is in the README's table with a "no" against it, because it is the
obvious guess and the answer costs money: on the free plan Pages serves public
repositories only. Worth stating where someone will look for it rather than
leaving them to find out.
And one line that is a decision rather than a fact, so it is written down:
publishing the viewer publishes the models. The bundle carries every builder in
it, minified. That is a different thing from the npm package — which ships the
source AS source, comments and all — but it is not nothing.
The piece the kit had been assembled around without having. The basket, the
vegetables, the sack, the crate and the coin pouch all belong on one, and a
market with none of it is a row of props on a floor. It is also the only thing
here with a roof.
The awning SAGS, and that is the whole model. A flat sheet on four posts is a
table with a lid; cloth slung between four points hangs, and the dip is what
says it is cloth before the colour does. It costs nothing:
`dishedSheetGeometry` bends its cross section by `curve`, and one turn about X
maps that bend onto the world's vertical, so the sag is the helper's own
parameter rather than anything hand-built. Per level, because the reference's
front edge dips further than its back — cloth pulled over a ridge and left
loose at the eaves does not hang symmetrically.
Three faults, each of them a thing I had drawn correctly and built wrong:
- The posts were one height, and the cloth is not one height. It is a
parabola across its length and deeper at the front than the back, so a
single post reached some corners and not others: the awning came away as
its own floating piece 67 mm above the post heads, held up in the drawing
and by nothing in the geometry. Each post is now sized by the sheet's OWN
arithmetic — same parabola, same per-level sag, same quarter turn, same
tilt — rather than by a second version of it that can drift.
- The legs ran up to the rail's top face and ended level with it. Two upward
faces in the plane y = 0.745, at all four corners. A leg is housed in a
rail, not flush with it.
- The posts stopped at the rail's underside, which put their feet in ITS
plane — the same fault one storey down. They now run past it and finish as
a stub below, which is the reference's detail as well as the fix.
And one thing the render corrected rather than the checks: the posts were
inboard, so the counter's boards had posts driven through them a hand's
breadth from each end. That is not how a stall is built and not what the
reference shows — the posts stand at the ends and the top lies between them.
The sag came down from 0.17 of the depth to 0.10 for the same reason: at 0.17
the cloth dropped a tenth of its own span and read as a hammock.
38 models, 32 700 triangles, all checks pass, 0 floating cases, GLB round-trip
clean, Conformant, documentation in step — including the showcase arithmetic,
which every new model moves.
The kit had an anvil, a forge and four edged tools, and nothing that put an
edge on any of them. This is the machine that did, and the only piece of
village equipment here with a working part in it — so it turns: crank it and it
runs for three revolutions over five and a half seconds, which is a heavy stone
on a wooden bearing rather than a flywheel.
The wheel is `limestone` and that was measured, not assumed: the reference
reads hue 32.5, saturation 0.196, lightness 0.511 against the trough's 36.6,
0.161, 0.500. A sandstone wheel and a weathered trough are the same rock as far
as this palette is concerned, and neither is the grey `stone` of a dressed
wall. The frame is `oak` taken well down in value — the reference's timber is
0.25 against the palette's 0.40 — but it stays oak, because a third wood key
for one model would be describing the weather rather than the material.
Three faults, all found by the checks and all of them things I had reasoned my
way past:
- The axle and its washers were built, merged into a local array, and then
never returned. The wheel hung on nothing and the crank floated beside the
machine at 0.60 m. An array you forget to use is not a type error.
- The axle sat 48 mm ABOVE the rails and touched them nowhere. The reference
answers this itself and I had left the answer out: there are bearing
blocks bolted on top of each rail with the axle running through them.
- The low stretcher ran down the centre line from end to end and touched
nothing on the way, because at that height the legs have splayed out to
z = 0.21 and it was at z = 0. It tied the trestles together in the drawing
and not in the geometry. There are now two rails, one down each side, and
where the legs are at that height is worked out from the strut's own two
ends so it stays true when the splay is changed.
The last two shared a symptom — the support check reporting two components,
one on the ground and one containing everything above it — and finding them
took a detour worth writing down. Sampling the frame's VERTICES to see where
the legs were at rail height found no legs at all, because a tapered box has
corners only at its ends. That is the third time in this project the same
mistake has been made with the same consequence: a measurement that returns a
confident number about nothing. Slicing the triangles instead put the leg's
cross section exactly where the arithmetic said it was, which is what proved
the gap was somewhere else.
What is NOT here, said plainly: the reference has a treadle and a linkage down
to a foot board. This has a hand crank instead. Both are period; the treadle is
three more moving parts to serve an animation the crank already carries.
37 models, 32 316 triangles, all checks pass, 0 floating cases at 224 voxels as
well as at 112, GLB round-trip clean, Conformant, documentation in step.
The kit could water nobody. It had a well, a bucket and a cauldron, and all
three are things you carry water IN; a trough is the thing a village keeps
water in, standing in the street outside the smithy.
Two additions to the core, both earned by measurement rather than by wanting
them:
- `limestone`, a palette key. `stone` was measured off dressed and rubble
masonry and is grey, because a wall is quarried and laid face out. One
block left in the open for a century reads hue 36 and saturation 0.16
against the wall's 44 and 0.05 — warmer, three times more saturated,
lighter. Both are right and neither replaces the other. The rendered block
now matches its reference to within a fifth of a degree of hue and 0.004
of lightness.
- `water`, a slot, and it is far less green than instinct says. Sampled
inside the reference's basin the film reads hue 39.5 — which is the hue of
the stone UNDERNEATH it, not a colour of its own. It darkens and slightly
saturates whatever it lies on and does that through the material's
opacity. Painted the pond green I was about to reach for, the composite
would have missed the reference by a hundred degrees.
