Pass 1 offers CountSketch for a count, but the executor cannot evaluate a CountSketch total #727

Problem, by example

stage_pipeline --table '{"name":"metrics","columns":[{"name":"ts","type":"timestamp","nullable":false},{"name":"service","type":"utf8","nullable":false}],"time_index":0}' \
  --sql "SELECT COUNT(*) FROM metrics" --epsilon 0.01 --delta 0.001

• Stage 1 lists L3 Q1 CountSketch.
• Stage 3 marks P3 valid ("accepted absolute_stream_l2 (CountSketch)", cost 0.00254/s, against 0.0000733/s for exact P1). P3 is only rejected because it costs more.
• If P3 were selected (for example, if the exact plan were priced higher), lowering would fail. validate_sketch_evaluation (crates/asap-physical-operators/src/capability.rs) accepts a bare PointCount { value: None } only for Count-Min, Hll, Kmv, DDSketch and UnivMon; for CountSketch it returns "evaluation is not implemented for this summary family".

Expected: no CountSketch candidate for a count, or a CountSketch total the executor can evaluate.

Cause

• realizations_for_intent (crates/asap-aware-mapping/src/realization.rs) maps AggIntent::Count to [Cms, CountSketch, UnivMon].
• The CountSketch accuracy model (count_sketch_l2_median_binomial_v1) then gives the total a single-item point-query bound, ε‖f‖₂, which is not a contract for a total.
• The family-capability table has no (Total, …) row for CountSketch. Its merged_error keeps a fallback default for unanswered statistics only so that this candidate's paned guarantee stays as it was.

Proposal

Drop CountSketch from AggIntent::Count. A CountSketch's signed counters don't add up to the total, and Count-Min and UnivMon already cover counts. Then remove the merged_error fallback, so a statistic the family doesn't answer fails closed (None).

The alternative is an executor kernel plus an accuracy contract for a CountSketch total, and then a (Total, AsIfBuiltOverUnion) row.

Scope / acceptance

• The example above lists no CountSketch candidate.
• merged_error returns None for any statistic outside the family's row, and the family_capability_table test asserts it for every row.
• Examples 1–4 select the same plans at the same costs.

Status / related

Found while reviewing #724, which keeps the fallback so behaviour is unchanged. Related: #696, #707.