fix(ir): top-k sketch readouts return selected rows; Example 1 expectations follow planner output

[zzylol]

Rebased on main d4869a7 (DF 54).

Why

#509 Example 1 runs end to end (1 → 24 → 24 → 1), but only 8 of the 24 Stage 2 candidates compiled in the runtime (physical_planner::compile). The acceptance tests still expected the spec's 6 candidates.

What

1. CountSketch+heap top-k. ASAPOp::output_schema derived SummaryEstimate{TopK} as partition keys + topk Utf8. No runtime operator produces or reads that encoding. The runtime's keyed evaluation, the exact Sort → Limit path and EvaluatePopulation{TopK} all return the selected rows, so keyed_evaluation rejected the IR's schema ("invalid keyed evaluation shape"). The IR now derives the selected rows: partition keys + item identity columns + Float64 value. With this, all three top-k implementations of Q2 return one row per selected series.
2. Count-Min+heap stays rejected. Q2 ranks sum_over_time of raw samples. Nothing in the workload or catalog declares http_requests_total non-negative, and neither existing proof (UnitCount, ResetAwareCounterDerivative) applies. The legacy rule also leaves this case UnknownOrSigned. Proving it would need a metric-type input (a counter or non-negative declaration) and a matching proof variant. The runtime and Stage 3 rules are unchanged and agree.
3. Expectations follow the planner (user decision). The acceptance spec and tests now expect 24 → 24 → 1. The exact-accumulator options account for the difference from 6. The spec lists all 24 candidates for manual review.

How

• crates/types/src/ir/operator/asap.rs: adds ranked_rows_schema for SketchStatistic::TopK. New unit test: structure_contract::topk_readout_derives_selected_rows.
• Regression test, which fails before the fix: stage2_count_sketch_heap_topk_compiles_in_the_physical_planner.
• stage2_runtime_compiles_exactly_the_candidates_stage3_finds_valid (replaces the ignored "every candidate compiles" test) checks that the runtime compiles the 16 valid candidates and rejects the 8 CMS+heap candidates, for the same reason Stage 3 gives.
• Un-ignored: stage1_has_24_candidates_covering_every_combination (was stage1_has_six…), stage2_keeps_every_logical_candidate, stage2_preserves_logical_choices, stage2_exact_topk_is_sort_then_limit (8 exact), stage2_summary_topk_is_build_then_estimate, and stage3_charges_each_node_once, which now covers the priced (valid) candidates.
• Still ignored, each naming the missing feature: stage1_q2_summary_families_are_heap_sketches_and_hydra (no Hydra), stage1_keeps_independent_and_shared_variants (Pass 2 sharing, Hydra), stage3_shared_input_is_not_costlier (Pass 2 sharing, Hydra).
• Fixture regenerated with stage_pipeline. Only the CountSketch/CMS estimate node schemas changed.

Before this PR: 8 of 24 Example 1 candidates compile in the runtime. The 8 CountSketch+heap candidates fail on the readout shape, and the 8 CMS+heap candidates fail on the non-negative weight rule. 10 acceptance tests are ignored.

After this PR: 16 of 24 compile, which is every candidate Stage 3 finds valid. The 8 CMS+heap candidates are rejected by both the runtime and Stage 3 for the non-negative weight reason. Stage 3 still selects P20 (all exact, 79.401 cpu ms), and every rejection reason is unchanged. 3 tests are ignored, each for a missing feature.

Follow-up: §5.2 of the docs/asap-primitive-schema branch still says topk Utf8 and should say "selected rows". This PR does not touch that branch.

🤖 Generated with Claude Code