Cross-consumer accuracy reconciliation for CSE sharing (near-duplicate AggIntents with different AccuracyTarget)

[zzylol]

Part of #33.

Gap

ReplacementStrategy candidate generation (SketchAlgorithmStrategy) sizes every sketch candidate from the AccuracyTarget embedded directly on the AggIntent at that one target — see accuracy_target/accuracy_budget/default_size_params in replacement.rs. That's correct and complete at single-target granularity.

At workload granularity, there is no reconciliation of accuracy requirements across consumers of what is conceptually the same computation. SharedSubtreeStrategy/PlanSpace only ever see a shared TargetSubDAG when pre-ASAP's share_common_subtrees has already merged two subtrees onto one Rc — and since AggIntent derives PartialEq over all its fields including accuracy, two otherwise-identical aggregates with different AccuracyTargets are never structurally equal, so they never merge. SharedSubtreeStrategy never gets a chance to propose "share" for them at all.

Concretely: if query A asks for quantile(0.99, x) at epsilon=0.01 and query B asks for the same quantile(0.99, x) at epsilon=0.05, this crate today builds two entirely independent candidate spaces for two entirely independent Rcs — there's no path to "build one sketch sized to the tighter bound (0.01) and let both consumers read from it," even though that would be strictly cheaper than building two.

Why this wasn't done in #259

Doing this correctly needs an accuracy-aware notion of "these two subtrees are close enough to share" that's looser than pre-ASAP CSE's exact structural PartialEq — a real design question (how loose is safe? does sharing at the tighter bound ever violate a looser consumer's own semantics — no, but does it change cost-model tradeoffs it should account for?) that #259 didn't have a driving use case for yet. Flagging as known future work rather than silently leaving it undiscoverable.

Possible shape (not a commitment, just a starting point)

• A pre-pass (or a ReplacementStrategy-adjacent mechanism) that recognizes two AggIntents equal in everything except accuracy, and represents that relationship without requiring pre-ASAP CSE's stricter equality to change (accuracy still matters for correctness elsewhere; this is additive, not a relaxation of share_common_subtrees itself).
• SharedSubtreeStrategy (or a new, narrower strategy) proposing "build once at the tightest of the group's accuracy requirements, all consumers read from it" as one more ReplacementSubDAG candidate — still exhaustive, still ranked by CostModel, not a forced choice.

[zzylol]

Related open work

• Accuracy-aware roll-up composition is tracked by Support accuracy-aware approximate Count roll-ups #278.
• Nested approximation propagation is tracked by Propagate end-to-end accuracy guarantees for approximate-over-approximate plans #172.
• Posterior error estimation is introduced by Add posterior (query-time) error estimation for CMS/CountSketch, per Chen et al. IMC'21 #239.
• The global candidate search that would consume reconciled requirements is tracked by asap-aware-mapping: Cascades/Volcano-style candidate-plan search engine over ReplacementStrategy #252.