Promql cpp functions

[cherep58]

Summary

Implement C++ pushdown optimizations for PromQL functions at the storage layer. When the query engine provides SelectHints.Func, the storage can now reduce data before it reaches the Go PromQL engine — returning only the samples needed for the specific function instead of the full series.

New decorator iterators

• Aggregation: MinOverTimeIterator, MaxOverTimeIterator, LastOverTimeIterator, SumOverTimeIterator (via OverTimeFuncIterator template with pluggable handlers)
• Counter/rate family: RateIterator (rate, increase), IRateIterator (irate, idelta)
• Transform: DeltaIterator, ChangesIterator, ResetsIterator
• Downsampling: DownsamplingDecodeIterator — interval-based sample reduction

Refactored decode iterator infrastructure

• CRTP-based DecodeIteratorTrait with seek(), seek_to(), invalidate(), set() methods
• SeekResult enum for composable seek logic (kUpdateSample, kNext, kStop, kUpdateSampleNextAndStop)
• UniversalDecodeIterator extended with seek(), seek_to(), invalidate(), set() that dispatch through std::visit
• Outer DecodeIterator variant wrapping all iterator types for the Go binding layer

Function dispatch

• gperf-generated perfect hash (FunctionNamesHash) for O(1) function name lookup
• create_decode_iterator() dispatches on SelectHints.Func to construct the appropriate iterator

Go bridge changes

• SelectHints passed through cppbridge via ABI-compatible GenericSelectHints<Go::String, Go::SliceView>
• downsamplingMs parameter added to Query() and ChunkRecoder paths
• DataStorageSerializedDataIterator control block simplified: direct Timestamp()/Value() accessors replacing raw field access
• Namespace change: entrypoint::head → entrypoint::series_data

Testing

• Parametric C++ unit tests for each new iterator (boundary conditions, StaleNaN handling, time interval filtering, reset scenarios)
• Go integration tests for all dispatched PromQL functions (downsampling, min/max/last/sum_over_time, rate, increase, changes, delta, irate, idelta, resets)

[vporoshok]

Review: WindowFunctionIterator sub-interval logic vs ADR-001

Analyzed the sub-interval splitting algorithm in WindowFunctionIterator against ADR-001 (range function iteration algorithm). The code works correctly for the most common case (s < w < 2s), but has critical issues in two other cases.

Summary of findings

#	Severity	Finding
1	Bug/Critical	advance_interval() breaks for w > 2s — produces invalid interval (min > max), potential infinite loop
2	Bug/Critical	Incorrect boundaries for w < s — aggregation captures samples from gaps between windows
3	Bug/Moderate	Possible double-count for sum_over_time at interval boundaries (depends on seek_to inclusive/exclusive semantics)
4	Design	count_over_time missing from dispatch (falls back to universal iterator — correct but no pushdown)
5	Design	sum_over_time result timestamp is last raw sample's ts, not sub-interval end
6	Observation	min/max sub-interval collapsing optimization from ADR not implemented (correct but sub-optimal)