Making it read as CARVED rather than sawn took three goes, and the two failures
are opposite errors worth keeping:
1. Each wall in two tiers, to give `roughenGeometry` more corners to move.
The tiers' corners sit at different positions, the hash moved them by
different amounts, and the render showed a lid on a base with a hard seam
right round it.
2. One box per wall, roughened harder to compensate. A box has eight corners
and every face is two triangles, so moving those corners folded each wall
along its own diagonal. Not stone: creased cardboard.
Both were trying to get subdivision out of the wrong thing. The walls are
`dishedSheetGeometry`, whose own docstring says the trough wants it: one
continuous body with levels along its length, so the rim undulates because
each level carries its own height, the wall thickens and thins because each
carries its own thickness, and there is no join for a seam to open in.
The z-fight checker earned its keep twice on the same plane, `0,-1,0 | 0`.
Four walls standing on the ground all have a face down there, and where the
corners overlap — which is where they are MEANT to overlap — two of those
faces share it. Lifting the basin floor only moved the pairing from
floor-against-wall to wall-against-wall. The answer was to stop there being
four things down there at all: a trough is solid below the water line, so it
is one block from the ground to the basin with the walls sunk into it, by
different amounts, because two buried faces in one plane are still two faces
in one plane.
Last, two defaults that the render corrected rather than the reference. The
walls at 0.27 of the width shut the basin so far that a three-quarter view
showed only its inner face, against the reference's opening of about three
fifths — now 0.21. And the water at 0.55 of the basin's depth sits below the
near rim and the trough reads as dry from every angle anyone will look at it
from; the reference stands a couple of centimetres under the rim, which is
0.85.
36 models, 31 748 triangles, fourteen material slots, all checks pass, 0
floating cases, GLB round-trip clean, Conformant, documentation in step.
Reported twice. The first time I planed the through-tenons from 42 mm proud to
3 and called it done, which answered the tenons rather than the question — the
tenons were never it. Measured again now: nothing stands more than 3.9 mm above
the seat at any slider setting, and the thing that was catching was on the
LEGS.
The leg board widens on the way down; that is the splay. The two feet, cut out
of its bottom edge by the arch, were built at `footWidth` — the board's width
at the FLOOR. But a foot starts at the top of the arch, a third of the way up,
where the board has not widened that far yet. So every foot stood proud of the
board it hangs from, on both sides, as a hard shoulder at shin height: 14 mm by
default, 20 mm on a wide bench, 32 mm at full splay. Setting the splay to zero
made it vanish, which is what finally named the cause.
The leg's outer edge is a straight line from `footWidth / 2` at the floor to
`legWidth / 2` under the seat, and a foot's outer face has to lie ON that line
rather than run vertically up from its own base. A tapered box narrows about
its centre, so taking twice the taper off the top width moves the outer face in
by exactly the taper — and moves the inner face out by the same, narrowing the
arch as it rises, which is also what a splayed leg does. 14.2 mm becomes 0.2,
32.3 becomes 0.4, and 0.2 mm is the sampling resolution rather than a step.
`scripts/probe-bench.ts` is how it was found and it took two wrong versions to
get there, both failing the same way — by measuring something and reporting a
number that was not the answer:
- The first sampled vertices into height bands. A tapered box has corners
only at its top and bottom, so the bands between them came out empty, were
skipped, and it reported a confident 0.0 mm step for a leg with a 17 mm
shoulder on it. It now slices the actual triangles, which gives a width at
any height whether a corner is there or not.
- The second scanned the full height of the leg part and found its biggest
drop at 86% — exactly where the board ends and only the tenon carries on
into the seat. That is what a tenon IS. A real fault of 14 mm was sitting
underneath a 66 mm reading that was not a fault at all, so the scan now
stops at the seat's underside.
Both media images regenerated, since the bench is in each of them.
The image both READMEs open with was a contact sheet whose FIRST cell — top
left, the one everybody's eye lands on — was the `kit` entry: all thirty-five
models arranged in one scene, squeezed into a 300 px cell, an illegible smear.
It is not a model and it does not belong in a grid of models. Rendered on its
own it is by far the better picture anyway, so it is now the picture, and the
contact sheet is a second image beneath it.
Rendering it on its own is also what showed the arrangement had quietly gone
wrong. The row width was a fixed 5.6 m, which was right for a kit of tankards
and barrels and became wrong the moment anything bigger than a row went in: the
oak is 8.6 m across and the mill 6.6, so each took a row to itself while the
small props packed four and five deep. The kit had stretched into a long thin
diagonal, and since a camera fits a bounding SPHERE, the sphere around a
diagonal line is enormous next to the things inside it — the kit occupied about
a quarter of its own picture and the rest was empty floor.
So the row width is measured from the models now rather than fixed: the square
root of the total footprint, which packs rows about as deep as they are wide
whatever the kit happens to contain, floored at the widest single model since
no row can be narrower than the thing standing in it. Swept for the factor:
1.15 gave 15.1 x 21.7 m, 1.45 gave 19.2 x 13.3 — the square is between them,
and 1.24 lands at 16.4 x 14.8, aspect 1.11, filling the framed disc better than
any of them.
It reads as a village yard now: mill and oak at the back, forge and cart in the
middle, the small props at the front. Which is what the entry is FOR — every
model on one ground plane at one scale is the only place a difference in scale
or tone has nowhere to hide.
REFERENCE.md was 751 lines long and said the kit had 27 models, 15,425
triangles and eleven material slots. It had 35, 31,404 and thirteen, and the
document named none of the last six models at all. A reference nobody can trust
the numbers in is worse than no reference, because a reader has no way to tell
which parts are still true.
