eval: export pooled quantiles and top-k from random manifests

[milindsrivastava1997]

Summary

• --random-manifest now exports RQEs that an ASAPQuery random workload marks pooled:
  • Pooled quantile: one quantile over all of a group's samples. The RQE and tree leaf stay at the query's grouping (one sketch per group), not per series.
  • Pooled top-k: topk(k, sum by (G) (...)), one global heap over G's groups. Keys = G's cardinality, read at the curves' 1,000 keys with an assumption noted when fewer.
• A pooled query's e2e entry is its tree alone (expr). ASAPQuery's saved planner then plans it without text (--expr-only). Its PromQL, which the other systems run, moves to client_promql.
• A pooled quantile's PromQL is checked against the stand-in shape quantile by (G) (phi, quantile_over_time(phi, data[W])). Non-pooled quantiles keep the avg by shape.
• Pooled and per-series RQEs with the same fields stay distinct (pooled is part of the sharing key).

Used by ProjectASAP/ASAPQuery codex/random-pooled (the pooled and dashboard-style random workloads).

Validation

• python3 -m pytest scripts/test_export_autosketch_eval_table.py: 56 pass, with 4 new tests:
  • pooled quantile leaf;
  • pooled top-k leaf;
  • per-group / global-heap table RQEs and tree-only e2e entries;
  • pooled vs per-series RQEs kept apart.
• Exported all 10 seed-862-pooled and 10 seed-862-dashboard manifests. The solver and ASAPQuery's --planner saved --expr-only plan every one with nothing left out.
• An end-to-end pilot on 5 systems: ASAPQuery's pooled answers agree with exact ClickHouse.

🤖 Generated with Claude Code

https://claude.ai/code/session_015KzRxtgjFfFJeyDy8fN2CB