[lmeyerov]

Summary

• Make empty temporal bounds raise typed E302 instead of silently resolving to NaT.
• Narrow naive datetime parsing to pd.Timestamp, reject pd.NaT explicitly, and localize only typed aware bounds.
• Cache optional pytz.exceptions discovery and run it only in the aware-localization path.
• Add pandas boundary and absent-dependency tests for empty timestamp/duration bounds, DST localization, and duration behavior.
• Add a compact review-skill requirement-to-evidence acceptance map, pressure-tested against the fix(gfql): preserve timezone-aware temporal equality #2234 missed-changelog, typed seam, GPU-claim, dependency, and final-diff scenarios.

Evidence

• Base: 5cb023e1ee524c1eafd10ef13e50a88e7b15100d
• Head: d468f52ef45d0d91e0b511a6339fb7dbabe4b632
• Focused new-boundary tests: 4 passed locally.
• Full temporal file on CPU host: 82 passed, 3 skipped, with 18 inherited cuDF failures caused by CUDARuntimeError: cudaErrorNoDevice during fixture construction; these are not counted as product evidence.
• DGX exact-source temporal + filter-by-dict: 118 passed, 3 failed; all three failed during fixture construction with CUDA std::bad_alloc while an external 60 GB process occupied the sole GPU. This is not counted as GPU proof.
• Ruff: passed.
• Exact-head CI/CodeQL/RTD to be updated here.