refactor(planner): move Pass 1 helpers out of the legacy search; drop show_post_asap_ir, erp and empirical_comparison

[zzylol]

Why

This is the first step in removing the legacy replacement search. The rule is that a legacy piece is deleted only once the new architecture implements it, and the group that reimplements a capability deletes the legacy piece. This PR deletes only the pieces that are already replaced (show_post_asap_ir → stage_pipeline) or that the user decided to drop (erp, empirical_comparison: Assumption 5's deployment-owned empirical models cover them, and sketch-bench keeps its own copy). It also moves the helpers Pass 1 shares with the legacy search out of replacement.rs, so that file can be deleted later without touching Pass 1.

What stays, and which group or issue removes it:

• recurrence, SharedSubDAGStrategy: G4
• accuracy::reconciliation: G5
• HLL confidence wrapper: Wire the HLL confidence model into Stage 1 sizing and plan selection #678
• function_rules, accuracy::composition, exact_composition: Compose accuracy guarantees through arithmetic and comparisons over approximate values #669, with an accuracy::allocation follow-up
• pane_sharing: G6/G10
• grouping: G7
• cost_model capability hooks: G9
• empirical_cost, empirical_resources, physical_plan_cost_model: Calibrate plan selection's costs from measured sketch costs #679
• rollup, topk_reuse, nested-sum: Pass 2 reuse rules: rollup, top-k reuse, nested sum merge #680
• population adapter: Offer maintained-population top-k candidates in Pass 1 #681
• dag_export, asap_types::{dag_export, cost}: the early viewer rebuild
• legacy core (replacement search, rewrite, explanation, propagate, PlanningModels.cost, the legacy integration tests and evidence types): last

What

• New asap_aware_mapping::realization holds summary_candidates, accuracy_target, accuracy_budget, default_size_params, DEFAULT_DELTA, column_ref, summarised_column and summarised_input. The bodies moved unchanged from replacement.rs. The one difference is that summarised_column is now pub(crate), because the legacy search still calls it. Pass 1 (per_computation_candidates), replacement.rs and the other callers import from the new module. RealizationError also moves to realization (re-exported from replacement), so the new module does not depend on the legacy search.
• Deleted the show_post_asap_ir bin, which stage_pipeline replaces. Renamed show_pre_asap_ir to show_logical_dag. Bins are auto-discovered, and no CI or tests referenced either name.
• Renamed the frontend-common resolver's fallback schema to usage_derived_schema.
• Deleted erp.rs and empirical_comparison.rs, plus empirical_cost::snapshot_prepare_cpu (only empirical_comparison used it) and its unit test.
• No Markdown changes.

Verification

• git grep -w finds no remaining reference to any deleted public item or to either old bin name outside Markdown.
• cargo fmt --all -- --check and cargo clippy --workspace --all-targets -- -D warnings: clean.
• cargo nextest run --workspace: 1609 passed, 50 skipped. All 23 removed tests were in deleted code: 10 in erp, 9 in empirical_comparison, 3 in show_post_asap_ir, and 1 for snapshot_prepare_cpu.
• stage_pipeline --example planner-layering-1 output is byte-identical to tools/dag-viewer/examples/planner-layering-example1.json.
• tools/dag-viewer/test_render.py: 29 OK. tools/verify_metricsql_external_consumer.sh: OK.

Human review

• The moved helpers and RealizationError are unchanged
• Deleting erp/empirical_comparison (and snapshot_prepare_cpu) matches the decision
• The ownership list above for what stays is correct

🤖 Generated with Claude Code

https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW