feat: add RADAR governance visualisation with Vela NPC

[ktkorpela-prog]

Summary

Adds visual governance to Pixel Agents: when a Claude Code agent calls the radar_assess MCP tool, the agent character walks to a Risk Assessment desk where Vela — the first NPC in the system — delivers a verdict (PROCEED / HOLD / DENY) with a colour-coded animation.

⚠️ Draft — testing in progress, animation pipeline verified end-to-end.

Demo flow

1. User places a radar_desk in the layout editor → Vela auto-spawns behind the desk
2. User places any chair adjacent to the desk → it becomes the visitor seat
3. Agent calls radar_assess → walks to visitor seat
4. Vela animates a stamp; verdict badge appears above the agent + stamp mark on the desk
5. After 3s, agent walks back to their own desk

What's included

Core

• Parser detection of mcp__radar__radar_assess (extracts verdict, tier, riskScore, triggerReason, recommended)
• CONSULT character state + walk arrival detection
• FIFO radarQueue.ts with deduplication and seat management
• NPC manager (npcManager.ts) — Vela lifecycle, stamp animation, T1/T2 visual distinction
• Canvas-drawn verdict overlays (stamp mark + badge + bold text label)
• Permanent "RISK ASSESSMENT" banner above the desk
• Hover tooltip: "Vela / Risk Architect"

Bonus fixes for upstream

• esbuild.js was pointing to providers/file/hooks/claude-hook.ts but the file is at providers/hook/claude/hooks/claude-hook.ts — fixed. The hook script was silently failing to build, causing Cannot find module errors when Claude Code invoked hooks.

Dependencies

Component	Required for	Notes
radar_desk placement	Vela visible	Furniture-driven; no radar_desk = no Vela = feature dormant
@essentianlabs/radar-mcp	Verdict animations	Without it, the feature is fully dormant — no errors, no warnings
LLM API key	Optional T2 deep assessment	Rules-engine assessment works without one

Tests

• 29 new unit tests (153 total, all passing)
• Coverage: extractRadarVerdict (18 tests), radarQueue (11 tests)
• Manual test verified end-to-end with real radar_assess calls

Placeholder assets

vela.png and RADAR_DESK_FRONT.png are placeholders — designed for drop-in replacement. A pixel artist contribution is welcome (no code changes needed).

Test plan

• Place radar_desk and adjacent chair in layout
• Verify Vela spawns and "RISK ASSESSMENT" banner appears
• Hover Vela → tooltip shows "Vela / Risk Architect"
• Trigger radar_assess from a Claude Code agent
• Agent walks to visitor seat, Vela animates, verdict displays, agent returns
• Without radar_desk placed: feature fully dormant, no errors
• Without @essentianlabs/radar-mcp installed: feature fully dormant, no errors

Backward compatibility

Fully backward compatible — all changes additive. New CONSULT state, NPC system, message types, and furniture entry don't affect existing behaviour.

🤖 Generated with Claude Code

[florintimbuc]

Hey @ktkorpela-prog, this is genuinely exciting work :D
We just merged a sizeable architecture refactor to main (for the v1.4 release) that changes how agents communicate and how state is managed - AgentRuntime + AgentStateStore + an AsyncAPI-defined WebSocket protocol, plus the four-package split (core/, server/, adapters/vscode/, webview-ui/).

That's going to cause meaningful rebase conflicts here, especially around the parser, character state, the hook-script paths etc. I'd really like to help you rebase this and work through the conflicts together!

I'll also spend some time thinking about how we can integrate this as a plugin/extension surface rather than a closed built-in feature, NPCs and governance overlays feel like exactly the kind of thing that should live alongside HookProvider, TeamProvider, Client, Adapter as a composable extension point (something like NPCProvider or OverlayProvider).

That way you keep ownership of Vela and the core stays lean. Let me sync after I take a closer look at the diff :)