AI assistant v3: flow-node behaviors + physics tools

[AlexZ005]

The assistant can now create behavior, not just objects: flow nodes always, and physics (bodies, joints, simulation control) behind a per-provider opt-in. Target prompt: "create a moving spider" — body + legs from primitives, grouped, patrolling the scene with bouncing legs, simulated.

What's in

• 'flownodes' history kind (flowGraphs.js) + recordFlowNodesEntry: node/edge create, data edits and removals as one undoable entry; entries carry SERIALIZED copies so replayed re-broadcasts hash identically on every peer (nodesync drift guard).
• setPhysicsFor(uuid, patch) (physics.js): the shared physics write path — props history + objectParameters replication + collider-viz poke + the CL-A A2 live mid-sim collider rebuild. Inspector setPhysics and enablePhysicsOnSelection refactored onto it, so all three callers (Inspector / quick action / AI) behave identically.
• src/lib/ai/flowTools.js (new): the five executors. create_flow_nodes / update_flow_nodes are always available (curated node-type enum + alias map, editor-identical node/edge construction incl. the handle-qualified edge-id format, implicit-owner leverage: one node + zero edges drives the graph's object). set_physics / create_joints / control_simulation gate behind the provider's Physics tools checkbox — as do the physics-family node types (mass/bounciness/friction/angularvelocity/motor + the CL-C collider override, onimpact/onenter/onexit triggers, velocity readout).
• tools.js wiring: schemas (built per call so the enum tracks the gate), dispatch, repairToolCall additions (5 names, invention aliases like add_behavior/start_simulation with action fill-in, shape inference incl. physics-only updates → set_physics), summarizeScene now carries per-object physics + compact flow (12-node cap) plus top-level sceneFlow + joints; system prompt gains the flow block (spider recipe) and a gated physics block (world-aligned jointing, auto sim start).
• assistant.js labels + tally (control_simulation counts 0); providers.js physicsTools flag; Settings → AI checkbox with a link to the new docs page.
• Docs: ai/local-models.md (qwen3_xml-not-hermes vLLM guidance, model-size table, checkbox semantics) already committed on the theprototype-docs repo.

Semantics worth noting

• One prompt = one undo, but sim start/stop is NOT part of the batch — undo never stops a running simulation (stop/reset first, then one undo reverts objects + nodes + physics + joints together).
• set_physics supports the CL-A A3 sensor flag; custom colliders and freeze axes stay editor-only.

Verification

• New suite tests/e2e/ai-flow-physics.test.cjs — 38 checks over a scripted mock endpoint: the full spider scenario (create → group → pathpatrol on the GROUP graph + bounce on a leg via refs → wired scene-graph pair → update_flow_nodes speed tune → static ground → hinge → sim start), broadcast capture, gating (disabled error + per-item physics-node rejection), invented-name repair, one-undo/redo-without-sim.
• Regressions green post-merge: ai-assistant, ai-local-model, flow-physics-collider, physics-joints, flow-object-graphs.
• npm run build green; npx svelte-check 438/62 — exactly the release.yml baseline gate.
• Merged release/next in (colliders v2 conflict resolved by moving physicsShapeChanged into setPhysicsFor).
• Manual check remaining: live smoke against the real vLLM box (--tool-call-parser qwen3_xml) — "create a moving spider" / "make the spider faster".

🤖 Generated with Claude Code