LTM prompt-cache, pin, and meta-distillation stability

Keep stable preference/entity LTM in system[1] byte-identical throughout a session, including idle resumes; defer newly curated stable knowledge until the next session. Carry changing context through durable deltas at a frozen safe tail position, not by rewriting stable system prefixes. Treat pins as diff baselines, compare entries by ID/content regardless of order, and keep persisted ltmCacheText synchronized with ltmPinText. Require isCacheWarm before preserving caches or skipping compaction; expired caches are cold. Avoid re-pinning when entry sets are unchanged, preserve uncertain warm state, and keep baseline metadata off the wire. Run meta-distillation only at an already-cold idle-resume boundary and track warm prefix rewrites separately.