refactor(planner): Stage 1 no longer depends on the cost model

[zzylol]

Why

#572 ("Update (2026-10-04)") records decision Q36(a): following #509, only Stage 3 uses the cost model. After #581 the planner facade already runs Stage 1 → Stage 2 → Stage 3, but the Stage 1 modules still imported cost_model and recurrence: candidate generation asked a CostModel how to order and size sketches, and the legacy whole-workload selection lived inside replacement.rs. That coupling also blocks moving Stage 1 into its own crate (Phase B, step B2).

What

Stage 1 no longer depends on the cost model. Planner facade behavior is unchanged: the #581 tests pass and stage_pipeline regenerates the Example 1 fixture byte for byte.

Before this PR (production imports):

replacement        ──► cost_model, recurrence
exact_composition  ──► cost_model
grouping           ──► cost_model
cost_model         ──► replacement            (cycle)

After this PR:

replacement        ──► plan_selection::candidate_selection   (GlobalSelection, TargetSubDAGSelection; a B2 blocker, see below)
exact_composition  ──► replacement, accuracy
grouping           ──► replacement, accuracy
plan_selection::candidate_selection ──► cost_model, recurrence, replacement
cost_model         ──► replacement

A new test, stage1_cost_independence, fails if a Stage 1 module imports cost_model or recurrence outside #[cfg(test)].

How

1. Cost-free candidate generation. realizations_for_intent(intent) lists sketch candidates in summary_candidates order and sizes them with accuracy::estimators::size_params. CostModel::size_params defaulted to the same function. Extension intents stay PassThrough, which was the previous default. CandidatePlanningInputs drops cost, and realize_child drops its cost argument.

2. Cost-free strategies. Strategies are now built without a cost model:

   • ASAPStrategies::default() and HydraGroupingStrategy::default();
   • ExactCompositionStrategy, now a unit struct;
   • new_with_planning_inputs*, without the cost argument;
   • default_strategies_with_evidence(evidence).

   The strategy-level runtime-support check is dropped, because selection already requires positive support evidence before it commits a composition.

3. Pure move. The cost- and recurrence-dependent selection block moves from replacement.rs to the new plan_selection/candidate_selection.rs (plan_selection.rs becomes plan_selection/mod.rs), together with its tests. Only visibility and imports change. Its deletion is tracked in Pass 2 sharing and Stage 1 coverage parity with the retired MajorPass #580.

4. Call sites. Call sites in integration-tests, asap-physical-operators, frontend-promql, planner and devtools use the new constructors, and docs/develop_docs no longer describes the removed hooks.

Moved to plan_selection::candidate_selection (1,718 production lines and 22 tests)

• ReplacementSubDAG::runtime_support_evidence
• CandidateLogicalASAPDAGs::{cost_sorted, cost_sorted_with_recurrence, recurrence_profiles, global_selection, global_selection_with_recurrence}
• RecurrenceProfileMap, RankedTargetSubDAGCandidates, rank_group, cse_preference, realize_one, sketch_kind_of, summary_grouping
• TargetSubDAGSelection, CompositionDecision, GlobalSelection (with assemble_*), composition_options, is_automatically_selectable, decide_*, multiplier, ReferenceDAG, topological_order, and their helpers

These are still used by tests, devtools dag_export --post-asap and integration-tests, so they are moved rather than deleted. The public types are still re-exported at the crate root.

Deleted

Item	Evidence or reason
CandidateLogicalASAPDAGs::recurrence_profiles_from_workload	No caller in the workspace
CostModel::{size_params, realize_extension, evaluation_extension}	Stage 1 no longer consults them, and they have no other caller
ASAPStrategies::{new, default_cost_model}, HydraGroupingStrategy::{new, default_cost_model}, ExactCompositionStrategy::{new, default_cost_model}, default_strategies_with	Replaced by the cost-free constructors (default_strategies_with(m) equals default_strategies() once the cost model is gone)

Ten tests were deleted, each because it exercised only a removed hook:

