WAL compaction for long-running orchestrations

[chernistry]

Description

For orchestrations running 3+ hours with thousands of ticks, the WAL grows unboundedly. Implement WAL compaction that: checkpoints committed entries, truncates the WAL to only uncommitted entries, and preserves a summary of compacted entries for audit. Different from ORCH-007 (WAL replay) and ORCH-019 (checkpoint/restore) -- this is about WAL size management.

Metadata

Field	Value
Priority	P1
Scope	medium
Complexity	medium
Role	architect

Implementation Suggestions

• Integrate with the existing merge queue (src/bernstein/core/merge_queue.py) and worktree management
• Extend the audit trail (src/bernstein/core/audit.py) and policy engine (src/bernstein/core/policy_engine.py)
• Build on the WAL (write-ahead log) in .sdd/runtime/ for replay capability
• Schedule cleanup operations during low-activity windows detected by the adaptive tick

Relevant Files

• src/bernstein/core/orchestrator.py
• src/bernstein/core/planner.py
• src/bernstein/core/plan_loader.py
• src/bernstein/cli/
• src/bernstein/core/wal.py

---

Backlog: road-047-wal-compaction-for-long-running-orchestrations.yaml

[chernistry]

Implemented.