feat: place summary state by backend-priced lifecycle choice

[zzylol]

Stacked on #793.

Why

Placement (precompute a summary vs. compute it at query time) was decided by enumerating materialization-key masks over counter/max leaves and fully compiling each mask as its own quote candidate. The owner-approved layering makes placement a summary-maintenance lifecycle choice priced by the backend's own costs.

What

• Repin ASAPPlanner to 9a13ae4 (integration/planner-for-backend; feat(precompute): separate idle and deadline closure #486/PromQL range warm path ignores step_ms and differential E2E does not detect it #487 PromQL Fallback physical coverage). Lockfile changes only the rev.
• compiler/placement.rs: for every unique summary state (after sharing), Planner's enumerate_summary_maintenance_lifecycles lists alternatives; one backend CostModel prices them — build, maintenance_per_update × ingestion rate, read × reads, retention, retirement — with demand from all consumers, so shared state is charged once. ContinuouslyMaintained wins by default.
• New lifecycle_costs.store_per_byte_second (default 0): retention adds estimated state bytes × retained panes (window ÷ evaluation interval) × partitions (input cardinality for per-series/grouped state) × price. An unpriced alternative never displaces retention.
• Ephemeral is offered only when raw series are readable from Prometheus at query time (mixed PromQL, backend-local target, not require_backend_local_execution). A query rebuilds all of its states or none. With no state left it runs natively: Planner's PromQL Fallback lowering compiles the query once over raw-series inputs, bound to range-selector Scans (the feat: bind query-time raw series inputs to the Prometheus endpoint #792 raw source). Otherwise counter/max leaves are externalized as before.
• Mask enumeration (materialization_candidates.rs, enabled_materialization_keys, MaterializationSearchCoverage) is removed; the inventory is the selected candidate plus the native exact alternative (plus existing forests, removed in the next PR).
• Manifests price a native program's raw Scan as an exact_backend source; decisions are recorded as deployment.lifecycle_placement trace events.

Before / After

quantile_over_time(0.99, m[1m]), default snapshot costs:

store_per_byte_second	Before	After
0	KLL maintained	KLL maintained (continuously_maintained)
1.0	field rejected	no materialization; backend reads {__name__="m"}[60000ms] from Prometheus and computes the quantile

Startup of the 44-query issue workload (issue_workloads_execute_warm_at_successive_evaluations snapshot, process start → /api/v1/health, debug build): 69.3 s → 38.6 s (164 → 84 candidates). Still above the 30 s budget; the next PR fixes it.

Behaviour differences

• Candidates no longer include per-mask variants; the report drops materialization_search_coverage and candidate identities no longer hash a mask.
• Sparse or costly maintenance can now move a state to query time. Window-layout tests that exercise retained state pin it with require_backend_local_execution.
• Quote/ERP path unchanged: manifests are exported and matched from the same inventory, by deep equality. Quotes no longer compare "retain vs. exact" per leaf; that choice is the lifecycle cost's.
• Ephemeral is priced as rebuilding the selected state per read with the backend unit costs; it is realized as an exact query-time program over raw series (or exact subtrees), which is at least as accurate.

Validation

• cargo fmt --all -- --check, cargo clippy --workspace --all-targets --locked -- -D warnings
• cargo test --workspace --locked --lib, cargo test -p control_plane --locked --tests (new tests/lifecycle_placement.rs: cheap/expensive store, no raw source, shared state priced once)
• cargo test -p data_plane --locked --test asapquery_compatibility_process_e2e -- --test-threads=1: 25/26; new lifecycle_placement_process e2e (store price flips placement) passes; issue_workloads_execute_warm_at_successive_evaluations fails its 30 s readiness budget, as it does on chore: repin Planner integration and take over storage formats and panes #793 (69 s there).

🤖 Generated with Claude Code