refactor: execute Planner precompute DAGs for raw heaps - #800

Stacked on #797.

Why

After #796/#797 every raw output except CMS/CountSketch heaps executes Planner's precompute graph. Heaps stayed on the backend's per-sample kernel (the last item/weight interpreter in raw_dag.rs) because stored heap readout could only decode the legacy integer heap kernel, not Planner's weighted-frequency state.

What

• Storage (backend-owned format): a Planner heap window is stored as a WeightedFrequencyV1 frame (sketchlib WeightedFrequency bytes, persisted with encoding tag 6). One helper picks a frame's encoding from its state for live appends, revision views and derived publications; carry-in checks treat it as a full frame.
• Readout: SummaryState::WeightedFrequency decodes such frames for their catalog family and matrix shape, merges them, and ranks items under the legacy heap key (;-joined parts; a canonical series identity renders as its series key). PointCount/total over these frames is reported unsupported.
• Execution: heaps build through the Planner graph like every other raw output; the backend item/weight interpreter is deleted. A heap's empty state is a Planner heap, so revisions install for heap outputs.

Before this PR

topk(3, count_over_time(top_endpoint_qps[5s])) with a CMS heap: each sample was keyed by the backend (series key string) and added to a CountMinSketchWithHeapAccumulator.

After this PR

The same output is built by Planner's KeyedSummaryBuild over the pane batch (item = LabelIdentity of the series' label set); registered_temporal_topk_cms_heap / _count_sketch_heap return the same ranked items (top_endpoint_qps{endpoint="gamma"} = 200, …) and range results as before.

Behaviour differences

• Heap scores are exact Float64 weights; the legacy CMS heap rounded each weight to an integer count.
• New heap windows are WeightedFrequencyV1 frames; a range mixing them with legacy heap frames of the same output fails with a family mismatch (a new plan generation starts new outputs). Older binaries do not read tag 6.
• Heap items that exclude identity labels (topk by (…) identity items) are rejected at install until readout can restore their group labels.

Remaining (backend computation still present)

• RawDagProgram::empty_state constructs empty family states (heaps via Planner's constructor; other families via the backend kernel factory) for known-empty counter panes and the revision codec check.
• PrecomputeMaterialization computation fields (aggregation_type, parameters, spatial_filter, aggregated_labels, sample_update_rule) are still cross-checked against the DAG at install; removing them changes plan identities.
• OTLP-delivered sketch panes are still merged with the kernel's merge_with, and late admitted corrections are coalesced with merge_with (both surface errors).

Validation

cargo fmt --all -- --check; cargo clippy --workspace --all-targets --locked -- -D warnings; cargo test --workspace --locked --lib; cargo test -p data_plane --locked --test asapquery_compatibility_process_e2e -- --test-threads=1 (26 passed, incl. both registered heap TopK tests); cargo test --workspace --locked --no-fail-fast (every test target) passes at the top of the stack. New tests: heap frame decode/merge/rank and shape rejection, heap empty state is a Planner heap frame. Reviewed by a separate reviewer agent; findings addressed in the last commit.

🤖 Generated with Claude Code