refactor(ir): remove legacy scaffolding and finish tooling migration

[zzylol]

Why

After the cutover (#542) and the move to the crate roots, the legacy QueryExpr and SummaryExpr IRs have no live consumers, but they are still compiled, documented and drawn by the DAG viewer.

What

• Remove the legacy IRs. Parameter types (GroupKeys, Reduction, Source, …) move to ir/operator_properties.rs, and their users are migrated. Deleted: pre_asap/{query_expr, cse, canonicalize, resolve, schema_resolver} and post_asap/{expr, cse, post_asap_dag}; QueryExpr paths are dropped from execution_data_state, maintained_population, agg_intent, column_resolution, scalar_type_rules and pre_asap/schema.
• Remove unused API around execution data states: the functions produced_data_state, validate_execution_data_states, validate_execution_data_states_at, assigned_child_data_state, plain_schema and exact_operation_output_schema, the struct ExecutionDataStateAssignment, the enum ExecutionDataStateEdge, the error variant AmbiguousKeepPreAsap, and the unused Source::data_model and ExecutionTiming::is_query_time.
• Docs describe the unified operator IR instead of the removed legacy IRs: export docs point at ir::physical_export and name the wire version by its constant (PHYSICAL_ASAP_DAG_WIRE_VERSION) rather than a number; physical-compile-coverage.md and planner-vocabulary-migration.md drop the removed Fallback payloads; links to deleted files in the decision docs are retargeted; stale cse / HAVING / post_asap references in code comments are fixed.
• DAG viewer uses the unified operator kinds and names OperatorNode. tools/dag-viewer/generate-sample.sh passes --data-ingestion-interval-ms and --default-cost, and dag.example.json is regenerated. crates/devtools/tests/viewer_contract.rs builds one instance of every operator variant and checks the viewer's kinds against their real kind_name() strings.
• SQL collection UDFs. sql/collection_planning.rs becomes collection_udfs.rs (CollectionUdf, CollectionFunction, MAP_UDF_NAME), with a module doc saying these are planning-only DataFusion UDFs for ClickHouse's Map / List / Struct functions that DataFusion never executes (from the feat(sql): lower SQL to unified operator and scalar IR #539 review).

How

Migrate first, then delete: every remaining user of a legacy type is moved to its ir counterpart in the same commit that deletes the type. Tests that covered behavior still present are ported to the IR: maintenance_binary_accepts_only_well_typed_series_identity (ir::timing), exports_summary_over_summary_as_typed_precompute_edges and physical_node_ids_serialize_in_deterministic_binding_order (physical_export.rs).

No planner behavior change is intended. Two kinds of visible change remain:

• Series-identity rule restored. The migration to ir::timing had loosened the maintenance-binary check, so a column named after the series identity was accepted with a Timestamp type. It must again be non-null Utf8 (InvalidMaintenanceBinary otherwise); the ported test caught this.
• Error messages and error types:
  • ConflictingTiming: "shared node is assigned conflicting timings: {first} and {second}".
  • UntimedNode: adds "; apply lifecycle timings before export".
  • UnimplementedOperator: "{operator} has no execution timing contract yet".
  • InvalidMaintainedPopulation: says "evaluation" instead of "readout".
  • ReadoutUnderMaintenance is folded into EvaluationUnderMaintenance { edge, child }.
  • ExactOperationSchemaError::{Schema(QueryExprError), Derivation} are merged into Schema(SchemaDerivationError).
  • AmbiguousKeepPreAsap is removed.

Before this PR

Two IRs were compiled side by side; docs and the viewer described the legacy one.

After this PR

asap-types has one IR (ir::OperatorNode), and the docs and the viewer describe it. Callers of the removed functions and error variants listed above must use the ir timing pass and the merged variants.

Evidence

Not applicable (removal; no planner behavior change). crates/devtools/tests/viewer_contract.rs (new) checks that the viewer knows every unified operator kind the export emits.

Verification

• Unit and integration tests: tests that lived in deleted legacy files are removed with them, except the three ported tests above; planner_vocabulary.rs is migrated off SchemaResolver.
• Viewer: crates/devtools/tests/viewer_contract.rs (new) and the tools/dag-viewer render tests.
• Other checks on the PR head: cargo fmt --check; workspace all-target Clippy with -D warnings; cargo test --workspace (1,635 passed, 0 failed, 2 ignored); dag-viewer test_render (31 passed); tools/verify_metricsql_external_consumer.sh OK.

Architectural decisions

• The summary maintenance lifecycle is removed in its own PR (refactor(planner): replace the summary lifecycle with #509 materialization options #667), because that changes when summaries are materialized, which is a product-semantics change, not a cleanup.

Limitations and follow-up

• Next PR (refactor(planner): replace the summary lifecycle with #509 materialization options #667): remove the summary maintenance lifecycle and replace it with the docs: propose workload-wide planning, summary sharing, and materialization #509 materialization options.
• Reorganize crates and modules by #509 stages and the #511 unified IR #572: arrange asap-types modules by docs: define unified operators and SQL/PromQL scalar boundaries #511 section (pre_asap / post_asap go away).

Stack: tracker #528 · main → #542 → #665 → #666 → #543 → #667 → #561.

Human review — do not complete with an agent

• The MVP boundary is correct.
• New conceptual layers or public interfaces are necessary.
• The before/after description matches the intended product behavior.
• Human reviewer:
• Decision and rationale:

🤖 Generated with Claude Code

https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

[zzylol]

Follow-up: the module/crate reorganization by #509 stages and the #511 IR (removing pre_asap / post_asap and splitting asap-aware-mapping) is tracked in #572 and starts after this cleanup lands.

🤖 Generated with Claude Code