scorecard: label a contended baseline and say where the regression check runs (#7494)

The 2026-10-05 baseline was recorded at load 224.9 on 14 cores; the scorecard now
says so without changing a number, shows the baseline's run count, and replaces
the claim that the nightly checks regressions with the machine class it actually
runs on.

docs/scorecard.md

@@ -18,9 +18,11 @@ Generated by `cargo xtask scorecard` from `perf/budgets.toml`, `perf/baseline.js
 
 ## Performance
 
-Budgets: `perf/budgets.toml` (targets from #209, #210, #211). Numbers: the committed baseline (`perf/baseline.json`, full run), updated only by `cargo xtask perf --update-baseline`. The nightly `perf-nightly` workflow runs `cargo xtask perf` on a fixed macOS runner and fails on a broken enforced budget or a p50 regression over 15 % against a baseline from the same machine class. "Over budget" rows are targets the code doesn't meet yet: they are reported, and only their regressions fail.
+Budgets: `perf/budgets.toml` (targets from #209, #210, #211). Numbers: the committed baseline (`perf/baseline.json`, full run), updated only by `cargo xtask perf --update-baseline`. The nightly `perf-nightly` workflow runs `cargo xtask perf` on a fixed macOS runner and fails on a broken enforced budget. Its p50 regression check (over 15 %, against the baseline's highest p50 when it was built from several runs) runs only on machine class `macos-aarch64-apple-m4-pro-apple-m4-pro`, the baseline's. "Over budget" rows are targets the code doesn't meet yet: they are reported, and only their regressions fail.
 
-Baseline: 2026-10-05, commit `f5d2bcb54a`, Apple M4 Pro (48 GB), GPU Apple M4 Pro, macOS 15.6; load average 224.90 / 226.31 / 196.03 (1/5/15 min) at the start of the run.
+Baseline: 2026-10-05, 1 run, commit `f5d2bcb54a`, Apple M4 Pro (48 GB), GPU Apple M4 Pro, macOS 15.6; load average 224.90 / 226.31 / 196.03 (1/5/15 min) at the start of the run.
+
+**contended baseline** (load 224.90 on 14 cores): these numbers measure machine load, not the code.
 
 | Id | Scenario | Target | p50 ms | p95 ms | max ms | Peak RSS | GPU | Status | Issue |
 |---|---|---|---:|---:|---:|---:|---:|---|---|

xtask/src/scorecard.rs

@@ -16,7 +16,7 @@ use std::collections::{BTreeMap, BTreeSet};
 use std::path::{Path, PathBuf};
 
 use serde::Deserialize;
-use serde_json::Value;
+use serde_json::{Value, json};
 
 use crate::perf;
 
@@ -560,12 +560,14 @@ here supersede estimates elsewhere.\n\n## Summary\n\n| Area | Done | Partial | M
 
