feat(compact): record checkpoint request usage telemetry

[bobzhang]

Summary

First slice of the compaction production-hardening plan (R0: measure before changing behavior).

The checkpoint summary request resends the whole projected history, making it the most expensive single request an agent issues — yet its response usage (including DeepSeek's prompt_cache_hit_tokens/prompt_cache_miss_tokens) and latency were discarded.

• generate_compaction_summary now returns a CompactionSummary struct: the summary content, the response usage, and the request's wall-clock duration_ms.
• Both compaction_finished (manual /compact, sessions compact) and auto_compaction_finished (ceiling checkpoints, compact-on-finish) log usage and duration.
• Wire-compatible: TUI and desktop decoders match on the event key and ignore unknown fields; absent usage still decodes as zeros.

Why

Grounds the next steps of the plan with data instead of speculation: whether prefix-cache pricing offsets the summarizer's isolated system prompt (a26456b), what checkpoint latency actually costs a pressured turn, and how summary size tracks coverage.

Testing

• New end-to-end test: ceiling turn against the mock server with fabricated cache counters in the checkpoint body; asserts auto_compaction_finished carries the usage fields and a non-negative duration (captured via the existing MemoryLogHandler pattern).
• moon test: 1045/1049 — the 4 failures are the known-flaky agent_tool/shell output-limit timing tests, failing on clean main in this checkout too.
• moon check --deny-warn clean; moon fmt + moon info applied.
• subal (codex) review of the commit: no actionable regressions found.

🤖 Generated with Claude Code