ContextManager

Orchestrates a 4-step context management pipeline (early-exit if budget is met at any step):

1. ToolDeterministicCompaction

• Run DeterministicCompressor.compressToCap() with compressToolOutputs first (subjective edits OFF).
• Goal: shrink/cap tool outputs (JSON/logs/binary) without touching user/assistant text.

2. HistoryCompression

• Run HistoryCompressor.compressToDigest(...): • Keep last K semantic blocks as is (K = config.maxSemanticBlocks). • Replace older blocks with a deterministic [HISTORY_SUMMARY] digest capped to config.summaryTokenTarget.

3. LLMHistorySqueeze

• If still over budget AND LLM enabled: • Compress the digest string only to summaryTokenTarget via LLMCompressor.squeezeDigest(). • No whole-conversation LLM compression.

4. FinalTokenTrim

• TokenWindow.trimToBudget() with headroom.
• Pin [HISTORY_SUMMARY] so it’s never dropped; pack remaining messages newest-first.