This is the third time the same failure has happened here. It is why
`catalog-table.ts` exists — the reference once said "four items" when the kit
had twenty-three — but regenerating the TABLE never touched the prose around
it, and the prose kept its own copies of every number.
So the counts are now checked the way the geometry is: against the models
themselves, by a script that fails. `check-docs.ts` reads the built kit and
requires each of the three documents to state the right model count and
triangle total, to list every model that exists, to name none that does not,
and to get the slot count right where it is written as a word.
Mutation-tested, as everything else here is — a wrong total, a dropped model
row and a stale slot word each fail it. And that test cost me the very fix it
was testing: `git checkout REFERENCE.md` to undo a mutation threw away the
uncommitted repairs along with it. Mutation-test after committing, not before.
Two things it caught immediately:
- The root README never stated a triangle total at all, so the number a
reader most wants for judging whether a kit is affordable was in two of
the three documents and absent from the front door.
- My own first version of the slot check. It looked for any word before
"slots", matched the ordinary phrase "material slots" several paragraphs
earlier, found that "material" is not a number, and silently checked
nothing — reporting nothing rather than failing. A check that passes by
not running is worse than no check, because it reads as evidence.
`bun run check:docs`, and it is in the README's verify list with the others.
The first thing in this kit that grew rather than got made, and the first one
built entirely against reference photographs from the start rather than after
three rounds of guessing.
Two references, generated with the same Codex tool the rest of the kit now
uses: one oak in full leaf and the same tree bare. Both of them contradicted
what I was about to build, and the default lowpoly tree -- a tall pole with a
ball on it -- turns out to be wrong in every measurement that matters.
crown width / tree height 1.32 to 1.43. An open-grown oak is half again
as WIDE as it is tall. Nothing else in this kit
is wider than it is tall.
widest at 24% of the height. Not halfway up: a quarter,
just above the ground, because the lowest limbs
are the longest and leave almost horizontally.
bole runs to 72% of the height as one continuous stem. It
does not fork low into equal leaders, which is
what the leafy photograph alone suggested and
the winter one disproved.
bole width / height 0.130 at 2% of the height, 0.073 at 10%. It
nearly halves in the bottom tenth and then
barely tapers at all for a quarter of the tree.
The taper table in the model is those rows, read off the photograph one at a
time. The limbs SWEEP -- out at a shallow angle, turning up at the end -- which
comes free from bending each limb along its own axis before rotating it into
place, and which cost a correction I had not expected: bending a limb swings
its whole chord up by half the arc, so every limb was arriving 0.21 rad high
and the crown's widest point had climbed to 38%. Aiming low by exactly what the
bend will lift is the fix, and it is a constant.
Measuring is the whole story of this one. `scripts/probe-tree.ts` prints the
four reference numbers off the built geometry the same way they came off the
photographs, plus one a photograph cannot give: how many separate pieces the
crown is in. A crown is one piece. This one was four, then one, then two, then
one again, and the flood fill was right every time my eye was vague.
Four things it caught that looking would not have:
- Forks throwing DOWNWARD. The rise of a fork is sin(splay) * sin(spin), so
the sign of the PRODUCT decides it. I constrained the spin and alternated
the splay, which left every second fork pointing at the floor. The
measurement said so plainly while the reasoning said the opposite.
- The crown torn in two, and not where I kept looking. Locating the loose
piece -- height 10-64%, bearing 228-313, reaching 5.7 m where the main mass
stopped at 4.5 -- ended four rounds of widening clumps and evening out
bearings, none of which were the gap. The gap was the MIDDLE: pushing the
leaves outboard to open the space under the crown had taken the inside out
with it, and nine lobes hung off a bare hub. The reference's crown is full
from the bole outward; the clear space is underneath.
- Occlusion, exonerated. Turning it off entirely changed the clipping by
nothing at all, which took two minutes and saved a redesign.
- My own metric, wrong twice. It counted every dark pixel in the crown's box
and 87% of what it found was the cast shadow, so it read 10.8% clipping on
a crown that was barely clipping. And it measured the reference by row
width but our own geometry by twice its radius, which is a different
quantity on anything that is not symmetric.
Colour is measured too, and the measurement contradicted me again: I was about
to bake a top-to-bottom gradient into the crown, and sampled in three bands the
reference reads 0.362, 0.336, 0.324 -- no gradient at all. What is there is
local, a spread of 0.45 in lightness within one band. Foliage varies leaf to
leaf, not top to bottom. Chasing that 0.45 with tint, mottle and occlusion all
at once put 2.2% of the crown through the floor into pure black against the
reference's 0.19%, which is the straw trap for the third time: a statistic taken
off fine detail cannot be added on top of the machinery that already
reproduces it.
The bark is measured off the WINTER photograph after the summer one gave a
colour 28 degrees too green -- a trunk under a full crown is lit through the
leaves. Low saturation made it worse than the hue did: a nearly neutral colour
takes its hue from its light, and this renderer's blue-grey sky dragged the
rendered bole to hue 119 on a tree whose bark is brown.
Finished: hue 79.6 against 77.9, saturation 0.386 against 0.381, lightness
0.386 against 0.373, 0.00% clipped. Crown 1.325 wide for its height, in one
piece, hanging from 8% where the reference hangs from 11%.
The honest limits. What fills a real winter silhouette is thousands of twigs
and no budget here reaches them, so `leafiness: 0` is a bare tree of LIMBS --
a good dead or winter tree, not a photograph of one. And the crown still
carries more of its width high than the reference does: widest at 50% against
24%, on a reference profile that is nearly flat from 24 to 44 so the peak is
softer than the number sounds.
2310 triangles, inside the kit's 2500 budget, which cost a side off every clump
and a level off every fork -- but not one clump, because the count is what
holds the crown together. 35 models, 31 404 triangles, thirteen material slots,
all checks pass, 0 floating cases, GLB round-trip clean, Conformant.
