refactor(legacy_expr): step α — legacy AggIntent → canonical AggIntent at field level

[zzylol]

Summary

Unifying first move in the revised legacy_expr migration plan. After Batch 2's A-lifts, the 3d/3e agents surfaced an architectural wall: legacy and canonical QueryExpr are unrelated types; consumer migration requires Schema threading first.

Step α takes a different cut: at the type-of-field level, replace legacy_expr::AggIntent with canonical intent_algebra::agg_intent::AggIntent in every field that carried legacy AggIntent. The wrapping legacy_expr::QueryExpr variants (SketchAgg, WindowedAgg, etc.) stay intact — they now carry canonical AggIntent as their op/agg field.

Unblocks Step β (Schema threading) and Step γ (variant-by-variant consumer migration).

Translation strategy: option (a) — re-export + PerPartitionWrap

legacy_expr.rs now does:

pub use crate::intent_algebra::agg_intent::AggIntent;
pub struct PerPartitionWrap { inner: AggIntent, keys: Vec<String> }

The legacy AggIntent enum + ExactAgg helper enum are deleted. Free-function helpers (agg_is_mergeable, agg_is_exact, agg_to_legacy_agg_type, agg_quantiles, agg_accuracy, etc.) replace the old AggIntent::method() API.

Translation table

Legacy	Canonical
Quantile { quantiles: Vec<f64>, accuracy: f64 }	fan-out N × Quantile { q, accuracy: AccuracyTarget }
Cardinality/Frequency { accuracy: f64 }	same kind, accuracy: Exact if 0.0 else Epsilon(eps)
Extrema { min: true, max: false }	Min
Extrema { min: false, max: true }	Max
Extrema { min: true, max: true }	fan-out: Min + Max
Exact(ExactAgg::Sum/Count/Avg/Min/Max)	flattened to canonical variant
PerPartition { inner, keys }	PerPartitionWrap { inner: canonical, keys } (carrier; collapses in Step γ)

Fan-out: F1 — Merge-wrap

legacy_lower::agg_func_to_intents returns Vec<AggIntent>; lower_aggregate wraps siblings in QueryExpr::Merge { inputs: [SketchAgg|WindowedAgg, …] }. SketchAgg/WindowedAgg shapes unchanged; consumers keep matching single-intent op: AggIntent.

Active fan-out paths:

• AggFunc::StdDev / AggFunc::Variance → Merge of 2 Quantile siblings (q=0.25, 0.75)
• R6 HistogramQuantileFusion when φ ≠ existing q → Merge of 2 Quantile siblings

R9 HydraConversion: DISABLED

R9's multi-key Partition + sketch fusion inlined PerPartition into SketchAgg::op — a shape canonical AggIntent doesn't support. R9's try_rewrite returns None; original logic preserved under #[allow(dead_code)] mod hydra_conversion_legacy with a Step γ TODO.

Disabling a cost-driven rule preserves correctness (unfused tree still produces the right result, just less efficient for multi-key Partition + sketch cases).

Sites migrated

• 13 construction sites across legacy_lower, legacy_expr, physical/{allocator,planner,stage_split}, optimizer/engine
• 9 match sites in sketch_catalog, allocator, stage_split, query_parser, optimizer/engine
• 2 fan-out emit cases (above)
• legacy_expr::AggFunc kept — it's L2 (AggItem.func) per Plan agent's guidance. 4 references all in query_parser/mod.rs::collect_agg_func*.

Build + test

• cargo build --release -p controller — clean
• cargo build --release -p query_engine_rust — clean
• cargo test -p controller --lib — 666 / 666 pass (no regression)
• cargo test -p controller --bin controller — 27 / 27 pass

Verification grep

• ::AggFunc\b → 4 hits, all in query_parser/mod.rs (L2 intentional)
• legacy_expr::AggIntent\b | ExactAgg\b → 1 hit (the pub use re-export in stage_split.rs) + comments
• Extrema { → 1 hit, EvalFunc::Extrema in physical/planner.rs (unrelated L5 enum)

4 TODOs flagged for Step γ

1. Re-enable R9 HydraConversion as canonical Aggregate { by: keys, aggs: [inner] }
2. R6 smarter dedup: today emits Merge-of-siblings; γ should make it Merge-aware
3. sketch_op_to_promql fall-throughs: route archive-only / TopK / Rate / Increase through their dedicated PromQL templates instead of count_over_time fallback
4. agg_to_legacy_agg_type taxonomy collapse: γ plumb canonical kind directly into planner; remove the coarse AggType enum

Diff: 8 files, +556 / -337

🤖 Generated with Claude Code