Add compaction_threshold_for_model_and_effort() with mode-aware soft
caps based on DeepSeek V4 paper Figure 9 recall-quality data:
Plan / off -> 64K (paper eval: 8K-128K)
Agent / high -> 192K (paper eval: 128K)
YOLO / max -> 384K (paper eval: 384K-512K)
Previously, the 80%-of-window rule gave 800K for V4's 1M window,
which is well past the point where MRCR MMR collapses (0.49 at 1M).
Non-V4 models keep the legacy 80% rule. None/unknown effort defaults
to agent-tier (192K).
Hunter Bown
1be18e691b