Improve MCP onboarding for local and weak models

[VelizarSeleznev]

Summary

• make OpenCode, direct LM Studio, and Antigravity independent MCP setup targets
• keep the existing complete MCP registry instead of introducing a third local profile
• add explicit project binding and fail-closed multi-editor selection
• make scene persistence compatible with the editor transport contract
• tolerate two narrow weak-model argument mistakes without adding tools: summer_remove_node parent/name aliases and unambiguous op-less AddNode/SetProp batch items
• ship one-paste setup guidance and dedicated LM Studio/Antigravity references for 2.8.0

Customer-path evidence

• OpenCode 1.18.12 + LM Studio 0.4.20 + google/gemma-4-26b-a4b-qat: playbook, context, tree read, add, save, read-back, diagnostics, remove, and final read-back all passed in one disposable project
• direct LM Studio with the same Gemma model: the same end-to-end mutation/verification/cleanup path passed without OpenCode
• a real follow-up request in the same LM Studio conversation added a SimpleCube with BoxMesh and a Sun; the resulting main.tscn was verified on disk
• the observed op-less summer_batch failure was reproduced, covered red-to-green, and then passed through the rebuilt LM Studio integration with create/read/remove/final-read verification
• Antigravity's normal interactive agy flow executed the MCP mutation path; current agy -p non-interactive mode remains an upstream client limitation and is documented as such

Verification

• npm test — 42 files, 405 tests passed
• npm run build
• npm pack --dry-run
• git diff --check origin/main

Rollout

This is the canonical package change for 2.8.0. The engine mirror and public docs/site PRs should stay coordinated with it. Do not merge the published setup text before the matching npm release is available.