feat: size shared summaries for the strictest consumer

[zzylol]

Why

#509, logical Pass 2, summary-capability rule: when several computations have the same summary input data and window, one summary should serve all of them, sized for the strictest accuracy requirement. After #515 and #516, identical states are already shared, including p50 and p99 at equal accuracy. Queries whose accuracy targets differ still produce different sketch parameters, so they never share. Example: p50 at ε=0.01 and p99 at ε=0.001 over lat[5m].

What

All in crates/asap-aware-mapping/src/replacement.rs, about +110 lines.

• Siblings. strictest_sibling_accuracy finds sibling targets: single-measure aggregates with the same child, grouping and filters, and the same intent apart from accuracy. A quantile's rank may also differ, since it is only a readout parameter. It returns the tightest ε/δ across them when that is stricter than the target's own.
• Pass 1. search_cse_workload_with passes the result through TargetSubDAG.strictest_sibling_accuracy. SketchAlgorithmStrategy then adds one extra candidate per sketch, re-sized with the existing size_params, but only if the parameters actually change. Each reader's guarantee is recomputed from the new parameters and its own readout.
• Sharing and cost. feat: share identical summary producers across queries after Pass 1 #515's CSE merges the equal variants and selection prices them: a shared state is costed once and split across its consumers. On its own, the larger variant costs more and is not chosen.
• Docs. share_common_summary_subtrees now documents that sharing identical states is the summary-capability rule.

UnivMon sizing is fixed, so it gets no variant. Its distinct, entropy and L2 readouts already share one state when certified.

Not changed: accuracy gates, capability.rs, and AccuracyReconciliationStrategy. No DDSketch Count candidate is added.

Candidate growth

A target gains at most one extra candidate per sketch algorithm, and only when a stricter matching sibling exists.

Tests

New or updated in crates/planner/tests/summary_sharing.rs:

• quantiles_share_one_producer_sized_for_the_strictest_consumer: p50 at ε=0.01 and p99 at ε=0.001 share one KLL. Its k is the ε=0.001 sizing, and both guarantees meet their own ε. Planned alone, the ε=0.01 query keeps its own sizing. The test fails when the new sibling field is forced to None.
• different_producers_are_not_shared: dropped the old "different accuracy is not shared" case, because that sharing is now intended. Added stricter-accuracy cases with a different window or selector, which still don't share.
• certified_frequency_readouts_share_one_univmon_state: distinct, entropy and L2 share one UnivMon state under a test accuracy model that certifies them. This test reproduces the MajorPass steps directly instead of going through e2e_plan, because MajorPass gives its accuracy model only to the final root check, not to the Pass 1 strategies. That will be tracked separately.

Notes

When siblings disagree on which of ε or δ is tighter, the variant takes the tightest of each. The result satisfies every sibling, though no single query asked for exactly that pair.

Validation

On Rust 1.99, all of these pass: cargo fmt --check, clippy with -D warnings, the vendored MetricsQL baseline, and cargo test --workspace.

🤖 Generated with Claude Code