R3: Message compaction after iteration 5

[VarunGitGood]

Problem

The ReAct loop in repi/investigation/react_loop.py lets the message history grow monotonically. By iteration 8+, the LLM is reasoning over many old observations it no longer needs — cost rises and signal-to-noise drops.

Scope

• After iteration 5, summarize observations from steps 1..N-3 into a single "findings so far" assistant message.
• Keep the most recent 3 raw observations un-summarized so the model still has fresh context.
• The summary must preserve every chunk_id referenced in the collapsed steps, so the final-answer grounding validator still passes.

Acceptance

• Token usage per investigation drops measurably on the longer eval scenarios.
• Final-answer chunk_id grounding still passes validation in investigation/schema.py.