mcp: add furniture assessments and public agent skills

[Aymericr]

What does this PR do?

Agents can now assess prospective furniture using exact dimensions without adding a temporary item to the scene. The collision tool returns scaled, rotated footprints, explicit clearance and units, skipped evidence, and unsupported checks; GLB export reports its unsupported status truthfully.

Adds standalone pascal-3d and furniture-fit skills, synthetic examples and evaluation fixtures, shared Claude/Codex Git marketplace packages, and CI package validation. Instructions preserve account and data boundaries, distinguish missing geometry from a passed check, and limit the standalone local HTTP service to one active agent client.

Claude Fable 5.1 reviewed the MCP implementation and package contracts. The corrected native Codex gate passes 20/20 frozen synthetic tasks, independently adjudicated by Fable; the original 17/20 result and grader errors are preserved. Final native Claude treatment passes 3/3 tasks and 22/22 criteria. These bounded trials do not establish a general performance uplift. A separate native task passes through a clean packed CLI installation and its managed HTTP connector.

Clean local type checking and builds pass; 356 MCP tests and 20 real stdio transport cases pass. The full suite passes with the repository's pinned Bun 1.3.14; Bun 1.4 timed out an unchanged canopy stress test. Both skills install and resolve all references from the public branch in skills.sh, Claude, and fresh Linux Codex flows. Exact source hashes and limits are recorded in skills/VALIDATION.md.

How to test

1. Use Bun 1.3.14, then run bun install --frozen-lockfile.
2. Run bun run check, bun scripts/validate-skills.ts, and claude plugin validate . --strict.
3. Run bun run check-types, bun run build, and bun run test --concurrency=2.
4. Run bun packages/mcp/scripts/furniture-fit-journey.ts. All 20 transport cases should pass, including exact-dimension candidates, rotations, units, unsupported checks, nonmutation, and reconnect persistence.
5. Review the source-specific native validation record. After merge, release MCP and CLI on the beta dist-tag and verify clean installation from npm and the default Git branch. Git installation is separate from official marketplace listing.

Screenshots / screen recording

Non-visual runtime, documentation, and packaging changes.

Checklist

• I've tested this locally with bun dev
• My code follows the existing code style (run bun check to verify)
• I've updated relevant documentation (if applicable)
• This PR targets the main branch

---

Note

Medium Risk
Changes the MCP collision tool contract and agent-facing error semantics; impact is mitigated by broad unit tests and the furniture-fit journey harness, but downstream agents on older servers may see narrower schemas.

Overview
Ships public agent workflows (pascal-3d, furniture-fit) with skills.sh install docs, shared Claude/Codex marketplace manifests, eval fixtures, and bun scripts/validate-skills.ts wired into CI.

MCP check_collisions is expanded into a bounded furniture-footprint assessment: optional minimumClearance, floorOnly, and a read-only candidate (never mutates the scene); structured results add status, units, source/effective dimensions, skipped items, unsupportedChecks, overlap vs clearance violation, and assessmentGraphHash. Unknown levels error instead of returning an empty pass; footprint inspection rejects missing/zero/tilted geometry as insufficient_evidence. Read-only READ_ONLY_TOOL_ANNOTATIONS are applied across inspection tools; export_glb now sets isError: true for not_implemented. measure adds areaUnits. A W02 stdio journey script exercises 20 frozen cases plus reconnect persistence.

Docs/README describe install paths and schema-version caveats for older hosted releases.

^{Reviewed by Cursor Bugbot for commit a89747b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[pascal]

I hit an error while handling your request (Model unavailable on AI Gateway free tier: Free tier users do not have access to this model. Upgrade to paid credits at https://vercel.com/d?to=%2F%5Bteam%5D%2F%7E%2Fai%3Fmodal%3Dtop-up for unrestricted…).

Please try again, rephrase, or reach out if it keeps failing.

Error id: 5dbc0aac-4feb-476b-96ad-99674b4c1278

[cd2dmzzd6p-ui]