massing: Skill for Claude Code

Verify a Massing frontend change

Standing directions for this repo: docs/roadmap-directions.md. Read those first.

The "preview stall" was diagnosed and FIXED in v0.3.703 — it was never the geometry loader. Five SSE endpoints polled the database inside async def gen(), blocking the event loop for every other client; two open tabs left the server wedged and every request timed out, including the app's boot api.health(). That left projectId unset, so the portal showed "No project open" while a model sat rendered on screen. Fixed with run_in_threadpool; measured >8s timeouts → ~20 ms. Full live verification now works — start both servers and drive the real app.

A SECOND "stall" was diagnosed and fixed 2026-08-02, and it was not the loader either — the canvas had width ZERO. Container 830x572, canvas 0x493, four meshes and 230 triangles built and visible, .frag 200, worker fine, console clean. The renderer sizes itself once at construction; if the container is not at final width yet, that first size sticks for ever. A ResizeObserver in createViewer now tracks it (apps/web/src/viewer/canvasResize.test.ts).

Twice now, "the geometry loader stalls" has meant something else entirely — once a blocked event loop, once a zero-width canvas. Before writing that sentence a third time, ask the page:

const c = document.querySelector('canvas'), k = document.querySelector('#container');
({ canvas: c.width+'x'+c.height, container: k.clientWidth+'x'+k.clientHeight })   // canvas ≈ container × devicePixelRatio

and count what is actually in the scene (v.viewer.world._scene.three, not v.world). A caveat nobody re-tests becomes folklore: this skill asserted the stall in its own frontmatter for weeks while its body said it was fixed, and the contradiction propagated into eight changelog entries.

Always

cd apps/web && export PATH="/c/Program Files/nodejs:$PATH"
npm run typecheck && npm run lint      # Node 24 (Node 18 breaks the build)
npx vitest run <path/if/covered>
npm run build                          # strongest compile check; ~1 min