feat(compaction): opt-in LLM semantic compaction summarizer (M3.2c / §1)

[wusijian007]

Add the long-designed semantic-compaction seam. Where the default compactMessagesTiered shrinks stale messages in place (deterministic, pointer- izing the whales), the new compactMessagesWithSummary REPLACES the whole stale region with a single LLM-written recap — recovering the reasoning/prose that pointers cannot. Root task + recent window stay verbatim.

Boundary safety: the recent-window start snaps earlier past any leading tool_result so a kept tool_result never has its originating tool_use dropped (no orphaned pairing). When there is no accumulated stale history yet (e.g. one big recent tool_result), it returns untouched and leaves the pathological case to the reactive prompt_too_long net — semantic compaction condenses HISTORY, not the recent window.

Opt-in and non-breaking: the deterministic pointer path stays the default. The proactive-compaction block branches on QueryOptions.compactionSummarizer and marks the reset as proactive_compaction_semantic. It is one amortized cache reset like any compaction (invariant #1); non-determinism is gated behind a scripted fake in tests (invariant #2).

• core: compactMessagesWithSummary + MessageSummarizer (context.ts), QueryOptions.compactionSummarizer, createModelCompactionSummarizer (model- backed summarizer kept in core next to the loop).
• cli: myagent agent --semantic-compaction (builds the summarizer from the agent's own client), help/usage text.
• eval: EvalTask.initialMessages + new semantic-compaction task (seeds a stale whale, asserts the summarizer seam fires + tokens drop); fingerprint updated to tasks=10 turns=21 in=14350 out=835.
• tests: 3 unit (recap replacement, boundary snap, under-target no-op) + 1 loop integration (proactive path uses the injected summarizer).
• docs: CLAUDE.md query-loop step 7 + roadmap M3.2c marked delivered.

214 tests pass (+4).