feat(cost): rank complete streaming deployments

[zzylol]

Why

A streaming lifecycle must be ranked against raw recomputation as a complete alternative. Selecting a locally cheap lifecycle first can hide bootstrap, source-read, readout, retained-state, or physical-implementation costs.

What

• Enumerate complete provider-owned physical alternatives for each summary candidate.
• Cross product each physical alternative with every legal lifecycle combination.
• Estimate and rank the complete horizon totals atomically against the same raw baseline.
• Return the selected provider physical_plan_id together with lifecycle decisions.
• Retain target and SummaryNode Rc values behind pointer-keyed bindings so allocator address reuse cannot alias evidence.

Correctness details

• Ephemeral direct build performs a full snapshot build and source read for every evaluation; it has transient working state but no retained state.
• Prepared delayed bootstrap charges its source read when the bootstrap actually occurs.
• The raw physical DAG validates and charges every reachable scan independently, while allowing several scans of the same semantic source and ignoring unreachable decoy nodes.
• Evaluation multiplicity comes from the same recurrence/horizon function used by physical costing.
• Missing, conflicting, or incomparable evidence makes the complete alternative unavailable.

Validation

• cargo test --workspace
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo fmt --all -- --check

Stack: depends on #335; #337 documents the planner/downstream boundary.