fix: generalize summary cost evidence across arrival modes

[zzylol]

Addresses the general evidence-model question raised in #528. Kept separate because analytical workload coverage goes beyond #511/#528's operator/scalar representation design. Based on main, with no #528 commits included.

Before this PR: StreamingSummaryInputs::from_workload rejected fixed snapshots, even though the lifecycle planner and comparison scope already supported AtRest. Generic operator evidence types were named as if they were stream-specific.

After this PR: a fixed 10-row snapshot can build a shared summary once and answer five evaluations with zero arrival updates. The actual lifecycle planner selects a completely costed at-rest summary against its matching raw scan. The same SummaryOperatorEvidence and SummaryNodeEvidence API serves continuous ingestion.

• Derive zero arrivals for AtRest; require fresh snapshot cardinality and retain mandatory fresh rate evidence for continuous ingestion.
• Reject mismatched workload/scope arrival modes, contradictory at-rest rates, and directly supplied raw/summary evidence that violates those semantics.
• Rename the evidence/input/window/alternative types without legacy aliases; bump the analytical estimator version to summary-maintenance-resource-v2.

Mixed/Unknown still fail explicitly: the workload schema does not separate backlog and arrivals. This PR does not add physical computation-placement policy (#530). #528 owns removal of “readout” and unified operator names; those representation edits are intentionally absent from this main-based diff.

Validation: regression reproduced UnsupportedDataArrival(AtRest) before the fix; all 1,524 workspace tests pass on Rust 1.99, including 43 summary-cost tests. Formatting and workspace/all-target clippy with warnings denied pass. Tests cover actual lifecycle selection, missing/contradictory evidence, direct-provider bypass attempts, and existing continuous/mixed behavior.