feat: add native physical operators and shared DAG runtime - #474

Stacked on #473.

Part 3 of the #462 split.

Before this PR

After #473, the Planner has kernels but no way to execute a physical DAG. Every deployment implements relational operators, summary build/merge/readout and execution separately.

After this PR

One shared runtime executes a DAG:

• plan, runtime and sources: physical DAGs, the RunContext, shared producers, isolated runs, bounded backpressure, cancellation, tracked memory reservations, and raw-source scans with a memory connector.
• expressions and operators: projection, filter, joins and semi-joins with certified pruning, aggregate/window, sort, limit, PromQL vector binary/window, temporal panes, current-series snapshots, and summary build, merge and readout.
• Summary readout is typed. ReadoutQuery::Sketch(SketchQuery) reuses the logical IR's query. ReadoutQuery::Exact(ExactReadout { statistic, lookback_ms }) carries the logical counter window, which is resolved to the run's evaluation range at execution.
• Operators derive serde, and deserialization validates their input/output contracts (operators/unchecked.rs). The byte format is left to the deployment.
• readout::exact_readout: merge exact panes and read one population (None = absent).

Operators are built directly. No physical planner yet.

Notes

• Runtime and operators come as one PR because they share one operator enum and the runtime dispatches on it.
• Four crate-private helpers used only by the physical planner arrive in feat: compile Planner selections into physical DAG candidates #475.

Validation: cargo test -p asap-physical-operators (55 passed) and cargo clippy --workspace --all-targets -D warnings.

Split stack: #472 (merged) · #473 · this PR · #475.

🤖 Generated with Claude Code