Committing these because the code should not be lost, not because they are
ready. Both pass every check and neither looks good enough to ship, and the
catalogue table and contact sheet are deliberately NOT regenerated so the
kit does not advertise them yet.
The shield opens a category the kit did not have at all -- no arms or armour
anywhere -- and its construction is right: planks cut to the chord of the
circle rather than squared off, a rawhide rim gripping their ends, an iron
boss over the hole the hand goes through, and quarters painted on the
diagonal by giving each plank's two halves different colours, which costs
no geometry because the halves exist anyway.
The forge is the other half of a smithy that has had an anvil and no fire
since early on: a raised hearth at working height, a bed that `setLit` puts
out without taking the charcoal with it, a tapering stack over the back,
and a bellows blowing through the side.
Four faults found and fixed on the way, and three of them are the same
mistake:
A two-rotation chain put the bellows boards on the wrong axis -- thickness
along Z, length along Y -- so both boards stood on edge in nearly one plane
with every face coplanar. The lever went the same way, ending up across the
bellows instead of over it and touching the post at neither end, which the
support check called a floating bar. And the shield's boss was turned -90
degrees about X instead of +90, which pointed the dome into the back of the
shield and left only its open rim showing from the front -- the reason the
boss rendered as a flat black ring.
The fourth was subtler and took three passes to see. Every plate of the
bellows -- two boards, a gusset, two ribs -- was built from one profile, so
all five ended at exactly the same x and began at exactly the same x. I
staggered the tails and the check moved to the noses; staggering both is
what fixed it.
What is still wrong, and neither is small: the forge's fire bed reads as a
black hole with a few orange specks rather than a bed of coals, and the
bellows still reads as loose boards rather than a bag of leather. The
shield's paint comes out pastel where the reference is deep, and it leans
further than it should.
34 models, 28 728 triangles, all checks pass, 0 floating cases, build clean.
Eight kinds mixed by two sliders: how many of them appear, and which one
leads. Turnip, cabbage, onion, leek, parsnip, carrot, tomato, potato.
The first six are what a medieval European kitchen had and the last two are
New World crops that only reach Europe after 1492 -- they are in because a
kit is more useful than it is a museum, and the carrot is orange rather
than the purple a medieval one would have been. The kinds are ordered so
that the period-correct set costs no thought: `lead: 0` with `kinds: 6`
gives exactly those six and never reaches the two after them.
Two parts, `bulbs` and `roots`, because those are the two ways a vegetable
rests -- on its base or on its side -- and a consumer wanting one should not
have to take the other.
Four faults on the way in, and every one of them was about placement rather
than shape.
The header claimed each vegetable is authored with its lowest point at the
origin, and I asserted that instead of measuring it: a leek's blades lean
four times its own radius away from the axis, and a cabbage's loose leaves
hang below its base. Whichever vegetable dug in deepest then defined the
model's floor and made every other one look as though it were hovering.
Reading the bounding box and setting each one down by what it actually
measures makes the claim true by construction rather than by an argument
that has to be right about eight shapes.
The leaves were flung off their vegetables outright. They were built
already in place and then rotated, and rotation turns a body about the
ORIGIN, so each one moved away by its own distance times the sine of the
angle. Origin, rotate, place -- which is the order the hoe's neck already
has a comment about.
Tightening the heap so it read as a pile rather than a display then made
the bulbs collide, twice. Standing vegetables were never turned about their
own axis, so every turnip presented its facets at the same phase and
neighbours met in parallel faces. And where the turnip's cream body meets
its purple shoulder -- two lathes, because `colourTop` interpolates across a
whole profile and bled the purple down to the root -- both pieces were
capping an end that is buried inside the vegetable. Two turnips touching
met on those hidden discs. They are not generated any more, which is the
same saving `bandGeometry` makes on the inside of a hoop.
32 models, 26 939 triangles, all checks pass, 0 floating cases, GLB
round-trip clean, Conformant, build clean.
Two models that each fill a hole the catalogue could name.
The kit had a torch and a lantern and no HEARTH, which is the wrong way
round: a cooking fire is what everyone in a medieval scene is facing and
the only light in most of them. The cauldron hangs from a tripod over a
ring of hearth stones, with embers that `setLit` puts out. Its belly is
wider than its mouth, it has three short feet because it stands in the
ashes as often as it hangs over them, and it hangs by a bail rather than by
its rim.
And there was a `cart-wheel` with nothing to put it on -- while every
container in the catalogue, crates and sacks and barrels and baskets, is a
thing you would MOVE. The handcart's axle sits aft of the bed's centre so a
load leans on the puller rather than away from him, and its shafts leave
the bed at an angle and rest on the ground, which is what lets one person
lift them to waist height without the load shifting.
The cauldron found a fault in the mill on its way in. Both build their
splayed legs the same way -- lean about X, swing about Y, translate out
along the radius -- and the lean was POSITIVE, which carries the top the
same way the translate does. The two add, so the head ends up at the full
outboard radius and the foot at the axis: the tripod stood on its apex and
the mill's quarter-bars stood on their heads, splayed over the cross-trees
and gathered at the post, which is precisely backwards for a brace whose
job is to carry a load out to the ground. Nothing had caught it in the
mill, because a bar joins post to cross-tree either way round and the
bounding box is identical; in the cauldron it separated the tripod from the
ground and the support check finally had something to say.
Lowering the pot to where it could actually cook then broke its chain, and
for a reason this session has seen before: the link radius was derived by
dividing the drop by the link COUNT while the links are laid out over
count-1 gaps, so each link came out a little smaller than half its own gap
and the chain came apart as soon as the drop grew. The tavern sign's chain
had the same slip.
