feat: lower and compile PromQL histogram_quantile

[zzylol]

Stacked on #494.

Rebuilt into the linear stack on main. New integrate: commit(s) fold in integration-branch resolutions this PR needs on top of the earlier stack: "bind the histogram input layout in series_labels". Conflict resolutions are recorded in the messages of: "feat(physical): compile classic histogram_quantile in PromQL fallbacks"; "docs: record classic histogram_quantile coverage".

Why

compile rejected PromQL histogram_quantile (coverage row 10). The IR was Aggregate{by (), [HistogramQuantile{q}]}, so the compiler could not tell which series form one histogram, and the output lost its labels.

What

• IR: AggIntent::HistogramQuantile { q, le: C } names the bucket-bound column. PromQL lowers the classic form to Aggregate{Reduce(without([le])), [HistogramQuantile{q, le}]}. The without keys seed le into the selector schema even when no matcher names it. The SQL asap_histogram_quantile bridge keeps its single-histogram by () and now names le.
• Types fix: without output schemas no longer keep a nested aggregate's renamed value (e.g. sum) as a label.
• Physical: a new SeriesHistogramQuantile operator. It groups rows by every label except le (keeping __name__, as Prometheus does), applies bucketQuantile, and drops le and __name__. Result label sets that become equal are an error, as in Prometheus. It works on series-identity rows and plain label columns, so it covers rate(x_bucket[5m]), sum by (le, job) (…) and bare selectors.
• bucket_quantile now matches Prometheus:
  • almost.Equal for the small-delta fix, so an infinite count is no longer merged into a finite one.
  • Divide-then-scale interpolation.
  • A NaN rank finds no bucket.
  • le values that ParseFloat rejects as out of range are skipped.
• Selection: candidate search still keeps this intent as one exact KeepPreAsap subtree for exact and approximate targets. No sketch candidates were added.

Before this PR

histogram_quantile(0.5, sum by (le, job) (rate(x_bucket[5m]))) → compile error: "vector aggregate has no native lowering". Even if it had compiled, the IR described one quantile over all buckets with no labels.

After this PR

With buckets {1: 2, 2: 6, 4: 8, +Inf: 10} per job=a, the query compiles through search and selection and returns {job="a"} 1.75. The hand-computed Prometheus result is 1 + (2-1)·((5-2)/4). A per-series rate(x_bucket[5m]) argument keeps {inst, job}.

Known limits:

• An argument that provably lacks le (sum by (job) (…)) is rejected at lowering. Prometheus returns an empty vector for it.
• histogram_quantiles and aggregating over the result still don't compile. See the coverage doc.

Validation

• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace --no-fail-fast pass (1401 tests).
• New end-to-end tests in asap-physical-operators/tests/promql_fallback.rs check hand-computed results for:
  • interpolation and the +Inf bucket;
  • q<0, q>1 and NaN q;
  • a missing +Inf bucket, fewer than two buckets, and zero observations;
  • unsorted buckets, unparsable and out-of-range le values, duplicate bounds and non-monotonic counts;
  • a lowest bucket bound ≤ 0;
  • rate and sum shapes;
  • duplicate output label sets;
  • search and selection.
• There are also frontend lowering and unit tests.
• Each fix has a regression test that failed before the fix.
• An independent agent reviewed the change. Its NaN-rank, ParseFloat-overflow and SQL-schema findings are fixed here.

🤖 Generated with Claude Code