feat: plan maintained populations through summary lifecycles

[zzylol]

Stacked on #485.

Rebuilt into the linear stack on main. New integrate: commit(s) fold in integration-branch resolutions this PR needs on top of the earlier stack: "read the population test frontier from lifecycle timing". Conflict resolutions are recorded in the messages of: "test: carry a chosen population lifecycle to physical compilation"; "docs: describe maintained populations as lifecycle-planned state".

Why

The lifecycle layer is meant to be the only source of node timing, but maintained populations were outside it: execution_timed_dag refused any plan containing MaintainPopulation (UnplannedMaintainedState), and MaintainedPopulationStrategy hard-coded the population at ingestion time.

What

• enumerate_summary_maintenance_lifecycles now emits one deployment for each unique maintained population that does not feed a SummaryAgg. It offers the usual alternatives (Ephemeral, Prepared, Shared, ContinuouslyMaintained), each costed through the caller's existing lifecycle hooks. Unknown cost stays unknown.
• A population that feeds a SummaryAgg is treated as that state's input and follows its timing, including when a readout also reads the population directly.
• select and execution_timed_dag time populations the same way as summary state. A retained population runs at ingestion together with its raw input. An Ephemeral population is rebuilt from the raw source at query time.
• The validator accepts either timing for MaintainPopulation. The structural contracts still hold: the population must read its matching raw input, and its readout must run at query time over a population that supports it. The strategy writes query time only as the initial layout.
• No new public types. deployments and the cost hooks may now carry MaintainPopulation nodes, and the docs say so. Cost models that price every node uniformly will now also price populations.

Before this PR

For topk by(job)(1, m) built by MaintainedPopulationStrategy, enumeration returned 0 deployments, and execution_timed_dag() returned Err(UnplannedMaintainedState). The population was always at ingestion.

After this PR

The same plan enumerates 1 population deployment.

• ContinuouslyMaintained: the timed DAG is raw@ingest → population@ingest → readout@query. The compiled candidate stores the population in precompute.
• Ephemeral: every node runs at query time and precompute is empty.
• Both choices return the same TopK (chosen_population_lifecycle_decides_precompute_contents).
• With a model where retaining is cheaper, Planner's own selection still keeps the population at ingestion.
• With DefaultCostModel (and SummaryMaintenanceCostModel, which has no population evidence), nothing about populations is selected.

Validation

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace --no-fail-fast: 1380 passed. The new lifecycle tests were first run against the base code and failed there.
• An independent reviewer agent examined the change. Its findings are fixed in the last commit: order-dependent handling of a shared population, which had also let its own lifecycle choice be overridden, plus a missing test and stale docs.

🤖 Generated with Claude Code