feat(cost): rank complete physical plan alternatives

[zzylol]

Why

Logical candidate ranking is valid only when raw and post-ASAP alternatives cover the same target, source contents, predicates, event-time range, recurrence, and workload horizon. Partial plans or statistics from different snapshots must not enter one comparison.

The implementation must also distinguish the physical plan being estimated from the analytical method used to estimate it. Physical operator identities and evidence are inputs; CPU, memory, and I/O formulas are analytical predictions rather than runtime measurements.

What

• Add PhysicalPlanCostModel at the final planner-selection boundary.
• Capture one immutable, versioned PhysicalEvidenceSnapshot per target; blank or whitespace-only versions fail before any evidence lookup.
• Lower the raw target and every rewrite into complete physical DAGs; require a complete provider-bound physical DAG for summary candidates.
• Compare separate CPU operations, peak-memory bytes, and source-read bytes with one validated calibration.
• Cache the raw baseline by retained target identity and consumer count.
• Make missing, stale, inconsistent, unsupported, or more-expensive candidates unavailable so the raw plan remains selected.
• Bypass local CSE structural heuristics when a complete-plan model is active.

Mechanism

The modules expose four distinct responsibilities:

• query_physical_lowering recursively maps QueryExpr and logical rewrites to physical operators.
• physical_operator_statistics defines the per-operator catalog/runtime evidence contract.
• analytical_cost applies algorithmic formulas to an EvidenceBackedPhysicalDag and composes CPU, peak-memory, and source-I/O estimates.
• physical_plan_cost_model compares complete raw and replacement physical alternatives at the planner-ranking boundary.

SummaryExpr candidates are bound by PlannerPhysicalPlanProvider, because implementation-specific summary operators and retained state require physical evidence. Raw and replacement DAGs are estimated against the exact same ComparisonScope and snapshot version. EvidenceBackedPhysicalDag is the single shared physical DAG representation; query lowering does not publish a second PhysicalDag alias. No structural node count or optimistic zero is used by this analytical selection path.

The self-contained design and formula contract are documented in docs/design_docs/asap-aware-mapping/analytical-resource-cost.md; physical lowering integration is documented separately in physical-plan-integration.md.

Validation

• cargo test -p asap-aware-mapping
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo fmt --all -- --check

Stack: based on #328; #331 adds streaming-summary resource modeling.