Wire the HLL confidence model into Stage 1 sizing and plan selection #678

ClassicHllConfidence (accuracy/estimators/hll.rs, kept by #676) gives classic HLL an (ε, δ) guarantee and picks its precision, given a deployment-supplied bound on the number of distinct items per evaluation (EstimatorContract::ClassicHll { max_distinct_per_evaluation }). The legacy search used it; the Stage 1 → Stage 3 pipeline does not. Today Pass 1 sizes HLL by error rate only, and HLL has "failure probability None", so plan selection rejects it for any (ε, δ) target unless the deployment supplies output evidence.

To do

• Accept the distinct-count bound as a deployment input. Keep the planner deployment-agnostic: the bound arrives as an input, not as built-in knowledge.
• Use it in Stage 1 composition to size HLL (precision) and to set its theoretical guarantee.
• Plan selection then accepts HLL when the guarantee meets the target.
• Test: Example 2's HLL candidate becomes valid with a bound and stays invalid without one.

Found in the #676 capability audit.