Resuming a session can reuse a precomputed compaction summary instead of recompacting

What

• A new path lets the agent loop reuse a previously precomputed compaction summary, built ahead of time and cached by session id, model, and boundary, instead of running compaction fresh.
• The cached summary is validated before use: its age, model match, and token growth are checked, with rejection reasons including session_mismatch, model_mismatch, bad_timestamp, too_old, grew_too_much, shrank_too_much, and preserve_uuid_missing. A successful reuse emits tengu_precomputed_compact_rehydrated.
• The session-resume warning dialog now shows a "Resume from summary (instant, recommended)" label instead of the usual "Resume from summary (recommended)" when a precomputed summary is available (hasPrecomputedSummary).
• The resume flow, gated by tengu_gleaming_fair_reuse (extending tengu_gleaming_fair), can skip recomputation entirely, or detect that a precomputed summary has gone stale.

Why Recomputing a full conversation summary on every resume is slow; reusing a precomputed one when it's still valid should make resuming a session noticeably faster. The feature is currently gated off.