types: evaluation-relative coverage time and mergeable summary families

[zzylol]

Restacked on #646 (coverage = definition + selection, derived). Chain: … #594 → #593 → #592 → #591 → … (Phase C tail, #528).

Step 3 of the #580 plan (item A prerequisites). Part of #509.

Why

Stage 1 tumbling windows (W1) compose N SummaryAgg panes with one SummaryMerge. #646 already derives where a pane sits in time: TimeRange(w) over TimeShift(s) gives the selection's relative time (-(s+w), -s], and a merge is accepted only for equal definitions and disjoint selections. Two facts are still missing:

• Whether a family may merge. SummaryMerge accepted any state family with matching schemas, including Rate/Increase accumulators and CMS/CountSketch heaps. W6 says merged heaps are rejected until there is an accuracy model.
• How the inputs' selections may relate. Disjointness is right for counting families, but idempotent families (HLL, exact Min/Max) give the same state when a row is seen twice, so overlapping inputs are fine (docs: ASAP primitive schema and summary coverage (definition + selection) #573 §4.4).

What

• FieldDataType::family_merges() returns true for exact Sum/Count/Min/Max, KLL, DDSketch, HLL, CMS, CountSketch and UnivMon. Everything else fails closed: Rate/Increase/IRate, CmsWithHeap, CountSketchWithHeap, KMV/Theta, Sample/Wavelet/StatModel and plain values. SummaryMerge::validate_inputs rejects a family that does not merge.
• FieldDataType::merge_relation() -> Option<SelectionRelation>, next to family_merges: Disjoint for counting families, OverlapAllowed for HLL and exact Min/Max, None when the family does not merge. SelectionRelation::Contained names what the reserved SummarySubtract will require; nothing produces it yet.
• SummaryCoverage::derive skips the disjointness check for OverlapAllowed families, and the union of identical selection boxes is one box.

The earlier version of this PR added CoverageTime::{Absolute, RelativeToEvaluation} to declared coverage; that commit is dropped because #646 derives relative time from the sub-DAG.

SummaryMerge(KLL us, KLL eu)          -> selection {region ∈ {us}} ∪ {region ∈ {eu}}
SummaryMerge(KLL us, KLL us)          -> PossibleOverlap
SummaryMerge(HLL us, HLL us)          -> accepted, selection {region ∈ {us}}
SummaryMerge over CmsWithHeap states  -> "summary merge over Sketch(CmsWithHeap…) is unsupported: the family has no sound merge"

Tests

Written first; idempotent_families_merge_overlapping_states and family_merge_selection_relation fail without the change.

• summary_merge_structure.rs: heap states are rejected and KLL states accepted (summary_merge_requires_a_mergeable_family); table tests of family_merges and merge_relation; HLL/Max states over the same rows merge and Sum/KLL ones do not.

The doc comments on SelectionRelation and merge_relation cite #573 §5.6, the doc's SummaryMerge section.

Open questions for review

• KMV and Theta merge soundly in the literature but were not on the W6 list, so they currently fail closed.

Gate

cargo fmt --all --check, cargo clippy --workspace --all-targets --all-features -- -D warnings and cargo test --workspace pass at this head.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW