The motion-blur snippet integrated translation only. Its output stage was one SVG filter per target whose copies were feOffset, and feOffset only translates, so a beat that only scaled or only turned sampled a displacement of zero and rendered sharp. No SVG filter primitive can give a copy its own affine transform, so the sampling loop was never the problem: the output stage was. Copies are now DOM clones of the element in an isolation:isolate group, each at 1/N opacity with mix-blend-mode plus-lighter, each carrying the element's resolved transform at its sample time. plus-lighter adds premultiplied colour, so N + 1 copies at 1/N sum to the average the shutter integral asks for; the isolation is load-bearing rather than tidy, because without a transparent backdrop of its own the first copy adds onto the page and blows a light background out to white. Reading the resolved matrix instead of a list of named GSAP properties is what makes every driver work at once: translation, scale, rotation, 3D rotation and skew arrive in the same one value. Measured against the After Effects export the shutter model was fitted to, per beat, previous stage to this one: scale +0.62 dB, rotate X +0.43, rotate Y +0.84, rotate Z +1.15, translate X -0.18, translate Y -0.04. The two small translate regressions are inherent and not tunable: the filter offset one rasterisation so every copy was pixel-identical, where each copy is now rasterised independently at its own fractional transform with its own subpixel antialiasing. Three consequences of the new stage, each carrying its own fix: - mix-blend-mode flattens preserve-3d, so a copy cannot inherit the parent's 3D context. Each copy carries the parent's perspective as its own first transform function instead. - A copy is styled by nothing that selected the original, since dropping the id also drops every #id rule that gave it size, colour and font. Each copy carries its own resolved style inline, read once and replayed. Those values are px, so a container-relative element needs them again when its box changes: a ResizeObserver re-snapshots rather than paying a subtree walk every frame. - A beat that moved and faded at once kept a full-strength smear behind a vanishing element. The element's opacity now rides the group; the 1/N weight stays on the copies. The axis option is gone. It selected which of x and y to sample, a question the resolved matrix does not have. shutter-slam is the new component: one word through six single-property beats, two translate slams, a scale punch, then a full turn about each axis, which exercises every driver the model covers in one mount. Easing is power2.inOut because it measured best, not because it was the default. A bezier fitted to the reference's own bbox centres looks like the better curve and renders worse: a blurred bbox centre is the shutter window's average rather than the instant, so on an accelerating segment the fit measures its own blur. Co-authored-by: Miguel Ángel <miguel.sierra@heygen.com>
Catalog artifact
Two files ship from here, and only these two. The CLI fetches them over HTTP
when a user opts into offline catalog search (catalog --query ... --on-device),
so they are served from the registry rather than bundled in the package.
| File | What it is |
|---|---|
local-vectors.json |
{ model, dimensions, names }. names is the row order of the binary. |
local-vectors.bin |
Float32, row-major, names.length * dimensions values, no header. |
Currently 168 rows at 384 dimensions: 258,048 bytes, one row per installable registry item.
Provenance
No digest, revision or build timestamp is recorded in either file, and the
runtime only checks that dimensions matches the model it loaded. model is a
label the build wrote, not a proof. So the only real provenance check is to
rebuild the rows and compare them, which works because both inputs are in this
repository:
- the text each row was embedded from is
itemRetrievalText(registry-item.json)(title, description, tags, joined by newlines, name deliberately excluded), overregistry/blocks/*andregistry/components/*sorted by name; - the model is the pinned quantized
bge-small-en-v1.5ONNX build thatpackages/cli/src/registry/localModel.tsdownloads.
Re-embedding in batches of 16, the batch size the build uses, reproduces the shipped rows exactly (cosine 1.000000). Batch size matters: the same text embedded alone differs at cosine 0.9969, because padding within a batch changes the quantized result. A rebuild that does not match this way was not built from this registry, or not with this model.
Rebuilding
bun scripts/catalog/build-local-vectors.ts
It reads registry/blocks/* and registry/components/* directly, so the
rebuild has no input outside this repository. You rarely need to run it by
hand: a lefthook catalog-index pre-commit command regenerates and re-stages
both files whenever a staged registry-item.json changes, and CI fails the
"Catalog: search index covers the registry" job if the index is ever missing an
item.
build-catalog-artifact.ts does not produce these files. It builds the
3072-dimension hosted artifact from an external shelf file, for the hosted search
tier that this repository does not ship. Its --shelf, manifest.json and
text-embedding-3-large have nothing to do with local-vectors.*.