test(promql): add a small o11y-bench-derived PromQL corpus

crates/frontend-promql/Cargo.toml

@@ -22,3 +22,7 @@ path = "tests/observability/awesome_prometheus_alerts.rs"
 [[test]]
 name = "promql_corpus"
 path = "tests/observability/promql_corpus.rs"
+
+[[test]]
+name = "o11y_bench_promql"
+path = "tests/observability/o11y_bench_promql.rs"

crates/frontend-promql/tests/observability/data/o11y_bench_promql.txt

@@ -0,0 +1,37 @@
+# PromQL query corpus derived from grafana/o11y-bench task-grading rubrics.
+# Source: https://github.com/grafana/o11y-bench (tasks-spec/**/*.yaml, `fact.query`
+#   values where backend == prometheus). Snapshot fetched 2026-07-17; deduplicated.
+# License: AGPL-3.0. Vendoring this corpus verbatim from an AGPL-3.0 source into a
+#   differently-licensed test suite has an open licensing question -- see issue #135.
+#   Do not treat this file's provenance as cleared; it is tracked, not resolved.
+# One real-world PromQL expression per line, pulled from LLM-agent-benchmark grading
+# rubrics (SRE-style incident/investigation queries: error ratios, burn rate, cache
+# lag, retry backlog, capacity). Used by tests/observability/o11y_bench_promql.rs to
+# pin parse + lowering behaviour over real-world queries (totality; no execution).
+(sum by (job) (increase(http_requests_total{status=~"5..",job=~"user-service|order-service|payment-service"}[6h]))) / (sum by (job) (increase(http_requests_total{job=~"user-service|order-service|payment-service"}[6h])))
+histogram_quantile(0.95, sum(rate(http_request_duration_seconds_bucket{job="order-service"}[5m])) by (le))
+histogram_quantile(0.99, sum(rate(http_request_duration_seconds_bucket{job="order-service"}[6h])) by (le))
+max((sum by (job) (rate(http_requests_total{status=~"5..",job=~"user-service|order-service|payment-service"}[6h]))) / (sum by (job) (rate(http_requests_total{job=~"user-service|order-service|payment-service"}[6h]))))
+max_over_time((sum(rate(http_requests_total{job="order-service",status=~"5.."}[5m])) / sum(rate(http_requests_total{job="order-service"}[5m])))[6h:1m])
+max_over_time(histogram_quantile(0.95, sum(rate(http_request_duration_seconds_bucket{job="user-service"}[5m])) by (le))[12h:1m])
+max_over_time(service_cache_refresh_lag_seconds{job="user-service"}[12h])
+max_over_time(service_retry_queue_depth{job="order-service"}[6h])
+max_over_time(service_retry_queue_depth{job="payment-service"}[6h])
+max_over_time(service_retry_queue_depth{job=~".+"}[6h])
+max_over_time(sum(process_resident_memory_bytes)[6h:1m])
+service_cache_refresh_lag_seconds{job="user-service"}
+sort_desc(sum by (job) (rate(process_cpu_seconds_total{job=~".+"}[6h])))
+sum by (job) (increase(http_requests_total{status=~"5..",job=~".+-service"}[24h])) > 0
+sum by (job) (rate(process_cpu_seconds_total{job=~".+"}[6h]))
+sum(avg_over_time(process_resident_memory_bytes{job=~".+"}[6h]))
+sum(increase(http_requests_total{job="order-service",status=~"5.."}[24h])) / sum(increase(http_requests_total{status=~"5.."}[24h]))
+sum(increase(http_requests_total{job="payment-service",status=~"5.."}[1h] offset 6h)) / sum(increase(http_requests_total{job="payment-service"}[1h] offset 6h))
+sum(increase(http_requests_total{job="payment-service",status=~"5.."}[1h])) / sum(increase(http_requests_total{job="payment-service"}[1h]))
+sum(process_resident_memory_bytes)
+sum(rate(http_requests_total{job="order-service"}[5m] offset 1h))
+sum(rate(http_requests_total{job="order-service"}[5m]))
+sum(rate(http_requests_total{job=~"user-service|order-service|payment-service"}[5m]))
+sum(rate(http_requests_total{status=~"5..",job=~"payment-service|order-service"}[6h])) / sum(rate(http_requests_total{job=~"payment-service|order-service"}[6h]))
+sum(rate(http_requests_total{status=~"5..",job=~"user-service|order-service|payment-service"}[1h])) / sum(rate(http_requests_total{job=~"user-service|order-service|payment-service"}[1h]))
+sum(rate(process_cpu_seconds_total[1h]))
+sum(up)

crates/frontend-promql/tests/observability/o11y_bench_promql.rs

@@ -0,0 +1,56 @@
+//! PromQL conformance over a small corpus derived from **grafana/o11y-bench**
+//! task-grading rubrics.
+//!
+//! Source: <https://github.com/grafana/o11y-bench> — 27 deduplicated `fact.query`
+//! values (`backend: prometheus`) pulled from `tasks-spec/**/*.yaml` grading
+//! rubrics (`tests/data/o11y_bench_promql.txt`). These are SRE-style
+//! incident/investigation queries: error ratios, burn rate, cache-lag and
+//! retry-backlog triage, capacity checks.
+//!
+//! LICENSE NOTE: o11y-bench is AGPL-3.0, unlike this repo's other vendored
+//! corpora (MIT/Apache-2.0). Whether copying these short query fixtures
+//! verbatim into a differently-licensed test suite is fine as-is is an open
+//! question — tracked in issue #135, not resolved by this file's existence.
+//!
+//! We *lower* (parse → L2 → L3), we do not execute. Totality: every query
+//! returns `Ok` or a clean `LoweringError` and never panics. Given the corpus
+//! is small and hand-picked from realistic incident-response queries, we also
+//! assert full lowering coverage — a regression here means a real pattern
+//! broke, not statistical noise.
+
+use asap_frontend_promql::{lower_promql, PromqlError as LoweringError};
+use asap_ir::types::AccuracyTarget;
+
+const CORPUS: &str = include_str!("data/o11y_bench_promql.txt");
+
+/// Non-comment, non-blank query lines.
+fn queries() -> impl Iterator<Item = &'static str> {
+    CORPUS
+        .lines()
+        .map(str::trim)
+        .filter(|l| !l.is_empty() && !l.starts_with('#'))
+}
+
+#[test]
+fn lowering_is_total_over_the_o11y_bench_corpus() {
+    let mut lowered = 0;
+    let mut rejected = Vec::new();
+
+    for q in queries() {
+        match lower_promql(q, AccuracyTarget::Exact) {
+            Ok(_) => lowered += 1,
+            Err(LoweringError::Parse(e)) => panic!("unexpected parse failure for {q:?}: {e}"),
+            Err(e) => rejected.push((q, e)),
+        }
+    }
+
+    assert!(
+        rejected.is_empty(),
+        "expected every o11y-bench query to lower cleanly, but {} were rejected: {rejected:#?}",
+        rejected.len()
+    );
+    assert_eq!(
+        lowered, 27,
+        "corpus size changed — update this assertion if the corpus was intentionally edited"
+    );
+}