Test	Removed hook
cost_model::size_params_default_body_matches_default_size_params	custom size_params
cost_model::custom_cost_model_can_override_sizing_independently_of_ranking	custom size_params
replacement::extension_intent_realizes_via_custom_cost_model	custom realize_extension
exact_composition::a_runtime_without_the_capability_gets_no_candidate	strategy-level support evidence
integration-tests exact_composition::a_runtime_without_mixed_execution_gets_no_composition_candidates	strategy-level support evidence (selection-time exclusion is still covered by unknown_runtime_capability_keeps_candidate_but_prevents_selection)
replacement::realize_with_custom_cost_model_overrides_default_summary_choice	strategy-level ranking in realize_child
replacement::custom_cost_model_still_enumerates_every_candidate_not_just_its_own_pick	strategy-level ranking (enumeration is still covered by approximate_quantile_enumerates_every_summary_candidate)
grouping::custom_cost_model_cannot_enable_unproven_hydra_kll	strategy-level ranking (still covered by quantile_has_no_hydra_candidate_without_a_modeled_error_bound)
explanation::custom_cost_model_changes_the_reported_sketch_kind	strategy-level ranking
empirical_cost::public_cost_model_changes_replacement_order_without_changing_guarantees	strategy-level ranking and size_params

One test was adapted: replacement::enumerating_the_targets_candidates_does_not_leak_into_a_nested_aggregate now expects the inner aggregate's static order [Kll, DDSketch] instead of a custom cost model's order.

Behavior note (legacy search path only). A custom cost model no longer reorders or resizes candidates during generation. It still ranks them at selection time through rank_candidates, candidate_cost and the other hooks. The in-tree production models (PhysicalPlanCostModel, the dag_export export model) already used the identity ranking.

Remaining B2 blockers (moving Stage 1 into a logical-optimizer crate)

1. Stage 1 imports Stage 3: replacement.rs:383 imports plan_selection::candidate_selection::{GlobalSelection, TargetSubDAGSelection}. enumerate_candidate_dags* builds them at replacement.rs:4240, replacement.rs:4249 and replacement.rs:4258 (assemble_target) to assemble the unpriced candidate inventory.

2. candidate_selection.rs adds inherent impls on Stage 1 types: impl ReplacementSubDAG (candidate_selection.rs:34) and impl CandidateLogicalASAPDAGs (:53, :199, :1211). These are not allowed across crates. They also use Stage 1 internals that are only pub(crate):

   • the fields groups, order and composition_plans;
   • PreparedComposition;
   • cse_candidate_pair, direct_child_counts and realize_child (candidate_selection.rs:28-29).

   Resolved by Pass 2 sharing and Stage 1 coverage parity with the retired MajorPass #580.

3. The shared accuracy module is used by both Stage 1 and Stage 3, and it imports Stage 1:

   • accuracy/mod.rs:31 imports exact_composition::ExactOperation;
   • accuracy/estimators/mod.rs:160 uses replacement::accuracy_budget;
   • accuracy/composition.rs:453 uses the private function_rules.

   It has to move with Stage 1, or be split.

4. cost_model.rs:52-53, :858 and :875 call the pub(crate) replacement::realize_child. The direction is fine, but this needs a public API once Stage 1 is a separate crate.

No Stage 1 module imports Stage 2 (physical_candidates) or the facade (pass).

Gate

• cargo fmt --all --check: passes.
• cargo clippy --workspace --all-targets --all-features -- -D warnings: passes.
• cargo test --workspace: 1,524 passed and 12 ignored. That is the 1,533 from feat(planner): run the #509 stage pipeline with tree-DP selection; retire MajorPass #581, less the 10 deleted tests, plus the new guard.
• python3 -m unittest discover -s tools/dag-viewer -p test_render.py: 29 tests OK, 6 skipped.
• stage_pipeline --example planner-layering-1 regenerates tools/dag-viewer/examples/planner-layering-example1.json byte for byte.

Part of #509. Follows #572 (decision Q36(a)). Follow-up: #580.

🤖 Generated with Claude Code