feat(graph): two-tier L3 alignment thresholds + reconcile co-mention hint — same-type 0.85, cross-type 0.92

feat(graph): cross-chunk relation reconciliation — top-K co-occurrence LLM pair check

feat(graph): map-reduce KG extraction — per-chunk two-pass under key-pool concurrency

fix(graph): drop extraction max_tokens 8192→5120 — Groq free-tier 8K TPM ceiling

fix(graph): entity-dense docs — bump max_tokens 4096→8192, tighten entity description spec

fix(graph): unstarve KG extraction — reasoning_effort=medium + bump tokens + domain-neutral prompt

deploy: accept .docx and .pptx uploads

initial backend deploy