[zzylol]

Why

Without retained raw data, the MVP benchmark's p99 query cannot be planned:

avg by (region, service, job) (quantile_over_time(0.99, data[100s]))

(every 1 s, max_error 0.05, failure probability 0.3, data a gauge declared "values_non_negative": true, raw_data_retained: false). Stage 3 rejects every candidate. The only candidates that read no raw data are the per-series DDSketch panes (P5, and P5-m1 built at ingestion). They were rejected with accuracy_target_missed, "Relative bound None", because the samples' sign never reached avg.

#780 attaches the sign of a metric's samples to a quantile estimated over the raw window (estimates_samples_of), but not to a quantile read from merged tumbling panes. A merged DDSketch keeps its alpha, and a quantile of merged panes is still one of the samples (or between two of them), so the sign carries over in the same way.

What

• A quantile estimate (φ in [0, 1]) read from SummaryMerge over a PaneChain of one definition now carries the samples' sign statistic, just as the same quantile over the raw window does. The rule does not depend on the family, so it covers KLL panes as well. KLL's rank error still has no rule through avg, so KLL candidates are unchanged.
• With the declaration, P5-m1 gets DDSketch's relative bound composed through avg and is selected. Without the declaration, it is still rejected.
• Unchanged: the φ ∈ [0, 1] guard, and every other stop. A filter, a value projection, abs, delta/deriv and an @ anchor still end the sign.

How

In estimates_samples_of, the quantile arm resolves its summary input through merged_chain (a SummaryMerge of exactly one PaneChain, or the chain itself). It then matches either a SummaryAgg or a PaneChain that has no item, weight SampleValue and no filter, and recurses into the chain's input (the input the range selector reads) with the existing strict rule. The guarantee composed for merged panes (merged_panes_guarantee) and the executor's pane expansion are not changed.

Key code interfaces added / modified

None. The change is inside the private estimates_samples_of in crates/logical-asap-optimization/src/per_computation_candidates.rs.

Usage (end to end, from crates/planner/tests/runtime_install.rs, run at this head):

// What the user writes: the benchmark's p99 query and the declared gauge.
let workload: PlanningWorkload = p99_avg_workload(/* values_non_negative */ true);
// The deployment keeps no raw samples.
let capabilities = DeploymentCapabilities { raw_data_retained: false, ..Default::default() };
// One Planner call. Stage 1 composes P5-m1 (DDSketch panes merged at query time).
// Its avg now sees values_non_negative[data] on the merged quantile, and Stage 3
// binds that statistic to 1 from the declaration.
let plan = plan_installable(UserInput::new(
    &workload,
    FrontendInput::Promql { histograms: None },
    PlanningModels::builtin().with_deployment_capabilities(&capabilities),
)).await?;
assert_eq!(plan.selection.unwrap().selected_candidate_id, "P5-m1");
// The root's guarantee once the sign and the 1,000 series are bound:
// Relative, error 0.05, failure probability 0 (DDSketch is deterministic).

stage_pipeline on the benchmark workload (q03, with values_non_negative: true) and its deployment_capabilities.json:

Candidate	Before	After
P1	raw_data_not_retained	raw_data_not_retained
P2, P3, P3-m1	no_accuracy_guarantee	no_accuracy_guarantee
P4	raw_data_not_retained	raw_data_not_retained
P5	accuracy_target_missed (Relative bound None)	raw_data_not_retained (now passes accuracy)
P5-m1	accuracy_target_missed (Relative bound None)	selected: accepted relative (DDSketch), 21.22 cost/s: DDSketch in 100 panes of 1000 ms for 1000 groups

The same workload without the declaration still has no valid plan: P4, P5 and P5-m1 give accuracy_target_missed, Relative bound None.

Before this PR

Without raw data, a per-series quantile sketch had no relative bound through avg/sum when it was built in panes, even over a declared non-negative gauge. The benchmark query had no valid plan.

After this PR

A declared non-negative gauge's quantile sketch, built in panes at ingestion and merged at query time, keeps its relative bound through avg and sum. The benchmark query plans as P5-m1.

Verification

Tests were written first and fail on main (8f5bf8e):

• a_quantile_of_merged_panes_shares_its_samples_sign (logical): a quantile over SummaryMerge[PaneChain] of m names values_non_negative[m] in its fold stats, and φ = -1 and 2 carry no sign. On main: None vs Some("m").
• a_non_negative_gauges_merged_ddsketch_panes_carry_the_relative_bound_through_avg (planner, plan_installable): P5-m1 is selected, with no query-time raw windows, a relative bound ≤ 0.05 and failure probability ≤ 0.3. It is then executed with the executor over 8 series in 4 groups, with 100 one-second panes built in ingestion runs. Its answers equal query_time_reference's and are within 5% of the exact averages of the per-series p99s. On main: "no valid candidate", with P5-m1 accuracy_target_missed.
• an_undeclared_gauges_merged_ddsketch_panes_get_no_relative_bound_through_avg: without the declaration, planning fails and names P5-m1.
• cargo test --workspace --locked: 1785 passed, 0 failed.

Limitations

• The rule follows only a SummaryMerge of one PaneChain, the shape the window-composition rule produces. A merge of several independently built states is not covered.
• The harness workload q03-quantile.json does not declare values_non_negative yet. It needs "values_non_negative": true on data to plan.