feat(core): converge grouped *_over_time onto positional Aggregate.by (closes #8)

[zzylol]

What

Closes #8 — the residual of PR #5's review D. Grouped *_over_time aggregates now converge onto a positional Aggregate.by, the same shape SQL GROUP BY and the rate/increase cases already produce — eliminating the last name-based Partition-as-grouping for cross-series aggregates.

sum by(instance) (avg_over_time(node_cpu_seconds_total[5m]))
  before: Partition{By([instance])} ∘ Aggregate{by:[],[Sum]} ∘ Window{Aggregate{[Avg]}}
  after:  Aggregate{by:[instance], [Sum]} ∘ Window{Aggregate{[Avg]}}   (no Partition)

Why it was deferred, and why it's now clean

The *_over_time reduction lowers to Window { Aggregate{ by:[], [reducer] } }, and that inner Aggregate collapsed its schema to [reducer], dropping the label x — so the outer cross-series aggregate had no x to resolve into Aggregate.by. The two blockers from review D both dissolve:

• Shared AggIntent variants (avg_over_time vs cross-series avg, both AggIntent::Avg): resolved structurally — the enclosing time Window is the per-series discriminator, no intent disambiguation needed.
• Sample-value resolution (the topk by(h)(avg_over_time(…)) regression caught in D): the preserved value column is kept named value, so SampleValue resolves it by name regardless of how many label columns are present.

How

One behavioral change in output_schema_in (query_expr.rs): a time Window over a single-agg, by:[] Aggregate is a per-series reduction → re-derive its schema label-preservingly (keep every label, replace only the sample value). The per-series logic is factored into per_series_reduction_schema, shared with the rate/increase (D) path. The converter is unchanged — with labels preserved, its existing key resolution lands by=[x].

Scope

Converges sum/count/avg/quantile-by(x) over *_over_time. topk/bottomk by(x) go through the generic Sort+Limit path, where grouping rides a Partition by a different mechanism — left as-is (and verified non-regressing).

Tests

• Updated the two flipped shape tests (quantile_over_time, count_over_time grouped → Aggregate.by).
• Added: a Window-over-Aggregate label-preservation unit test; a sum by(instance)(avg_over_time(...)) conformance test.
• topk_over_avg_is_generic_sort_limit still green (no regression).
• Full workspace green (135 tests), clippy --all-targets -D warnings + fmt --check clean.

🤖 Generated with Claude Code