refactor!: align DAG APIs and share candidate graphs

[zzylol]

Stacked on #508.

Rebuilt into the linear stack on main. The 12 docs commits before the former build: merge were written against main; each resolves docs/design_docs/architecture/input-output-workflow.md to that commit's version, as integration merge e59640f did. That merge is no longer needed. As in e59640f, this PR's text of that file replaces the #485 sentence about lifecycle-assigned timing and the #508 renames there.

Why

The architecture's DAG names did not identify the Rust APIs, and callers had to compose a separate lifecycle enumerator and physical candidate wrappers between layers. The public pipeline now uses named candidate collections end to end, with shared logical graphs and physical compilation.

What

• Stack on refactor: rename PlanSpace to CandidatePostASAPDAGs #508 and retain the workload facade already on main.
• Align individual graph names: PreASAPDAG, PostASAPDAG, PhysicalDAG; distinguish nodes and runtime-bound BoundPhysicalDAG. Remove compatibility aliases for the old names.
• Expose lower_pre_asap_dag_candidates returning CandidatePreASAPDAGs with normalized workload entry IDs.
• Give CandidatePostASAPDAGs a timed stage. with_timing_for_root encapsulates logical realization, lifecycle enumeration, shared indexing, assignment budgets and lazy iteration. Existing assembled graphs use from_post_asap_dag. The lifecycle enumerator is now crate-private.
• Make compile_physical_dag_candidates directly consume the timed logical collection and return CandidatePhysicalDAGs<Id>. The physical collection owns shared compiled graphs and cut descriptors; the public PhysicalDAGCandidate wrapper and tuple-based compilation API are removed.
• Preserve workload/candidate IDs, lifecycle metadata, logical assembly rejections, timing failures and compilation errors. Unknown costs stay unknown.
• Reuse physical compilation only when logical identity, Binary timing, input contracts and requested roots agree. Materialize typed precompute/query cuts on demand.
• Reuse the existing lifecycle enumeration, graph validator, operator compiler and cut implementation. Explicit selection helpers remain available.
• Update the implementation plan, DAG naming table and breaking API migration guide.

Before this PR

A deployment used Rc<QueryExpr>, Rc<SummaryNode>, PostAsapDag and CompiledPhysicalDag for graph stages. It separately enumerated lifecycle choices, attached timing and assembled physical candidate wrappers.

After this PR

CandidatePreASAPDAGs → CandidatePostASAPDAGs → CandidatePostASAPDAGs with timing → CandidatePhysicalDAGs

A caller invokes logical.with_timing_for_root(...), then passes the resulting collection directly to compile_physical_dag_candidates(...). It can inspect physical graphs and diagnostics through iter() and obtain typed execution cuts with materialize(index). Compatible assignments share one compiled graph; no intermediate stage selects a winner.

Generation can be scoped to a workload root without expanding independent query combinations. Deployment selection must still coordinate shared state across roots. Missing window evidence must be resolved before installation. Downstream consumers must migrate the breaking API names before repinning.

Validation

• cargo test --workspace: 1509 passed, 0 failed, across 91 test groups.
• cargo clippy --workspace --all-targets -- -D warnings and formatting checks passed.
• Regressions cover collection-to-collection handoff, workload entry identity, lazy assignment budgets, unknown costs, graph sharing, Binary timing, contract/root-sensitive compilation reuse, transport equivalence and error preservation.
• Alignment documentation links/fences and git diff --check passed.

🤖 Generated with Claude Code