Misleading Recency in Temporal RAG: Intent-Conditioned Evidence Packets Under Fixed Retrieval

A query about Python 3.8 can retrieve both the requested tutorial and a newer Python 3.12 deprecation notice: publication chronology favors the notice, whereas semantic coherence with the requested version favors the tutorial. Correct handling is not a single recency correction. Wrong-Newer, Missing-Timestamp, Branch-Specific, and Partial-Rollback conflicts require different decisions about whether retrieved endpoints should be rejected, traversed, or retained together for a requested state, history, branch, or comparison. ChronoPacket makes that decision with typed claim relations and a bounded, intent-conditioned packet objective: supersession collapses revisions for state queries, while comparison and branch intents preserve the requested alternatives. To isolate packet commitment from retrieval, our 1,800-query protocol admits only examples whose shared top-$20$ pool contains a complete support set and identifies 240 misleading-recency test cases by timestamp-first incompatibility before any learned-selector output is observed. Diagnostic gains over full-pool reading are 5.9 points on Wrong-Newer, 4.6 on Missing-Timestamp, 6.3 on Branch-Specific, and 4.1 on Partial-Rollback queries. Across all misleading-recency cases, ChronoPacket reaches 66.4 $\pm$ 0.8% accuracy versus 65.1 $\pm$ 0.8% for Reader-Conditioned Selector (paired difference 1.3 points, 95% CI [0.4, 2.3]); the no-abstention control also lowers stale citations from 9.1% to 8.1% and unresolved contradictions from 10.7% to 7.0%, while answered-query median post-retrieval latency falls from 171 to 61 ms. Collapsing coexistence removes 8.7 Branch-Specific points and shuffling validity intervals removes 6.6 aggregate points; source and entity holdouts, independent transfer cases, faster incident updates, and BGE-M3 retrieval preserve the ordering. Under fixed answer-bearing access, these results identify intent-sensitive conflict topology as the decision unit for choosing which retrieved revisions should jointly govern a temporal RAG answer.