31 models, 25 221 triangles, all checks pass, 0 floating cases, GLB
round-trip clean, Conformant, build clean.
A village well: a drystone kerb, a timber frame over it, a windlass with
its collars and crank, and a rope and bucket that lower down the shaft.
`setDepth` and `setWinding` drive the bucket; its part origin is the
roller's axis, so winding is a translation and costs nothing per frame.
The kit had no masonry at all, which for a medieval catalogue is a hole
rather than an omission -- a well, a trough, a millstone, a boundary wall
and a hearth are all stone before they are anything else. `stone` is the
twelfth slot. Its colour sits close to weathered oak on purpose: what
separates the two at a glance is not hue but that stone scatters light
completely flat, and roughness is not something vertex colour can carry,
which is exactly the argument the slot table makes for keeping iron and
steel apart.
Four faults on the way in, and three of them are ones this session has seen
before in other models.
The blocks were written footprint-first as [wall, arc]. `rotateY` carries a
footprint's +Z out along the radius and its +X round the circumference, so
every block sat across the ring rather than along it -- each one sticking a
block's length out of a wall a block's width thick. The kerb came out as
rubble thrown at a circle.
The rope started at the roller's axis, which put its whole surface inside
the barrel's hollow interior touching nothing. Then it started in the eye
of the wound coil, which is also a hole. It has to begin within the coil's
material, between 1.0 and 1.34 of the roller radius, so that it crosses the
coil's outer face on the way down.
The actions set `anchor.position.y = 0`, which destroys what `origin` had
just done and dropped the whole rope and bucket by the height of the
roller. Everything the actions touch is now an offset from where the origin
put the anchor.
And the masonry was over the triangle budget at 2 784, almost 1 900 of
which were chamfers on forty-four blocks -- edges the roughen pass softens
anyway, and stone is the one material here where a sharp arris is not
wrong. Plain tapered boxes bring the model to 1 376.
Two things about the wall itself, both from looking at it. Varying each
block by a tenth in arc and thickness put every stone at its own radius and
its own width; a dry wall is irregular the way a hand-cut stone is
irregular, which is a few percent. And blocks cut to exactly their share of
the arc cannot close: a flat face on a curve leaves a wedge of daylight at
every joint, so they are cut long and bite into each other -- safe here
because each is rotated to its own angle and no two adjoining faces are
parallel.
29 models, 22 043 triangles, all checks pass, 0 floating cases, GLB
round-trip clean, Conformant, build clean.
Three defects reported from looking at the models, and the first of them
says something about the check that was supposed to catch it.
Every third coin was lifted to "land on top of the previous one" while its
position stayed independently scattered, so it was almost never above
anything and simply hovered about 3 mm off the ground. A stacking coin now
takes the previous coin's PLACE as well as its height. The tilted ones were
lifted by half their thickness plus their circumradius; a coin is a
nine-sided prism and can come to rest on a face rather than a corner, so
the lift uses the apothem and its own radius, erring towards bedding into
the ground rather than floating above it.
The support check missed all of that, and the reason is worth writing down:
its voxel is the model's extent over the resolution, so at 64 a 0.19 m coin
pouch gets a 3 mm voxel -- exactly the size of the gap it existed to find. A
check coarser than the defect it hunts is worth very little. At 112 the
whole suite still runs in fifteen seconds.
Raising it immediately turned up a real fault nobody had seen. The log
pile's sawn ends were cut to exactly the body's end radius and placed
exactly at its rim, so disc and lathe met along a seam of no width: they
touched and never overlapped, and there was nothing joining them but two
edges in the same place. The caps are now three percent wider and seated
slightly inside, which is two millimetres here and true of a real log
anyway, since bark wears back from a cut face.
And the bench's through-tenons stood 42 mm proud of the seat -- a block of
oak the size of a thumb sticking straight up out of the surface you are
meant to sit on. A real through-tenon is wedged and then planed back, so
what shows is end grain and a hairline of shoulder. Three millimetres
proud, ten housed below.
28 models, 20 547 triangles, all checks pass at resolution 112, 0 floating
cases, build clean.
The largest thing in the kit, and the one whose name is its whole design:
the entire body -- millstones, gearing, miller and all -- is balanced on one
massive vertical post and turns bodily on it, so the sails can be pointed
into whatever wind there is. The miller does that by walking the tail
ladder round, which is why the ladder is not a way up but a lever, and why
it reaches so far behind the mill.
Four parts, in the order the thing is built: `trestle` (two cross-trees,
the post standing on them, four quarter-bars), `body` (the boarded buck and
its pitched roof), `sails` (a windshaft leaving the front gable, with its
own origin so `setTurning` spins it alone), and `ladder`. `setTurning`,
`setAngle` and `update` drive the sails at about ten revolutions a minute
-- a working mill turns slowly, not in the blur people draw.
The sails are where the triangles went, and deliberately. A sail is not a
blade, it is a LATTICE the miller dresses with canvas according to the
wind. Drawn solid the silhouette is wrong in the one place everyone looks,
so the bars are modelled and the budget is taken from the buck's boarding,
which reads perfectly well as vertex colour. 1 516 triangles.
Four faults on the way in, all found by measuring rather than by looking:
The roof pitches were built `slope * 2` long and tilted by the complement
of the roof angle, which made two panels 2.79 m across a building 1.80 m
wide, standing up like the covers of an open book. Each pitch covers half a
roof, and the tilt is measured from the horizontal because
`chamferedBoxGeometry` takes an X-Z footprint.
The rails of the tail ladder were tilted the wrong way round. `run` is
positive -- the head sits at a larger z than the foot -- so rotating by
`-lean` laid the rails along the opposite diagonal to the one the rungs
were placed on. They crossed near the middle and diverged towards the ends,
and the bounding box looked correct the whole time because the translation
put the ladder where it belonged and only its internals were wrong. The
support audit found it as loose rungs, and found it only at the small end
of the height slider, where the voxel grid is fine enough not to bridge the
gap: the joint was never real at any size.
