[zzylol]

Why

#1144 made the group selection MILP (#1095) Stage 3's default search. The old engine (candidate graph, cost bounds, the evaluator-backed MILP refinement loop and the bounded objective traversal) stayed behind ASAP_STAGE3_SEARCH=old. It is no longer the plan we run, it is about 21.8k lines, and keeping it doubles every Stage 3 change.

What

• Deletes crates/plan-selection/src/search/ (the old engine and its tests) and the ASAP_STAGE3_SEARCH switch.
• select_plan moves into crates/plan-selection/src/select.rs and always runs the group selection MILP. Its signature is unchanged.
• The whole-plan selector (selector.rs, WholePlanSelection) is now compiled only for tests and the reference-pricer feature (its remaining users: stage_pipeline --exhaustive/--price-enumerated, planner_fuzz, integration tests).
• Comments and docs that referred to "today's engine", "the new search" or the switch are updated; planner-layering.md's "How the search uses the memo" paragraph and code locations now describe the group selection MILP; the crate's cost-model docs say select_plan supports cost model 1 only (Memo search solver: MILP over the memo for all cost models #888).

Capability lost, and where it is tracked

Old-engine capability	New implementation or issue
Stage 3 search under cost model 1	select::select_plan (group MILP, #1095)
Cost model 2 (peak) and cost model 3 (cluster capacity) search	#888 (MILP over the memo for all cost models); today these return UnsupportedCostModel, as on main since #1144
Measured operation cost table without a frozen calibration factor	#888; UnsupportedCostModel as on main since #1144
ASAP_STAGE3_SEARCH=old escape hatch	none (removed by request)

How

Move the entry point and its doc comment from search/select.rs into select.rs (cost-input validation first, then the existing MILP path), drop the env switch, git rm search/, and gate the whole-plan selector on cfg(any(test, feature = "reference-pricer")) so the default build has no dead code.

Key code interfaces added / modified

crates/plan-selection/src/select.rs, crates/plan-selection/src/lib.rs

/// Stage 3's only entry: select one plan among the physical memo's candidates
/// with the group selection MILP. Errs on invalid or unsupported cost inputs
/// (cost models 2/3), an infeasible catalog, no plan by the deadline, or a
/// realizer rejection (a Stage 2 catalog bug).
pub fn select_plan(
    physical: &PhysicalASAPMemoCandidates,
    entries: &[QueryWorkloadEntry],
    data: &DataWorkload,
    accuracy_checks: &AccuracyCheckInputs<'_>,
    cost_inputs: &CostInputs,
    realizer: &dyn AssignmentRealizer,
    deadline: Option<Instant>,
) -> Result<Selection, ErrorSelectingPlan>; // moved from search/select.rs; signature unchanged

/// Env switch to the old engine.
pub(crate) const STAGE3_SEARCH_ENV: &str = "ASAP_STAGE3_SEARCH"; // removed
pub(crate) fn milp_search_enabled() -> bool; // removed
pub(crate) fn select_plan_with_milp(/* ... */); // removed: now select_plan

crates/plan-selection/src/report.rs

/// The whole-plan selector's result; stays, now only with the
/// `reference-pricer` feature or in tests.
#[cfg(any(test, feature = "reference-pricer"))] // new gate
pub struct WholePlanSelection { /* fields unchanged */ }

Usage (unchanged for callers): the facade's run_planning calls asap_plan_selection::select_plan(&physical, &entries, &data, &checks, &cost_inputs, &realizer, deadline) and gets back the Selection (selected DAG, PlanCost, accepted accuracy per query, pruned candidates, SearchReport) exactly as on main with the default switch.

Major algorithm changes

None. Main already ran the group MILP by default; this removes the non-default engine.

Before this PR

select_plan ran the group MILP unless ASAP_STAGE3_SEARCH=old was set, which ran the old candidate-graph engine (and was the only way to search under cost models 2/3).

After this PR

select_plan always runs the group MILP. ASAP_STAGE3_SEARCH is ignored. Cost models 2/3 return UnsupportedCostModel until #888.

Verification

• cargo check --workspace --tests: clean, no warnings; also cargo check -p asap-plan-selection --tests --features reference-pricer.
• cargo test -p asap-plan-selection --lib: 142 passed, 0 failed.
• cargo test -p asap-devtools --test devtools stage_pipeline: 5 passed, 0 failed, 46 ignored (already #[ignore] on main: wall-clock or slow tests).
• cargo fmt --all.
• Screenshots / measurements: not applicable (no behavior change on the default path).

Limitations and follow-ups

• Memo search solver: MILP over the memo for all cost models #888: cost models 2/3 in the MILP.
• Remove unnecessary deadlines and caps in the group selection MILP #1185: remove the MILP's deadlines and inexact-costing caps (next PR, stacked on this one).
• The "candidate-connections" search label (stage_pipeline's search field, planner_fuzz --search candidate-connections) still names the deleted engine; rename in a follow-up.
• ErrorSelectingPlan::EvaluationDeadlineExceeded is now produced only by the reference pricer; revisit with Remove unnecessary deadlines and caps in the group selection MILP #1185's deadline removal.
• Historical design/develop docs (joint-planning-search.md, joint-planning-implementation-plan.md, evaluation notes) still describe the old engine as history; not rewritten here.

🤖 Generated with Claude Code