feat(agent-core): harden sendAgentMessage (sanitize sender label, dedup targets)

[WinterYukky]

Summary

Hardens sendAgentMessage with two small, self-contained safety improvements.

Motivation

The inline [Message from <name> (<id>)]: prefix that wraps an agent-to-agent
message is fed into the recipient LLM's prompt. A stray newline or bracket in the
sender label could break out of that envelope (prompt injection). Separately, a
caller passing the same target session id more than once would deliver to — and
wake — that target multiple times.

Changes

• Sanitize the sender name and worker id embedded in the inline prefix via
  sanitizeSenderLabel (already used elsewhere in prompt.ts), falling back to
  agent / unknown when sanitisation empties the value.
• De-duplicate targetSessionIds ([...new Set(...)]) so each unique target is
  delivered to / woken exactly once.
• Add focused unit tests for both behaviours (no prior test existed for this module).

Testing

• npm run build -w packages/agent-core (tsc) — pass
• vitest run new tests — 3 passed
• full agent-core suite — 271 passed
• prettier format:check (agent-core) — pass

Notes

• sanitizeSenderLabel already exists upstream in prompt.ts (its JSDoc even
  references sendAgentMessage); this wires it into the previously-unsanitised
  call site.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.