feat(dag-viewer): show planner cost and benefit annotations

[zzylol]

Closes #286.

Why

The planner now produces evidence-backed CPU, peak-memory, and source-I/O comparisons, but those decisions are difficult to audit from a raw plan alone. The export and viewer need to show the selected physical alternative, its equally scoped raw baseline, the resulting benefit, and the evidence provenance without recomputing costs in JavaScript.

This PR is intentionally the top of the analytical-cost stack: it consumes the completed planner/streaming cost contracts and adds presentation. It does not define planner formulas or structural ranking.

What

• Adds structured cost annotations with units, source, baseline, delta, benefit ratio, model version, benchmark identity, and inputs.
• Exports baseline, selected, and benefit annotations for the exact candidate chosen by whole-plan global_selection.
• Adds per-query and workload-wide summaries with explicit deduplication identity.
• Annotates genuine shared-DAG edges without inventing multi-hop path costs.
• Renders graph badges plus node, edge, decision, lifecycle, and workload cost details.
• Adds a complete physical-evidence input contract through --planner-cost-json.
• Keeps the output schema backward compatible by making annotation fields optional.

How

The planner evidence document contains calibration, an exact target QueryExpr, comparison scope, exact query-node selectors with PhysicalNodeEvidence, and candidate-specific evidence. Logical rewrites are recursively lowered from their query-node evidence; summary replacements carry a complete bound PhysicalDag. Target, query-node, and candidate matching uses full structural equality, not a hash or strategy name.

Candidate generation is bound to the supplied analytical model, and final selection uses PlanSpace::global_selection. Viewer annotations are produced from the selected candidate's same complete raw-versus-candidate comparison. The exporter does not use DefaultCostModel, structural node counts, or a Top-K query shortcut in production.

Unknown fields, duplicate selectors, missing evidence, unused target/candidate/query evidence, invalid physical DAGs, conflicting identities, unsupported algorithms, or incomplete source coverage fail closed. Without a complete evidence document, --post-asap exports the raw graph only. The old --analytical-cost-json spelling accepts the new document as an alias; its former compact aggregation payload produces an explicit migration error.

group_count belongs only to aggregate operator evidence, while k belongs only to a Top-K operator. Neither is promoted into a global viewer/export input.

Before this PR

Cost explanations were not attached to the exported DAG. Earlier versions of this branch used relative structural units, but that production path and its generated fixture have been removed.

After this PR

The viewer displays only explicit modeled, measured, or unavailable annotations. It never estimates costs client-side. A complete physical evidence document can select and explain generic non-Top-K, relational, PromQL, shared-DAG, and streaming-summary alternatives; incomplete evidence preserves the raw plan.

Verification

• cargo test -p asap-aware-mapping --all-targets — 313 passed.
• cargo test -p asap-devtools --bin dag_export — 17 passed.
• python3 -m unittest discover -s tools/dag-viewer -p "test_*.py" — 21 passed.
• cargo clippy -p asap-aware-mapping -p asap-devtools --all-targets -- -D warnings — passed.
• cargo fmt --all -- --check — passed.
• git diff --check — passed.
• Independent final correctness review — no remaining blocker.

The exporter tests include JSON-to-provider-to-planner-to-annotation coverage for a generic non-Top-K DAG, selection of the numerically cheapest of two complete candidates, and fail-closed duplicate, missing, unused, unknown-field, and invalid-DAG cases.

Viewer visualization

These captures demonstrate the pre/post lanes, decision badges, graph-element linkage, edge details, and workload comparison UI. They are presentation references only; the previous structural fixture shown in the captures is no longer generated or accepted as planner evidence. Current cost values must come from --planner-cost-json and are displayed with their explicit modeled units and provenance.