Unify argument parsing between MCP handlers and agent tool executor #1060

Problem

Issue #1040 established the shared ops/ layer so both MCP handlers and local agent tool executors delegate to the same business logic. However, the argument parsing layer above ops still diverges:

• MCP handlers (packages/core/src/mcp/handlers/nodes.rs): Accept flat properties (e.g., {"id": "...", "status": "done"}) and pass them through to the ops layer, which handles namespacing.
• Local agent tool executor (packages/agent/src/local_agent/tools.rs): Previously required the model to wrap properties in a "properties" object (e.g., {"id": "...", "properties": {"status": "done"}}). This was partially fixed inline but not systematically aligned.

This divergence means the 8B model's flat tool calls fail on the agent path but would succeed via MCP.

Specific Gaps

1. Argument normalization: MCP handlers and agent tool executors parse arguments differently — no shared "args → typed input" adapter.
2. Property flattening: The agent executor now collects unknown top-level keys into properties, but this logic is duplicated inline rather than shared.
3. Error handling: MCP handlers return MCP error codes; agent executor returns ToolError variants. Error messages differ for the same underlying failure.
4. Tool definitions: The agent defines tool schemas in tools.rs (def_create_node(), etc.) while MCP defines them in handlers/. These can drift apart.

Proposed Solution

Extract a shared argument adapter layer that both MCP handlers and agent tool executors use:

Model/Client → Argument Adapter (shared) → ops/ layer → NodeService

Options:

• Option A: Move argument parsing into the ops layer itself (ops functions accept Value and parse internally)
• Option B: Create ops::args module with parse_create_node_args(Value) → CreateNodeInput, etc.
• Option C: Generate tool definitions + argument parsers from a single schema definition

Prior Work

• Issue Refine local agent system prompt for better tool usage and response quality #1040: Created ops layer (core business logic shared)
• Issue Semantic skill discovery via find_skills tool for 8B model and external agents #1051/1055: Unified find_skills via skill_ops (model for future unification)
• Issue Unify skill/prompt seeding with markdown import pipeline #1056: Unify seed/import pipelines (related but different scope)

Acceptance Criteria

• create_node and update_node argument parsing is shared between MCP and agent executor
• Flat properties are accepted uniformly (no "properties" wrapper required)
• Tool schema definitions are generated from a single source (no drift)
• Error messages are consistent between MCP and agent paths
• find_skills pattern (shared ops function) extended to remaining tools