executor: support tumbling pane merges (per-entity TimeShift panes, temporal merge timestamp)

[zzylol]

Rebased on main d4869a7 (DF 54).

Wave 1 chain: #594 → #593 → #592 → #591 → #595 → #596 → #597 → #598 → #599 (on #589)

Why

#580 will plan a tumbling window as N pane SummaryAggs merged by SummaryMerge (#509 Example 3/4, Pattern B). Pane i reads TimeRange(w) over TimeShift(i·w) over Scan. For per-series (PromQL) queries, the executor could not run that shape. This PR ports the executor parts of the parked #566 and #569 onto the renamed crates/executor.

What

1. Per-entity build over TimeShift (from Generate exact window composition and materialization candidates #566). The per-entity SummaryAgg path required Scan directly under TimeRange. It now looks through one TimeShift. The deployment supplies each pane's shifted raw rows.
2. Temporal pane merge keeps the evaluation timestamp (from Preserve evaluation timestamps in native temporal pane merges #569). A SummaryMerge whose output has a time_index now lowers to [union] → summary_merge (group by series, no ts) → scope_timestamp, the same pattern as per-series SummaryAgg. Pane build timestamps are grouped away, and the run scope supplies the time column. Non-temporal merges are unchanged.

How it is tested

crates/executor/tests/tumbling_pane_merge.rs takes the selected single-build plan for sum_over_time(m[5m]) (exact) and quantile_over_time(0.99, m[5m]) (KLL, k=52). It rewrites the plan into 5 one-minute panes under a SummaryMerge and runs both versions over 2 series × 20 samples, at evaluation time 300 000 ms. Pane coverage uses absolute regions [T−(i+1)·60 000, T−i·60 000). The pane plan runs two ways:

• B2: panes rebuilt from raw rows at query time.
• B1: panes built in Scope::Ingestion for their own window, so they have distinct build timestamps (60 000 … 300 000), then retained and merged at query time.

Before this PR:

exact_sum_over_tumbling_panes_matches_single_build ... FAILED
  Invalid("per-entity summary requires a resolved source")
# with only fix 1 applied:
  Invalid("node 20: invalid DAG: native output type differs from Planner output")

After this PR: both tests pass. Each path returns one row per series, ts = 300 000, and the output is identical to the single 5-minute build: Sum [190, 2190], p99 [19, 119]. 20 samples < k, so the KLL result is exact.

Notes for the planner side (#580)

• The tests set timing = QueryTime on each node by hand. apply_materialization_timings (types) still returns UnimplementedOperator { "SummaryMerge" }, so a planner-emitted pane plan cannot be exported until that rule is added.
• Coverage regions are absolute here. The relative-time coverage (W2) is being added to types separately.

Gate: cargo fmt --all --check, cargo clippy --workspace --all-targets --all-features -D warnings, and cargo test --workspace pass: 1,519 passed / 10 ignored, compared with 1,517 / 10 on #589. The +2 are the new tests.

Origin: #566, #569. Part of #580 (item O), #509.

🤖 Generated with Claude Code