Vector-Free Salience versus Local Dense Retrieval for Grounded Longitudinal Clinical Summarisation

Longitudinal social and clinical records grow for years; handover, continuity, and audit require timelines no single visit can reread. Dense embedding RAG adds pre-decode latency, opaque rankings, and compute or API cost, while still risking missed mandatory events or wrong visit order on dated charts. This work presents a vector-free alternative: deterministic salience from catalogues, row flags, safeguarding signals, and a recency tail, then profiling, character-budget windowing, one decoder pass with at most one length repair, and hard validation of bracketed row identifiers. Real de-identified longitudinal data is hard to obtain under ethics and governance; a 100-patient cohort is a reasonable scale to stress-test pipelines without population claims. On every patient in the read-only evaluation database (10 charts, 503 case-note rows), this study compares head-to-head against dense RAG with ephemeral ChromaDB and local all-MiniLM-L6-v2 embeddings (no paid embedding API), with fixed decoder and LLM-as-judge settings. Vector-free salience yields higher mean readability (7.90 vs. 7.50), understandability (7.90 vs. 7.70), and factual credibility (6.30 vs. 5.90), and lower mean context time (0.19 vs. 34.03 s) and end-to-end latency (15.29 vs. 31.89 s). Row-id validation and grouped evidence support audit beyond lexical metrics. The approach fits clinical AI for healthy communities and innovation infrastructure under tight compute budgets in African and under-resourced settings.