     // Performance.
     s.push_str("\n## Performance\n\n");
+    let class = inp.baseline.as_ref().and_then(|b| b.get("machine_class")).and_then(Value::as_str).unwrap_or("none");
     s.push_str(&format!(
         "Budgets: `perf/budgets.toml` (targets from #209, #210, #211). Numbers: the committed baseline \
 (`perf/baseline.json`, full run), updated only by `cargo xtask perf --update-baseline`. The nightly \
 `perf-nightly` workflow runs `cargo xtask perf` on a fixed macOS runner and fails on a broken enforced \
-budget or a p50 regression over {:.0} % against a baseline from the same machine class. \"Over budget\" \
-rows are targets the code doesn't meet yet: they are reported, and only their regressions fail.\n\n",
+budget. Its p50 regression check (over {:.0} %, against the baseline's highest p50 when it was built from \
+several runs) runs only on machine class `{class}`, the baseline's. \"Over budget\" rows are targets the \
+code doesn't meet yet: they are reported, and only their regressions fail.\n\n",
         inp.budgets.settings.regression_pct
     ));
     match (&base_date, full) {
@@ -573,12 +575,22 @@ rows are targets the code doesn't meet yet: they are reported, and only their re
             let load = perf::fmt_load(f.get("load_avg_start"));
             let commit = f.get("commit").and_then(Value::as_str).map_or_else(|| "unknown".into(), |c| c.get(..10).unwrap_or(c).to_string());
             let ram = machine.and_then(|m| m.get("ram_bytes")).and_then(Value::as_u64).map_or_else(|| "?".into(), |b| format!("{} GB", b >> 30));
+            let runs = f.get("runs").and_then(Value::as_u64).unwrap_or(1);
+            let runs = if runs == 1 { "1 run".to_string() } else { format!("{runs} runs (median p50; the latest run's date and commit)") };
             s.push_str(&format!(
-                "Baseline: {d}, commit `{commit}`, {} ({ram}), GPU {}, {}; load average {load} (1/5/15 min) at the start of the run.\n\n",
+                "Baseline: {d}, {runs}, commit `{commit}`, {} ({ram}), GPU {}, {}; load average {load} (1/5/15 min) at the start of the run.\n\n",
                 ms("cpu"),
                 ms("gpu_adapter"),
                 ms("os_version")
             ));
+            let cores = f.get("cores").or_else(|| machine.and_then(|m| m.get("cores")));
+            let contended = f.get("contended").and_then(Value::as_bool) == Some(true)
+                || perf::run_contended(f.get("load_avg_start"), Some(&json!({"cores": cores})), inp.budgets.settings.max_load_per_core);
+            if contended {
+                let one = f.get("load_avg_start").and_then(|l| l.get(0)).and_then(Value::as_f64).map_or_else(|| "?".into(), |x| format!("{x:.2}"));
+                let n = cores.and_then(Value::as_u64).map_or_else(|| "?".into(), |c| c.to_string());
+                s.push_str(&format!("**contended baseline** (load {one} on {n} cores): these numbers measure machine load, not the code.\n\n"));
+            }
         }
         _ => s.push_str("Baseline: **not run** (no `perf/baseline.json` full run yet).\n\n"),
     }
@@ -808,6 +820,29 @@ const PHOTOSHOP: Source = Source { label: "photoshop oracles", dir: "corpus/phot
         assert_eq!(anchor("File compatibility"), "file-compatibility");
     }
 
+    #[test]
+    fn scorecard_labels_a_contended_baseline() {
+        let mut inp = gather(&crate::root()).unwrap_or_else(|e| panic!("{e}"));
+        let base = |load: f64, extra: Value| {
+            let mut full = json!({"date": "2026-10-05T00:00:00Z", "commit": "f5d2bcb54abd", "load_avg_start": [load, 1.0, 1.0], "scenarios": {}});
+            if let (Some(f), Some(x)) = (full.as_object_mut(), extra.as_object()) {
+                f.extend(x.clone());
+            }
+            json!({"machine_class": "mac-m4", "machine": {"cores": 14, "cpu": "Apple M4 Pro"}, "modes": {"full": full}})
+        };
+        inp.baseline = Some(base(224.9, json!({})));
+        let a = render(&inp);
+        assert!(a.contains("**contended baseline** (load 224.90 on 14 cores)"), "labelled from the load the baseline already has");
+        assert!(a.contains("runs only on machine class `mac-m4`"));
+        assert!(a.contains("2026-10-05, 1 run,"));
+        inp.baseline = Some(base(2.0, json!({"runs": 5})));
+        let a = render(&inp);
+        assert!(!a.contains("contended baseline"));
+        assert!(a.contains("5 runs (median p50"));
+        inp.baseline = Some(base(2.0, json!({"contended": true, "cores": 3})));
+        assert!(render(&inp).contains("(load 2.00 on 3 cores)"), "a baseline recorded with --allow-contended");
+    }
+
     #[test]
     fn staleness_check_and_determinism() {
         let root = crate::root();