RAG-3: Inject relevant Live doc chunks into system prompt

[nonogv]

Context

Depends on #67 (RAG-2). Part of the v2 RAG milestone.

What

Before building the system prompt, detect Live/device queries and inject the top relevant doc chunks.

Files: src/agent/chat.ts, src/server.ts

Logic:

1. If isLiveOrProductionQuery(latestUserMessage) is true, call retrieveLiveDocs(query, 3)
2. If chunks are returned, inject them into the system prompt as a fenced block:

```

Relevant Live documentation

Title
body text…

Title
body text…
```

Constraints:

• Total injected context ≤ 1000 tokens (truncate body if needed, never truncate mid-sentence)
• Only inject for learning/question requests — skip for pure action commands ("set tempo to 120")
• retrieveLiveDocs must run with a 50ms timeout; on timeout or error, silently continue without injection — never delay the agent turn

Tests in src/agent/chat.test.ts:

• System prompt includes doc chunks when retrieveLiveDocs returns results for a Live query
• System prompt is unchanged when retrieveLiveDocs returns []
• System prompt is unchanged for pure action commands even if isLiveOrProductionQuery is true

Acceptance criteria

• Asking "how does Operator's FM ratio work?" injects the Operator manual section
• Asking "set tempo to 120" does not inject anything
• A retrieveLiveDocs timeout never adds latency to the agent response
• Injected context stays under 1000 tokens