Fixing the ladder's foot, which sat 0.10 m below the cross-trees so the
mill hovered above whatever it was placed on, then displaced the rails
perpendicular to their own axis by almost exactly a rung's half-depth --
correcting the rails alone was worse than not correcting at all. Rails and
rungs now share one lift.
And the lowest sail tip came within 0.27 m of the ground, which reads as a
mill about to plough its own field.
28 models, 20 523 triangles, all checks pass, 0 floating cases, GLB
round-trip clean, Conformant at 30 items, build clean.
The blade was a `chamferedBoxGeometry` put through `bendGeometry`. That has
four levels in Y, and once the blade was deep enough to look like a hoe the
bend had to work a 50-degree arc across those four: the cutting edge came
out as a row of steps rather than a curve. `dishedSheetGeometry` exists for
exactly this and is what the shovel's blade already uses -- one seamless
concave surface, curved at every level. It also removes the last
bend-a-box from this model, which is the pattern the helper now refuses.
The blade is deeper and properly wedged. A hoe is narrow where the neck
carries it and wide where it meets the ground, and at 0.115 of the tool's
length with sides nearly parallel it read as a tab riveted to a stick.
Two mistakes on the way, both mine, both caught by looking:
I lowered `neckSweep` from 112 to 95 after reading a render in which the
blade happened to sit edge-on to the camera, and mistook foreshortening for
a wrong angle. Reverted.
Then `curve` was written as `-0.34 * dish`. It is an absolute rise in
metres, not a ratio, so that asked for 340 mm across a 230 mm blade --
seven times its own half-width -- and the sheet came out as a pair of
wings. The shovel scales it by its own blade width and this now does the
same. The dish sliders on the two models run over different ranges, so the
coefficient differs; the quantity does not.
One thing deliberately NOT changed. The reference generated for this model
is a plain strap hoe, while this one is a gooseneck -- a curved forged neck
carrying the blade ahead of the shaft axis, as its description says. Both
are period-correct, and pulling the model towards the photograph there
would be redesigning it rather than fixing it. The blade's depth and wedge
are a different matter: those are true of a hoe whichever neck it hangs on.
27 models, 18759 triangles, all checks pass, 0 floating cases, build clean.
Three faults across two tools, and two of them come down to a model being
named after a material it was not made of.
The pitchfork had three steel tines, 0.24 m on a 1.5 m shaft -- 16% of the
tool against roughly 22% in the reference -- fanned out like a cultivator.
A hay fork has four or five, they are the working length of the thing, and
they stay close to parallel so a forkful stays on them. They are also
wood: the model is called wooden-pitchfork and a hay fork really was cut
and steamed from a branch, because iron is heavy at the end of a long shaft
and hay needs no edge. The socket stays iron, since that is the ferrule.
The shovel's blade was a sheet of polished steel. The strap that runs up
behind it has been modelled as an iron shoe all along -- a shoe for a board
that was being rendered as steel. A medieval digging shovel is a shaped
board with iron along its lip, so the blade moves to oak and the shoe
becomes the only metal on the head.
And it had no handle: the shaft simply ended in a swelled knob. A digging
spade is worked with both hands with the lower one round a crossbar, which
is what separates a spade from a stick with a plate on the end, and it sits
right at the end of the silhouette where the eye lands. The bar crosses the
grip swell rather than butting onto it.
Two notes for later. The wider tine spread pushed the fork past its size
guard at 0.35 m across; 30-40 cm is the head of a real hay fork, so the
limit moved rather than the geometry. And `createTinter` returns one Color
and mutates it while `steelTint` and `ironTint` allocate per call -- the
new code copies out of the first before the next call, but the two helpers
disagreeing about this is a trap in itself.
27 models, 18599 triangles, all checks pass, 0 floating cases, build clean.
The file header has claimed this model "comes to ~0.75 m with a real anvil
stump" for as long as it has existed, while shipping a 0.34 m anvil on its
own, lying on the floor. An anvil is bedded into the end grain of a log:
the block kills the ring, and it puts the face at the smith's knuckle. The
stump is its own part, so anyone wanting the bare anvil for their own bench
hides `parts.stump` -- the same arrangement the bell's frame uses.
Two things went wrong on the way in, and the second corrects the first.
Passing `woodTint('oak', ...)` and `woodTint('oakEnd', ...)` as two
arguments of one call hands the callee the same object twice: the tinter
returns one Color and mutates it, which its own docstring warns about. The
block was painted end-grain on every face. That is a real bug and the fix
stands -- but it was not why the stump looked wrong, and the measurement is
what said so. Against the bench and the table on the same oak slot our
stump is LESS saturated, not more (0.35 against 0.45 and 0.49), so the
colour handling was never anomalous.
The actual gap was in one axis: value 0.54 against the reference's 0.28. A
chopping block is old, split and weathered nearly black, which is a
genuinely darker piece of wood than fresh oak, not a rendering artefact.
Sweeping the lift and measuring lands at hue 30 / saturation 0.28 / value
0.34, against 31 / 0.29 / 0.28.
The block is also wider than first drawn. The anvil overhung it, and a
block the anvil overhangs looks like it would go over the first time
anyone swung at it.
27 models, 18517 triangles, all checks pass, 0 floating cases, build clean.
Three defects, two of them mine.
The weave had gaps you could see the contents through. The band height was
a flat 0.11 of the model height while the spacing between rows is 1/rows --
0.167 at the default six -- leaving a 12 mm hole between every course. It
was a slatted crate, not a weave. The height now derives from the row
count so rows meet at any setting.
