feat(promql): lower histogram_quantiles as a per-φ Merge fan-out (#109)

[zzylol]

Part of #109 — the histogram_quantiles half. The non-literal scalar k/φ half is split out to #120.

Left open for review, not merged.

Change

histogram_quantiles(v, "label", φ₀, φ₁, …) is the experimental multi-quantile form: histogram_quantile(φᵢ, v) evaluated once per φ, each output series tagged with label = φᵢ. It parsed but was rejected as UnsupportedFunction (19 corpus lines).

It lowers to a Merge of one Relabel-wrapped quantile branch per φ:

Merge {
  Relabel { q = "0.5", Aggregate{[Quantile(0.5)]} over <v> },
  Relabel { q = "0.9", Aggregate{[Quantile(0.9)]} over <v> },
}

Each branch reuses the single-quantile decision — classic le-buckets interpolate (HistogramQuantile), native histograms and raw samples take the sketch-able Quantile (#43 / #79) — so the two functions cannot diverge. The choice is a property of the argument, not of φ, so it is made once.

This is the first front end to emit L2::Merge, previously marked reserved.

Two details that were easy to get wrong

Label formatting. Prometheus uses labels.FormatOpenMetricsFloat, not Go’s default: 1 → "1.0", 0 → "0.0", and %g exponent form below 1e-4 (0.00001 → "1e-05"). I read model/labels/float.go rather than guess; Rust’s Display would have produced 1, 0, and 0.00001.

Union compatibility. Each branch aliases its value column to value rather than the intent-keyed name (quantile_0_5, quantile_0_9, …). Merge derives its schema from the first child, so branches disagreeing on a column name would make the merged schema silently misdescribe every branch but one. My first version had exactly that bug — merged schema = ["quantile_0_5", "q"] while branch 2 was ["quantile_0_9", "q"]. The quantile belongs in the label, which is where Prometheus puts it.

Corpus

testdata lowered 1502 → 1512; rejected 86 → 76.

Of the 19 histogram_quantiles lines, 10 now lower. The 9 that remain are pre-existing, deliberate rejections and not regressions:

• 7 out-of-range φ (-0.1, 1.01, NaN) — identical to histogram_quantile(1.001, …); whether we should follow Prometheus and return ±Inf with an annotation is the open question already noted in PromQL: histogram_quantiles() and non-literal scalar k/φ parameters unsupported #109.
• 2 __name__ regex selectors — PromQL: __name__=~/!~ regex selectors mislowered to a literal metric name #67.

The coverage tripwire is ratcheted from 1485 to 1495, which is what its own comment asks for when coverage lands.

Tests (6 new)

test	pins
histogram_quantiles_fans_out_over_native_histograms	one branch per φ, sketch-able Quantile
histogram_quantiles_over_classic_buckets_interpolates	_bucket → HistogramQuantile, never a sketch
histogram_quantiles_branches_are_union_compatible	the bug above
histogram_quantiles_uses_the_given_label_name	"phi", not a hardcoded "q"
histogram_quantiles_formats_small_quantiles_like_prometheus	1e-05
histogram_quantiles_rejects_an_out_of_range_quantile	one bad φ fails the call

Verification

cargo test --workspace      # 379 passed, 0 failed
cargo clippy --all-targets  # clean

cargo fmt --all not run — main carries pre-existing rustfmt diffs. Each touched file has the same violation count as on main.

🤖 Generated with Claude Code

[zzylol]

@milindsrivastava1997 @Selvomega Interesting "Merge" node.

[milindsrivastava1997]

Not sure I understand. Can you give some context?