Pass 2: window-composition rule (tumbling windows) — blocked on #511

[zzylol]

Goal

Implement the window-composition rule of logical Pass 2 in #509 (docs/design_docs/proposals/planner-layering.md, Example 3 Pattern B), starting with tumbling windows. Computations with the same summary input data share one tumbling-window summary. Each query's window is answered by merging the tumbling windows it covers. The tumbling length must divide both the query window and the evaluation interval. For example, [5m] evaluated every 1m uses 1-minute KLLs and merges 5 of them.

Blocked on

#511 (unified operators). Today a summary's window lives inside its input (SummaryAgg { KeepPreAsap(TimeRange{w, …}) }), so the window is part of the state's identity. A clean window-composition rule needs KeepPreAsap removed first, which is part of the #511 implementation.

Current state on main (107ab28)

• The evaluation interval isn't visible to planning. It exists only in RepeatedDemand (crates/types/src/workload.rs). Pass 1/2 (search_workload_with_targets) receive only (id, expr, accuracy).
• The window framework (tumbling/sliding/EH) is always None in production. selected_window_framework comes only from provider evidence, which only tests register (summary_maintenance_cost/model.rs, bind_window_framework_candidate).
• The planner never builds a SummaryMerge. Its doc says "inserted by a deployment's stage allocator". The physical planner already compiles SummaryMerge as union + merge.
• Queries with different windows never share. feat: share identical summary producers across queries after Pass 1 #515 shares only structurally identical SummaryAggs, and the window is inside the child. different_producers_are_not_shared (crates/planner/tests/summary_sharing.rs) covers this.
• select_shared_panes (pane_sharing.rs) is dead code. It has no callers outside its own tests and duplicates feat: share identical summary producers across queries after Pass 1 #515's class pricing.
• Merging is lossless for KLL (fixed k) and for exact Sum/Count/Min/Max. Rate/Increase need counter-reset handling across window boundaries. Offline empirical evidence rejects merges.

Proposed design (to revisit after #511; recommendations, not final)

1. Merge count: give SummaryMerge the query window, so it merges W/g tumbling windows. No new node type. Drop the "stage allocator only" note.
2. Tumbling length g: represent it in the state's input window so feat: share identical summary producers across queries after Pass 1 #515 identity applies unchanged, and set selected_window_framework = Tumbling from the rule itself. Exact representation depends on docs: define unified operators and SQL/PromQL scalar boundaries #511's post-KeepPreAsap IR.
3. Evaluation interval: pass RepeatedDemand into Pass 1/2. g = gcd of the windows and evaluation intervals, or of the windows alone when no interval is known.
4. Phase/origin: the planner records only g. Window origin and phase belong to physical planning (stage 2).
5. Delete select_shared_panes.
6. Different accuracy across siblings: reuse the strictest-consumer sizing from Pass 2 rule 2. Window accuracy is Exact for mergeable families only.
7. Families: start with KLL and exact Sum/Count/Min/Max. Exclude Rate/Increase and non-mergeable intents.
8. A single repeating query (Pattern B) gets a tumbling candidate through the same path, with a group of one. The independent candidate is kept, and selection compares the two.
9. Candidate growth: one tumbling candidate per maximal sibling group, not every subset.

Tests to add

• quantile_over_time(0.99, lat[5m]) every 1m plus quantile_over_time(0.5, lat[15m]) every 5m. Expect one shared 1-minute state, a per-query SummaryMerge, and independent candidates still present.
• Single-query Pattern B.
• Update different_producers_are_not_shared, using costs under which sharing loses.

Later

Sliding windows and Exponential Histograms. EH also needs query-time materialization and the summary subtract/delete design, which is a TODO in #509.