The fruit was invisible, and that was a regression I introduced. An earlier
pass fixed produce that hung level with the rim over a gap by resting it on
the base. That was right; what did not follow was that a basket deeper than
its contents then needs enough contents to fill it. Nine apples in two
layers topped out 12 cm below a 10 cm rim -- correctly supported, and
completely out of sight. The basket is shallower, holds fifteen, and the
layers nest at 0.95 of a diameter rather than 1.5, so the heap now mounds
over the rim, which is the only state a fruit basket is ever drawn in.
Closing the weave then exposed two collisions. Overlapping rows meet in
coplanar faces at the nodes where the undulation crosses zero -- on a
cylindrical basket every row shares one radius -- so alternate rows now sit
slightly proud of and behind each other, which is how a weaver's rod really
passes the stakes. And the rim's inner surface sat at 0.16055 against the
liner's 0.16065: a tenth of a millimetre apart, by coincidence, and the
same 22-sided prism. The rim is now the heavy rod it should be and clears
the liner.
Colour: the basket drew from the `straw` palette and rendered at hue 41,
saturation 0.59 against a reference at 29 and 0.50 -- bright yellow
plastic. Willow is a wood, so both the colour and the material slot move to
oak, which measures 7 and 0.03 away instead of 12 and 0.09.
It still does not look like the photograph, and the measurement cannot say
why. Filtering to pixels above 0.18 saturation discards precisely the grey,
weathered pixels that give the reference its character, so the statistic
reports agreement that the eye does not see. Setting the lift to 0.12 to
make willow "paler than oak" turned it salmon -- raising HSL lightness
holds saturation, so a saturated brown climbs towards peach. Sweeping the
lift over 0.00 / 0.04 / 0.08 moved the measurement by 0.01, which is the
clearest evidence that this axis is not where the remaining gap lives.
27 models, 17853 triangles, all checks pass, 0 floating cases, build clean.
Generates one photographic reference per model through the Codex CLI's
image tool, into an untracked references/ directory. The prompt asks for
the same shot every time -- three-quarter view, whole object, plain dark
ground, even light -- because these images are measuring instruments
rather than art: a reference lit from a dramatic angle hides the very
silhouette detail we need to compare against.
Models whose name alone sends a generator somewhere useless carry an
explicit hint. 'Phial' produces perfume bottles; 'sack' produces modern
hessian; 'broom' produces a household brush rather than a besom.
Two rules were missing, and each one had been quietly failing for a while.
The first is structural. Every check in this repo asked about surfaces --
winding, coplanarity, triangle counts -- and none of them could see a piece
of a model hanging in the air with nothing holding it up. A chest lid
separated from its chest passed everything we had. scripts/support.ts
voxelises the model, finds the connected components of occupied space and
requires each one to reach the floor; scripts/audit.ts runs that across
default, min, max, both sweep ends, actions and the midpoint of a morph, so
a variation cannot break what the default gets right. That audit started at
22 failing cases: the bell and the sign had no supporting structure at all
and now have a frame and a post, the lantern was in four pieces, the sack's
collar floated above its body, the basket's fruit hung level with the rim
instead of resting on the base, and the ladder's rungs were sized by a
formula that disagreed with where the rails actually were. It now reports 0.
The second is that we had no target. The feedback loop was an offline
render of our own output, which can only answer "is this broken", never
"is this right". Given a reference photograph of a cask, five defects were
obvious in seconds that three rounds of self-review had missed: four hoops
where there should be six, no rivets at all, a chime cut so shallow it did
not read, and a stave count low enough to leave a visibly polygonal
silhouette. All four are fixed here.
Two things the rivets taught us, both kept as comments rather than as
scars: a stud that tapers outward tilts its side normals back towards the
axis, which the radial check correctly reports as reversed winding -- a
real rivet head flares, so the honest shape is also the passing one. And
once seated deep enough to look right, the stud's buried inner cap falls
inside the shell the check inspects, so it is not generated at all, exactly
as bandGeometry already declines to build the inside of a hoop.
27 models, 16973 triangles, all checks pass, 0 floating cases.
The kit is going public, so everything a reader sees is now English:
27 model sources, the core vocabulary, meta.ts (titles, descriptions and
every slider label), the viewer UI, all script output, REFERENCE.md and
the READMEs.
Rules applied: only comments, doc comments and human-facing strings were
touched. Identifiers, numbers, logic, registry item ids, part names and
material slot names are unchanged — those are a published contract and
were already English.
Verified rather than assumed: for all 53 changed .ts files the code
skeleton (source with comments stripped and string contents replaced by a
placeholder) is byte-identical before and after. Full gate is green —
typecheck, registry conformance, ~510 model checks, GLB round trip and a
production build.
Also fixed here:
- index.html and viewer.html were missed by the translation pass because
the file list only covered .ts/.md/.css. Both had lang="tr" and Turkish
titles.
- The viewer formatted triangle counts with toLocaleString('tr-TR'), which
prints 15.425 rather than 15,425.
- REFERENCE.md's model table header did not match what
scripts/catalog-table.ts emits, so the next regeneration would have
silently changed it. Aligned to the generator.
- Removed scripts/probe.ts and scripts/zdiag.ts; they were throwaway
diagnostics that got committed by accident.
Üçü de render'a bakılarak elden geçirildi. Ortak sebep: her birinde
modelin KONUSU eksikti.
çit — birleşim yoktu. Dört kare çubuk ve önlerine konmuş iki lata;
derinlik/boy oranı 54:1, yani karton. Artık gerçek zıvana var: direk iki
yanak + köprü bloklarından kuruluyor, delik geometrik olarak var oluyor,
kiriş içinden geçip uçlarda karşı yüzden taşıyor. Ayrıca kiriş dağılımı
formülü kaç kiriş olursa olsun hep 0.28–0.78 aralığını dolduruyordu —
üstü ve altı doldurmak yapısal olarak imkânsızdı. 440 → 860 üçgen,
derinlik 0.09 → 0.31 m.
