Phase 0 spike results (2026-10-14)

Local toolchain: Bun 1.4.2 (pinned target 1.4.3), TypeScript 7.0.2, macOS arm64. The spike code was removed in Phase 5; it is preserved in commit 02fa947 under each package’s spikes/ folder. D1 was re-verified against Foldkit 0.167.0, the workspace pin, after consolidation.

Spike	Question	Result	Evidence
D1	Which Foldkit seam gives update-in-place, typed events and scoped release?	Hybrid works: CustomElement.define for typed props + Schema-decoded events, Mount.define (one-shot) for lifecycle.	d1-adapter/d1.test.ts passes in Chromium
D2	Can the reconciler hot path follow the import-gated lint rules?	No, at acceptable cost. Effect-style (Option/HashMap/Array) diff is 11–21× slower; kernel style lints clean under all rules.	d2-lint-layering/bench.ts
D6	Can Bun drive deterministic browser visual tests?	Yes. Playwright library under bun test, headless shell + SwiftShader: two renders pixel-identical (0 diff).	d6-playwright/d6.test.ts
Lint	Do oxlint jsPlugins and type-aware rules need Node?	No. bunx --bun oxlint runs linteffect/* and typescript/no-unnecessary-type-assertion.	bun run lint
TS 7	Do Foldkit, three, Effect and Playwright types work under TS 7.0.2?	Yes. 1,012 files checked in 0.17 s, no errors.	bun run typecheck
Core purity	Can core use three with no DOM lib?	Yes. three scene graph types check and run headless in Bun; a document reference fails compilation.	packages/effect-three/tsconfig.json (lib: ["ES2023"])
Docs	Does doc tooling support TS 7?	TypeDoc no (peer typescript ≤ 6.0.x). API Extractor yes (bundles TS 5.9.3, reads emitted .d.ts).	npm registry

D1 details

• Mount must be Mount.define, not Mount.defineStream. A completing stream closes the Mount scope immediately and runs the finalisers while the element is still mounted. define keeps the scope open until the element unmounts.
• Prop is not exported from foldkit/html’s public surface. The public route to a diffed DOM property is a CustomElement.define property factory, which sets the property without Schema-validating it on each render.
• Foldkit re-sets the property with a fresh object on every render, even when the scene is unchanged. In the spike, an unrelated model change (nodeClicks) produced a redundant update:red. The runtime must short-circuit structurally equal descriptions cheaply.
• Pointer events go through element.dispatchEvent(new CustomEvent("scene-pointer", { detail })). Foldkit decodes detail with the declared Schema and maps it to the user’s typed Message through OnScenePointer. The Mount’s own message is just a lifecycle acknowledgement.
• Mount.define’s execute has R = never. Library services (renderer, frame source, resource registry) cannot come from Foldkit’s resources Layer. They are provided inside the Mount from an adapter-built Layer (see ADR-002).
• Not yet covered (Phase 3): time-travel viewStateChanges → Paused suppressing events.

D2 numbers (1,000 keyed nodes, mitata, Apple Silicon)

Case	Kernel (plain TS)	Effect style	Ratio
no-op	24 µs	515 µs	21×
one change	25 µs	533 µs	21×
all change	48 µs	541 µs	11×

The Effect-style no-op alone (0.51 ms) exceeds the accepted 0.2 ms budget.

D6 details

• Launch args: --use-angle=swiftshader --enable-unsafe-swiftshader --force-device-scale-factor=1. Use preserveDrawingBuffer: true for screenshots.
• Leak oracle works: renderer.info.memory.geometries dropped 1 → 0 after dispose(). textures stayed at 1 with no user textures, because three allocates an internal texture. Baselines must be captured after the first render, not assumed to be zero.

Lint findings that affect later phases

• Idiomatic Foldkit code fails several rules in Effect-importing files: { ...model } (no-naked-object-state-update), ternaries in update/view (no-ternary), string comparisons (no-magic-domain-string). The demo app must use Struct.evolve/modifyFields, Match and Schema literal unions.
• no-effect-wrapper-alias flags any const x = Effect... alias. Public Effect values should be functions that return Effects.

Labels under WebGPU (2026-10-19)

Question: can effect-three draw troika labels under WebGPURenderer, where troika’s own material (GLSL patched in with onBeforeCompile) cannot run?

Result: yes, by keeping troika’s text pipeline and replacing only its shader. troika’s getTextRenderInfo lays text out (fonts, wrapping, bidi, colour ranges) and builds the glyphs’ signed distance fields into an atlas on its own canvas, independent of the renderer. A TSL node material ports troika’s fragment logic: one instanced quad per glyph over its bounds, the atlas cell from the glyph index (four glyphs per texel, one per channel), exponential decoding to a signed distance, and anti-aliasing from fwidth. In headless Chromium on SwiftShader’s WebGPU adapter, WebGPURenderer reported its WebGPU backend and drew two labels (50 glyphs) cleanly, colour ranges included.

Two findings to carry into an engine:

• Round the glyph index. Interpolated across the quad, an index of 5 can arrive as 4.9999, and mod(index, 4) then picks the wrong channel: whole glyphs come out as neighbouring ones. troika’s GLSL rounds (floor(ch + 0.5)) for this reason.
• Copy the atlas as bytes. Uploaded as an image, troika’s WebGL canvas goes through premultiplied alpha, which marks the R, G and B glyphs where the A glyph is thin; readPixels into a DataTexture (troika’s own Safari workaround) is exact. The atlas grows as labels add glyphs, so an engine must re-read it when troika reports new ones.

Not yet covered: outlines, blur and curves (troika’s vertex and outline passes), clip rectangles, a lit material, and depth with other transparent objects. Evidence: packages/effect-three-troika/spikes/tsl-labels/ (bun spikes/tsl-labels/run.ts).

Outcome (2026-10-19): built. troikaLabels from @opsydyn/effect-three-troika/webgpu draws labels this way behind LabelEngine.createTsl, with outlines and blur (a first pass, as troika draws them), curves, colour ranges, lit materials and picking. Two more findings from building it:

• troika’s glyph indices are integers. Documented as a Float32Array, they are the typesetter’s glyph id array; given to WebGPU’s WebGL 2 backend as is, it declares an int input and the draw fails (“Vertex shader input type does not match”). Copy them to floats.
• A superseded layout must still draw again. troika answers a layout whose glyphs an earlier, unfinished layout is still building at once, before their fields have landed in the atlas. Showing only the latest layout is right, but the earlier one’s answer must still ask for a frame, or a scene that draws on demand keeps an empty atlas copy.

Against the WebGL path on the samples, glyphs, colour ranges, outlines and layout match; the clearcoat highlight on the gallery’s curved, lit title is narrower under WebGPU.