Improve compaction messages to enable better post-compaction agent behavior

[tlongwell-block]

Summary

Updates the continuation messages injected after context compaction to explicitly inform the agent that its context was compacted. This gives the agent accurate information about its own state, enabling it to act appropriately after compaction events.

Problem

The previous wording was passive and indirect:

"The previous message contains a summary that was prepared because a context limit was reached."

This phrasing obscures what actually happened. The agent doesn't clearly understand that:

1. Its context window was compacted
2. It's now operating from a summary, not full conversation history
3. It may need to re-verify assumptions or reload context

Solution

Updated all three continuation messages to use explicit, direct language:

"Your context was compacted. The previous message contains a summary of the conversation so far."

This accurate framing allows agents to:

• Understand their current state
• Make informed decisions about whether to reload context (e.g., re-read files, check TODO)
• Avoid assumptions based on "remembered" details that may have been lost in summarization

Changes

• CONVERSATION_CONTINUATION_TEXT — normal conversation flow
• TOOL_LOOP_CONTINUATION_TEXT — compaction during active tool execution
• MANUAL_COMPACT_CONTINUATION_TEXT — user-triggered /compact

Testing

All existing context_mgmt tests pass. No behavioral changes to compaction logic—this is purely a wording improvement to the injected messages.

[DOsinga]

it would be even better if we just injected the reason into the template we render so this is just inside an {% if %} statement there, but we can do that some other time.