fix(l3): thread the input column onto Quantile and Cardinality intents (#115)

[zzylol]

Closes #115.

The bug

AggIntent::Quantile and AggIntent::Cardinality carried no input column, so two aggregates over different columns lowered to byte-identical L3:

SELECT approx_percentile_cont(v, 0.5),   -- Quantile { q: 0.5, accuracy: Exact }
       approx_percentile_cont(w, 0.5),   -- Quantile { q: 0.5, accuracy: Exact }   <-- equal
       count(distinct v),                -- Cardinality { accuracy: Exact }
       count(distinct w)                 -- Cardinality { accuracy: Exact }         <-- equal
FROM t;

while Sum/Avg/Min/Max/StdDev/Variance all carry col: Option<ColumnId> and distinguish correctly.

Three consequences, all silent:

1. CSE miscompilation. plan::cse dedupes on AggIntent equality, so a query computing median(v) and median(w) collapsed into one.
2. Sketch bound to the wrong column. plan::bind picks the summarised column via input_col(). With no column it always fell through to ColumnRef::SampleValue, so every SQL COUNT(DISTINCT c) built its HLL over the wrong input.
3. Expression arguments silently dropped. approx_percentile_cont(v * 8, 0.95) succeeded and discarded v * 8, because the column slot that would have rejected it did not exist. SUM(v * 8) correctly errors.

Root cause

crates/l2/src/lower.rs. col is already in scope and threaded onto the six reducers above it, but was simply not passed to these two — an omission at the L2→L3 boundary, not a deliberate deferral to L4 (the L2 AggFunc does carry the ColumnRef).

The fix

• col: Option<ColumnId> on Quantile and Cardinality, #[serde(default)] so pre-L3: AggIntent::Quantile / Cardinality / TopK drop their input column — distinct aggregates compare equal #115 serialized L3 still deserializes (as None). None keeps its existing meaning: the PromQL sample value.
• Threaded through l2::lower, and both added to input_col() — which fixes plan::bind for free.
• TopK deliberately keeps no col: it ranks by the aggregate output rather than a base column (Do we need AggIntent::TopK? #13 / L3 IR structural inconsistency: SQL and PromQL heavy-hitter topk produce different tree shapes #25). Noted in the type.
• SQL-side, approx_percentile_cont / COUNT(DISTINCT …) / approx_distinct now resolve their argument through reducer_col, so an expression argument is rejected exactly like SUM(a*b). A SQL query has no sample value to fall back on, so col: None there means "we lost it". This makes agg_col_ref dead code — its only purpose was that fallback — so it is removed.

Tests (6 new)

test	pins
quantile_carries_its_input_column	distinct-column quantiles are not equal
count_distinct_carries_its_input_column	ditto for cardinality
quantile_and_count_distinct_over_an_expression_are_rejected	consequence 3
cse::quantiles_over_different_columns_do_not_dedupe	consequence 1
bind::sketch_binds_the_intents_input_column	consequence 2, incl. None → SampleValue
agg_intent_serde_reads_pre_115_payloads	#[serde(default)] back-compat

Verification

cargo test --workspace      # 363 passed, 0 failed
cargo clippy --all-targets  # clean

The diff is deliberately kept to semantic changes. main is not currently rustfmt-clean (159 pre-existing diffs), so cargo fmt --all was not run; each touched file has the same rustfmt-violation count as on main.

🤖 Generated with Claude Code