[rudycelekli]

Lexical timestamp sorting puts snapshots with different UTC offsets in the wrong chronological order, changing the evidence ingestion sequence for temporal benchmarks. Compare valid dated snapshots by their instants and retain the filename fallback for undated corpora.

Validation: the actual CLI fails the offset-order regression on baseline and passes both dated and undated regressions with the fix. Synthetic local corpus only; no downloaded data or live API calls.

[WaylandYang]

Thank you for going through the codebase this carefully. We are taking eight from this batch (#1140, #1143, #1147, #1148, #1157, #1163, #1166, #1168), each a behaviour someone using Utopia would hit.

We are closing this one: it guards against an input we have not seen in use, and we keep the code and its tests to what the product needs today. For the scripts under scripts/, CI runs only scripts/bench/lib.test.mjs, so the added tests would not run.

If you see this happen in a real deployment, please open an issue with the case and we will take it. For future sweeps, one issue listing the findings first would let us pick together before you write the fixes.

Thanks again for the time you put into this.