Asset library: two entries, and scope library repos to the game that installed them

[duolahypercho]

What

Adds two entries to the asset registry and scopes library repos per-project, the way Unity and Unreal do.

The entries

Repo	License	Why
dimforge/rapier	Apache-2.0	Standalone WASM physics engine. Framework-agnostic, so it drives plain three.js meshes.
pmndrs/three-stdlib	MIT	Typed stand-alone builds of the three.js example modules — controls, loaders, postprocessing.

pmndrs/use-cannon was the original suggestion and is rejected: it is React-Three-Fiber-only with no non-R3F entry point (this editor is raw three.js — grep -rl "@react-three" client/src is empty), it has no LICENSE file at all (gh api repos/pmndrs/use-cannon/contents/LICENSE → 404, .license → null), and it was last pushed Feb 2024.

Browse globally, use locally

search_library was the one source taking no slug while its sibling search_local correctly branched on it, so every game saw every repo as equally usable. That is the one place CaliCode diverged from the Unity/Unreal model:

Layer	Unity	Unreal	CaliCode
Storefront (browse all)	Package Manager	Fab	asset_catalog
Add to this project	→ manifest.json	→ Content/	asset_pick
Consumer sees only installed	compiler	Content Browser	✗ → fixed here

Now:

• hits carry installed, read from this project's settings.assetRepos
• detail.url only once installed. Repos are metadata-only pointers, so the url is the actionable payload — handing it out for an uninstalled repo let the agent build against something the project never took on.
• detail.currentSettings reports the game's tuned values. It previously returned catalogue defaults, so a game that set gravity to 0 still read -9.81. That was an outright bug.
• uninstalled hits carry a hint pointing at asset_pick, which already was the "Add to Project" step and is unchanged.

Driven against a real core on a scratch port, two projects:

1. Game A, before install
   rapier-physics   installed=False  url=(withheld)  + hint
2. asset_pick → {attached: True, settings: {gravityY: -9.81}}
   then tuned gravityY → 0
3. Game A again
   rapier-physics   installed=True   url=https://github.com/dimforge/rapier
                    currentSettings={"gravityY": 0}
4. Game B (never installed)
   rapier-physics   installed=False  url=(withheld)

Card layout fix

three-stdlib is the first entry with enough tags to wrap, which exposed this: the grid card is a <button>, and a bare button centres its content box. Content taller than the grid-stretched row overflowed top and bottom into overflow-hidden, so the gradient header bled over the rounded corner.

        card top  visual top
card 1     137       148
card 2     137       148
three-stdlib 137     138     ← 10px up-shift

flex flex-col on the card plus shrink-0 on RepoVisual — all three now measure identically.

Verification

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, 888 Rust tests (+2 new) — run in a scratch worktree containing only this commit's changes, not the full working tree
• npx tsc -b --noEmit, 903 client tests
• Driven headlessly in light and dark across grid / detail / installed / attached-with-settings / narrow / empty, no console or page errors
• Confirmed in the packaged app after pnpm desktop:build

Known issue — not fixed here

With more than one registry entry, workflow.spec.ts:312 fails intermittently: 7 of 12 runs with this change, 0 of 7 at baseline.

I instrumented the two obvious suspects and disproved both: the editor_bridge broadcast never lags (Lagged count 0 across failing runs) and core-side editor tool calls never time out. Evidence points at a page.evaluate / page.reload() race inside the spec itself rather than a product defect — but that is not proven, and the spec is deliberately left untouched rather than edited until it passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_014YTKbJdRjJtAHEuLxkeFkm