çapa — kaz boynu yoktu. Ağzı sapın ekseninden öne taşıyan kıvrık boyun
olmadan model "kürsü / yol tabelası" diye okunuyordu. Boyun latheGeometry
ile kurulup bükülüyor, ağız boyun ucunun teğetine hizalanıyor; ikisi de
hesapla, gözle değil.
süpürge — demet içi boş bir koni kabuğuydu: bütün levhalar tek halkada.
Üç eşmerkezli halka (6/10/16) kütleyi kuruyor. Bağlar artık demetin dış
yarıçapına oturuyor; eskiden ayrı bir tahmin formülüyle levhaların İÇİNDE
kalıyor ve hiç görünmüyorlardı.
Ölçerek bulunan iki gerçek hata:
- bendGeometry'nin etkisi geometrinin y=0'a göre nerede durduğuna bağlı.
Orijinde ORTALANMIŞ bir gövde simetrik bükülüyor: iki uç aynı yöne
gidiyor, siluet değişmiyor. Çapanın ağzında Z aralığı 0.0337'den
0.0327'ye DÜŞÜYORDU — yani yorumda "sorunu çözüyor" dediğim kavis hiç
yoktu. Tabanı orijinde kurulunca kaçış 44 mm. Yardımcıya yazıldı.
- Süpürgede açılma işareti tersti: rotateX(+f) levhanın ucunu -Z'ye
atıyor, rotateY onu EKSENE doğru çeviriyor. Levhalar ekseni geçip karşı
tarafa savruluyordu; demetin kabuk olmasının asıl sebebi buydu. Gözle
değil ölçüyle yakalandı (genişlik 0.23 beklerken 0.13).
Sistemik: kapanma yarıçapı ve alaca hücresi modelin en büyük ölçüsünden
türetiliyordu. 4.89 m'lik çitte hücre 0.27 m çıkıyor, yani bütün direk tek
hücreye düşüyor ve doku sistemi hiçbir şey yapmıyordu; kapanma yarıçapı
0.69 m ile her şeyi battaniye gibi karartıyordu. İkisi de mutlak sınırlara
alındı — ahşabın damar lekesi nesnenin boyundan bağımsızdır.
Doğrulama: test yamalarının gerçek config alanlarına işaret ettiği artık
denetleniyor. Çapa yeniden yazılırken `bladeAngle` alanı kalkmıştı ama
test hâlâ onu yamalıyordu, yani o varyant sessizce varsayılan modeli
sınıyordu.
render.ts: --angles ile turntable. Tek bir 3/4 açı yanıltıcı — çit ancak
tam yandan bakınca cılız görünüyor.
Kullanıcının prop listesindeki son madde meyve/sebzeydi; sepetle birlikte
kapatıldı çünkü kitte zaten sepet eksikti.
Örgü, biçimi yapımdan türetiyor: yatay halkalar önce düz bant olarak
üretilip köşeleri açılarına bağlı olarak içeri-dışarı itiliyor
(yarıçap × (1 + genlik·cos(dikeySayısı·açı + faz))), ardışık sıralarda faz
π kaydırılıyor. Dikey çubuk başına iki segment yeterli — dalga tam olarak
her çubukta bir artı bir eksi örnekleniyor.
Halka başına iç yüz üretmek altı sırada ~800 üçgen tutuyor ve modeli
lowpoly bütçesinin dışına çıkarıyordu; yerine 44 üçgenlik tek parça iç
astar var.
Yeni `produce` materyal yuvası (mumsu kabuk, roughness 0.52). Meyve rengi
`hue` alanından geliyor: aynı modelden elma, şalgam ya da lahana sepeti.
Modeller (13 → 26): chest, bench, torch, lantern, bell, sign, hay bale,
sack, broom, tankard, book, phial, coin pouch. Kürek, çapa ve yaba
render'lara bakılarak yeniden yazıldı.
core:
- createKitModel artık tipli `actions` ve `update(dt)` taşıyor
- BuiltPart.extras: bir parça birden çok materyal yuvası kullanabiliyor
(sandığın kapağı = meşe tahta + demir kayış, birlikte dönmek zorunda)
- BuiltPart.origin: parça başına dönme merkezi (menteşe, sarkaç)
- bendGeometry / roughenGeometry / mottleGeometry
- 5 yeni materyal yuvası: steel, brass, straw, cloth, leather, glass
- bandGeometry({ inner }) — serbest halkalar kapalı katı olabilsin
my-registry/meta.ts: katalog metadata'sı artık build.ts ve viewer için
tek kaynak. Doğrulama modelle metadata'nın uyuştuğunu denetliyor ve bu
eklendiği gün dört modelde gerçek bir sapma yakaladı.
Doğrulama (~500 kontrol):
- metadata uyumu, eylem semantiği, kare hızı bağımsızlığı
- scripts/render.ts: tarayıcısız yazılım rasterleyici. Bütün doğrulama
geometrikti; "bu kürek küreğe benzemiyor" diyebilmek için modele
bakmak gerekiyordu
- scripts/verify-glb.ts: dışa aktar → geri oku → karşılaştır
GLB dışa aktarımı: viewer düğmesi ve toplu CLI aynı kodu kullanıyor.
Viewer:
- parça paneli (solo/izole) ve materyal yuvası paneli (canlı renk override)
- kit görünümü: 26 model tek sahnede, ölçek ve ton tutarlılığı için
- configure() sonrası gölge bayraklarının düşmesi düzeltildi
