Implement Context Compaction

PROJECT CONTEXT:
My agent dies or degrades on long tasks as the context window fills. Analyses of production harnesses show the fix: monitor usage, and at a threshold compact old turns into a structured summary that preserves the load-bearing facts.

GOAL:
Implement threshold-triggered compaction: measure context usage, summarize old turns into a structured digest, and continue the session seamlessly.

REQUIREMENTS:
- Continuous token accounting for the assembled context
- A compaction trigger at a configurable threshold (default ~75% of window)
- A structured summary format: task state, decisions, file/tool facts, open items
- Recent turns kept verbatim; only older turns are compacted
- Post-compaction continuity: the agent resumes work without re-asking or re-doing

CONSTRAINTS:
- The summary is produced by a dedicated summarization call with its own prompt, not ad hoc
- Critical anchors (original task, active plan, key constraints) must never be summarized away

DEFINITION OF DONE:
- A session exceeding the window completes via at least one compaction
- Post-compaction, the agent correctly answers questions about pre-compaction decisions
- Token accounting logs show usage dropping after compaction and staying under the window
- The original task statement and active plan survive verbatim

COMMON FAILURES TO AVOID:
- Compacting everything including the task itself, leaving an agent with amnesia
- Triggering at 100% so compaction itself fails for lack of room
- Freeform prose summaries that drop file paths, decisions, and constraints
- Summarizing tool results the agent still needs at full fidelity