feat(summary_executor): support SketchQuery::TopK via a two-shape Value

[zzylol]

Summary

Step 1 of 3 toward retiring the legacy sketch_reducer.rs query path in favor of the plan-based SummaryExecutor merged in #411 — the user wants no duplicate/parallel query-execution implementations, and this was the one functional gap (of two) blocking a full cutover.

SummaryExecutor::Value was Vec<(i64, f64)> — one scalar per point — which can't represent SketchQuery::TopK's answer (K ranked (item, count) pairs per point), so it errored Unsupported. Value is a fully deployment-opaque associated type (no bounds in the upstream ASAPController trait), so this is 100% local to data_plane:

• New SummaryValue::{Points(Vec<(i64,f64)>), TopK(Vec<(i64, Vec<(String,f64)>)>)}, replacing the old Value type.
• readout_cumulative/readout_per_window each branch once on matches!(query, SketchQuery::TopK) to decide which shape to produce (merge loop itself is unchanged).
• New topk_ranked(rs, k) helper: reads SummaryState::topk_items() (already existed), sorts descending by value (load-bearing — the heap's backing array isn't actually ordered despite its own doc comment's claim), caps at k.
• No new merge/sketch-math anywhere. merge_same_family already reconciles heaps correctly across sids (asap_sketchlib's CountMinSketchWithHeap/CountSketchWithHeap::merge re-queries every candidate key against the merged matrix) — top-k readout just reads the final merged state's heap, reusing the exact same cross-sid merge pipeline every other query already goes through.

Blast radius confirmed via full-crate grep before starting: this executor isn't wired into engine.rs's live ASAPQueryEngine yet, so nothing outside this one file/its own test module is affected.

Test plan

• cargo test -p data_plane --lib summary_executor — 15 tests (12 existing + 3 new), all passing:
  • single_cms_with_heap_sid_topk_readout_sorted_and_capped — proves sort+cap.
  • two_cms_with_heap_sids_same_group_topk_merges_cross_sid — two sids with disjoint keys, proves the merged answer contains both — the actual cross-sid merge proof.
  • per_window_matrix_topk_produces_per_window_ranked_lists — the one genuinely new code path (readout_per_window's TopK branch), two windows with different top keys each.
  • topk_query_against_non_heap_sketch_is_unsupported (renamed from ..._is_explicitly_unsupported_not_silently_wrong) — still errors for a heap-less family; now documented as a family limitation, not "unimplemented."
• cargo test -p data_plane --lib — full lib suite: 911 passed, 2 ignored (908 + 3 new).
• cargo build -p data_plane — clean.
• cargo clippy -p data_plane --lib --tests — zero warnings in the touched file.
• rustfmt scoped to the one touched file.

Not in this PR (separate, later steps per the user's own sequencing): wiring SummaryExecutor into engine.rs's live ASAPQueryEngine, the named-key PointCount gap, and deleting sketch_reducer.rs.

🤖 Generated with Claude Code