[LinuxDevil]

What changed

Pure, immutable graph mutations plus the single validation entry point for the canvas graph v2 (no canvas/Inspector wiring — that is the C24 cutover).

• graph/v2/mutations.ts: NewNode, ownerForNewNode (selected agent → selected node's owner → root; root-only kinds always agent, notes null), addNode (dedupeId(baseNodeId(...)) ids; note:n<ts><counter> for notes), removeNode (root returns same reference; inline agents remove their whole owner-chain subtree plus matching emptySections/unplaced), duplicateNode (root + notes refused, KIND_RULES[kind].duplicateData ?? structuredClone, +40/+40), moveNode, setNodeData, patchNode (merges into data.item / data.fields for agents / data.step for flow / data for notes; undefined deletes the key), setCollapsed, connectNodes (canConnect refusal verbatim, then KIND_RULES[source.kind].connectFrom ?? default re-own). Node ids are never recomputed after creation — nodeIdMap rekeys at save.
• graph/v2/reown.ts: the re-own effect (reownNode — id-stable, {x:24,y:48} entering a group, absolute position leaving one). Kept in a leaf module so kinds/*.ts connect effects can share it without an import cycle through kindRules.
• graph/v2/kindRules.ts: ConnectResult, KindRules<K> (validate / connectFrom / duplicateData), and the KIND_RULES registry starting with agent.
• graph/v2/kinds/agent.ts: agentRules — connectFrom = re-own target to the source agent (C03 adds validate, C04 duplicateData).
• graph/v2/validateGraph.ts: validateGraph(graph): Problem[] — a CORE_CHECKS table (one function per row: no main agent, duplicate node id, missing owner, root-only kind on a non-root owner, duplicate mcp/subagent keys per owner, duplicate memory names, non-unique guardrail/permission/hook/trigger/abort item ids across the graph, second output on one agent) plus KIND_RULES[kind].validate per node in orderedNodes order. specPaths comes from one graphToSpec call with a tolerant mirror for graphs it cannot re-emit (duplicate record keys, second output). Order: graph-level problems first, then node walk order, errors before warnings within a node.
• types.ts: added ROOT_ONLY_KINDS (trigger, abort, flow — derived from ROOT_ONLY_SECTIONS).

How it was tested

• npm run typecheck --workspace apps/agent-forge — clean
• npm test --workspace apps/agent-forge -- --run src/graph/v2 — 1444 tests pass (124 new mutation cases incl. the 100-seed mutation round-trip property, 16 validateGraph cases — one per core-check row with exact message/nodeId/fieldPath, ordering, clean worked example → [])
• npx vitest run (apps/agent-forge) — 2412 pass · npm run typecheck:server, npm run test:server — 1121 pass · root npx tsc --noEmit — clean
• npm run lint (root) — clean · Studio lint — 0 errors (31 pre-existing warnings, none in these files)
• npm run test:coverage — 7246 pass; 6 failures in src/spec/v2/templates.test.ts README title checks (known CRLF line-ending environment issue, unrelated; create-lousho dist was built so its e2e passes)
• npm run fallow — clean (no dead code/exports, no cycles, no new clones)
• npm run docs:ratchet, npm run api:check, npm run docs:llms:check, npm run size-check — all pass

Acceptance criteria

• Every mutation returns a new Graph and never mutates its input (deep-freeze the input in tests)
• removeNode(g, 'agent') returns g unchanged (same reference)
• Removing an inline subagent removes its whole subtree
• connectNodes returns canConnect's refusal verbatim; agent → tool re-owns the tool
• Each core check row produces its message for a minimal graph and nothing for a clean one

Generated with Devin