refactor: execute Planner precompute DAGs for live raw ingest

[zzylol]

Stacked on #795.

Why

Live raw ingest evaluated summary computation in the backend: RawDagProgram interpreted each SummaryAgg's SummaryInputExpr per sample, and merge_*_panes_for_window "merged" panes (always a single pane) while silently dropping merge errors. Under the agreed layering, Planner's physical compile owns computation; the backend owns ingestion, routing, pane assignment, lateness and storage.

What

• Repin Planner to integration/planner-for-backend@ca422df, which merges feat(physical): compile summaries over raw-sample precompute boundaries ASAPPlanner#488: precompute::compile accepts a raw time-series scan as a boundary (rows [$population label map, $timestamp, value]).
• Control plane retains, for every raw stored output, Planner's compiled precompute graph in InstalledPostAsapDag::native_programs. Validation admits raw sample inputs (bound as ingestion inputs); pricing is unchanged (raw outputs are still priced by their stored state).
• Live worker keeps each open pane's admitted samples; when the backend closes a pane (watermark, flush, shutdown) or forwards a late sample, RawDagProgram::build runs the installed Planner graph over that pane as one typed batch (time-ordered, one sample per series and timestamp). Pane assignment, completeness and lateness are unchanged in the worker.
• merge_panes_for_window / merge_sketch_panes_for_window are replaced by a plain remove.

Before this PR

sum_over_time(asap_demo_gauge[5s]): each sample went through RawDagProgram::apply → update_single(weight) on a backend-owned updater; the Planner DAG only chose the kernel.

After this PR

The same output installs with its Planner graph (Project → SummaryBuild → ScopeTimestamp) and each closed pane executes it. For the 16-query probe (exact / Epsilon / EpsilonDelta, 75 raw outputs), every raw output now compiles through Planner (11/75 before, via the general compiler). live_panes_execute_planner_dag_with_identical_states feeds one batch per scrape through the worker for every distinct raw output of those plans (Sum, Count, Min, Max, Rate, Increase, DDSketch, grouped variants; ≥50 outputs) and asserts byte-identical stored states versus the previous per-sample kernel path.

Behaviour differences

• A pane's state is built when the pane closes, not per sample; open panes hold their samples (memory is O(samples in open panes), each build bounded by Planner's default 64 MiB limit).
• Samples in a pane are delivered in timestamp order with one sample per series and timestamp (first kept). Counter results no longer depend on arrival order, and a resent identical-timestamp sample is not double-counted.
• Plans installed without raw Planner graphs (compiled before this PR) no longer install.

Remaining

• CMS/CountSketch heaps still use the backend heap kernel (stored heap readout does not yet decode Planner's weighted-frequency state); the item interpreter in raw_dag.rs remains only for them.
• Backfill and Remote Write revisions still use the per-sample interpreter; the next stacked PR moves them.
• PrecomputeMaterialization computation fields (aggregation_type, parameters, spatial_filter, aggregated_labels, sample_update_rule) are still cross-checked against the DAG and are not removed here.

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); cargo test -p control_plane --locked --test native_rate_topk. New tests: live/Planner equivalence, counter ordering (fails without the ordering fix), late forwarded correction, raw output requires its Planner graph (fails on the base). Reviewed by a separate reviewer agent; findings addressed in the last commit.

🤖 Generated with Claude Code