[zzylol]

Why

With a candidate heap, WeightedFrequency::update is the per-sample hot path of a keyed top-k build. Each call cloned every retained candidate (its identity tuple and its hash key) to rebuild the heap, and scoring each candidate collected its row estimates into a fresh Vec. At capacity 100 that is about 200 heap allocations per sample.

What

update returns the same state as before, about twice as fast at capacity 100. No interface changes.

Stacked on #226 (shape and weight checks), which is stacked on #225. Retarget to main after those merge.

How

• The update moves the candidates out of the heap (std::mem::take(..).into_vec()) instead of cloning them, and adds the new key only when it is not already retained. Before, the duplicate was added and then dropped by the sort-and-dedup in retain, which keeps the earlier entry, so the result is the same.
• Scoring takes a caller-owned buffer: retain reuses one buffer for every candidate's CountSketch row estimates, and Count-Min takes the row minimum with no buffer.
• Every retained score is still re-read after each update, since the update may have moved cells the candidates share.

Key code interfaces added / modified

None. (estimate_key_with is a private helper.)

Before this PR

Release build, 200,000 updates over 1,000 two-label Utf8 identities, width 2048, depth 5:

algorithm	capacity	per update
CMS	100	65.9 us
CountSketch	100	76.9 us
CountSketch	0	518 ns

After this PR

algorithm	capacity	per update
CMS	100	32.6 us
CountSketch	100	43.2 us
CountSketch	0	509 ns

The remaining cost is rescoring: each update re-hashes every retained key (capacity x depth hashes) and re-sorts the candidates. Caching each candidate's row cells (not serialized) would remove the hashing at capacity x depth x 16 bytes of memory; that is left as a follow-up.

Verification

• planner_v1_snapshots_remain_compatible (golden bytes) and every existing weighted-frequency test pass unchanged.
• New retained_scores_track_every_update: on an 8 x 3 shape with 7 keys and capacity 3 (heavy collisions), after each of 60 updates every retained score equals the point estimate and no key is retained twice. It passes before and after this change; it guards the rescoring behavior the optimization must keep.
• The table above comes from a temporary #[ignore] benchmark test run with cargo test --release, removed before commit.
• cargo fmt --all -- --check; cargo clippy --workspace --all-targets --all-features --locked -- -D warnings.

[zzylol]

Closing: main now provides f64 Count-Min and Count Sketch with heaps (#232), ASAPv1 serialization as the only format (#241) and TupleKey group-by keys (#237), so WeightedFrequency's own kernel and ASAP-WFREQ-1 format are redundant. The remaining gaps (shape and weight checks, an atomic checked f64 update, a full clear, opt-in rescoring of retained candidates, and estimated_memory_bytes) move to new PRs on main's heap sketches. The old head is kept on backup/perf/weighted-frequency-update-without-candidate-clones-20261020-161432.