Refactor prompt system: extract generic interactive agent prompt from telegram-specific template

Currently telegram/system.mustache contains ~205 lines of non-partial content, but only ~40 lines are actually telegram-specific. The rest is generic 'ava as interactive agent' instructions that should be reusable across channels (CLI, web, other agents).

What to extract into a shared partial (e.g. partials/interactive_agent.mustache):

• Skills system (~15 lines)
• Memory usage (~12 lines)
• Long-running work / projects.md (~9 lines)
• Workspace (~4 lines)
• Dates & time (~5 lines)
• Tone & style (~16 lines)
• Web citations (~9 lines)
• Tool usage / task workflow (~27 lines)
• Subagent delegation (~24 lines)
• Domain knowledge / PIL (~10 lines)

What stays telegram-specific:

• Message protocol (send_message/send_photo, AGENT_DONE, private vs group)
• Telegram formatting (legacy Markdown rules)
• Group chat response rules
• Heartbeat system

What already exists as partials:

• base_identity (4 lines)
• instruction_hierarchy (7 lines)
• safety_rules (8 lines)

Goal: telegram/system.mustache becomes shared partials + telegram-specific sections. A future CLI or web agent template reuses the same shared partials with its own channel-specific sections.

Approach: 1. Create partials/interactive_agent.mustache with the generic content 2. Refactor telegram/system.mustache to include the new partial 3. Verify no behavioral change in telegram agent 4. Document the partial system so new channels can compose their own system prompt