v0.20.2

The consumption half of the long-session fix (v0.20.1 removed the 300-turn import cap): every LLM reader of session history now iterates context-budget-sized chunks via the new llm_chunking map-reduce layer — excerpt truncation removed (no more "…sessions omitted" data loss), single unbounded narrative/decision prompts split into PART i/N calls with hierarchical merge, session-graph chunking made size-aware so one huge tool-dump turn can't overflow the window. --max-turns is a scan memory guard only — it no longer bounds what the model reads. Verified: 600KB of history → 82 bounded calls → one merged narrative, zero loss. Full notes: docs/release-notes/v0.20.2.md (7 translations included).