feat(lower): support *_over_time over a sub-query (#27)

[zzylol]

What & why

Follow-up to #27. The *_over_time / quantile_over_time-over-sub-query family — the canonical

max_over_time(rate(m[5m])[1h:])

— was rejected:

unsupported feature: expected a range-vector (matrix) argument, got Discriminant(4)

extract_matrix only accepts a (parenthesised) matrix selector, not the PromQLSubquery a sub-query argument lowers to. This was pinned as over_time_of_subquery_is_rejected__GAP (the last un-issued item on #27's checklist after topk-bare-by and SQL derived tables landed).

How

crates/lower/src/promql.rs — a new walk_call intercepts a *_over_time/quantile_over_time call whose argument is a sub-query, lowers the sub-query recursively (PromQLSubquery), and reduces it per series. Non-sub-query calls fall through to the existing flat lower_inner_call template unchanged:

max_over_time(rate(m[5m])[1h:])
→ Aggregate{Max} → Subquery{1h} → Aggregate{Rate} → TimeRange{5m} → Scan

over_time_reducer reuses inner_func (no AggFunc drift); quantile_over_time's φ is read from arg 0, the sub-query from arg 1.

crates/core/src/intent_algebra/query_expr.rs — an empty-keys single aggregate over a Subquery is now recognised as a label-preserving per-series range reduction, mirroring the existing TimeRange marker.

Why the per-series marker is sound

An empty-keys aggregate over a Subquery would otherwise be treated as a cross-series reduction and collapse every label. That's safe to change because a genuine cross-series aggregation operator over a range vector — sum(rate(m[5m])[1h:]) — is a PromQL type error the parser already rejects (expected type vector in aggregation expression, got matrix). So an Aggregate directly over a Subquery only ever arises from this per-series *_over_time shape.

Confirmed by sum by (job) (max_over_time(rate(m[5m])[1h:])): the inner Max preserves the job label so the outer sum by (job) resolves and groups on it.

Tests

Flips over_time_of_subquery_is_rejected__GAP into three passing conformance tests:

• over_time_of_subquery_reduces_per_series — per-series Max over a Subquery, no group keys, inner Rate present.
• quantile_over_time_of_subquery_carries_phi — φ = 0.9 threaded from arg 0, sub-query from arg 1.
• aggregation_over_over_time_of_subquery_keeps_labels — outer sum by (job) groups on the label the inner reduction preserved.

Regression-checked: plain max_over_time(m[5m]), heavy-hitter topk(k, count_over_time(...)), and parser-rejected sum(subquery) unchanged. Full workspace suite green; cargo clippy --all-targets clean.

Notes

Branches off latest main (includes #31/#32). Advances #27 — the remaining checklist items there are now the SQL scalar/IN/EXISTS subqueries and the scalar/negation operands (tracked by #35/#36).

🤖 Generated with Claude Code