Measure saturated upstream block load

Author: Arbousier1

.github/workflows/upstream-runtime-comparison.yml

@@ -930,11 +930,16 @@ jobs:
           mean_nonregression = float(mean["ratioBootstrap95Ci"][1]) <= 1.05
           p95_nonregression = float(p95["ratioBootstrap95Ci"][1]) <= 1.05
           p99_nonregression = float(p99["ratioBootstrap95Ci"][1]) <= 1.10
+          candidate_tps_healthy = all(
+              19.9 <= float(value["observedTps"]) <= 20.5
+              for value in metrics_by_variant["B"]
+          )
           scenario_passed = formal and all((
               primary_improvement,
               mean_nonregression,
               p95_nonregression,
               p99_nonregression,
+              candidate_tps_healthy,
           ))
           if not formal:
               conclusion = "exploratory-no-winner"
@@ -954,11 +959,20 @@ jobs:
               "",
               f"A is official upstream Jenkins #163; B is production candidate "
               f"`{candidate_source[:7]}`. ",
-              "Smoke evidence is exploratory and never declares a winner.",
+              ("Formal evidence is evaluated against the pre-registered gates."
+               if formal else
+               "Smoke evidence is exploratory and never declares a winner."),
               "",
               "| Metric | Upstream median | Rewrite median | Median B/A | Ratio 95% CI | Registered use |",
               "|---|---:|---:|---:|---:|---|",
           ]
+          if formal and scenario == "block-active":
+              lines[5:5] = [
+                  "Load-regime rule: sustained block load may saturate official upstream A "
+                  "when the 120-second window still contains at least 100 tick samples; "
+                  "candidate B must remain in the 19.9-20.5 TPS regime.",
+                  "",
+              ]
           for metric, unit in metric_specs:
               analysis = analyses[metric]
               a_median = statistics.median(float(value[metric]) for value in metrics_by_variant["A"])
@@ -986,7 +1000,12 @@ jobs:
               "",
               f"Scenario conclusion: **{conclusion}**.",
               "",
-              "`observedTps` is capped near 20 on a healthy server; MSPT is the primary effect-size evidence.",
+              ("Candidate `observedTps` must stay near 20; saturated upstream throughput is "
+               "reported directly for the sustained block load. MSPT remains the primary "
+               "effect-size evidence."
+               if formal and scenario == "block-active" else
+               "`observedTps` is capped near 20 on a healthy server; MSPT is the primary "
+               "effect-size evidence."),
               "",
               f"- Upstream source/artifact: `{upstream_source}` / `{upstream_artifact}`",
               f"- Candidate source/artifact: `{candidate_source}` / `{candidate_artifact}`",
@@ -1025,6 +1044,17 @@ jobs:
                       "ratioCiUpperMaximum": 1.10,
                       "passed": p99_nonregression if formal else None,
                   },
+                  "candidateHealthyTps": {
+                      "minimum": 19.9,
+                      "maximum": 20.5,
+                      "passed": candidate_tps_healthy if formal else None,
+                  },
+              },
+              "loadRegimePolicy": {
+                  "upstreamSaturationAllowed": formal and scenario == "block-active",
+                  "minimumUpstreamTickSamples": (
+                      100 if formal and scenario == "block-active" else None
+                  ),
               },
               "metrics": {
                   metric: {

tools/perf/run-upstream-runtime-once.sh

@@ -561,17 +561,31 @@ if data.get("seconds", 0) < measure - 2 or data.get("seconds", 0) > measure + 3:
     raise SystemExit(f"comparison measurement duration is invalid: {data.get('seconds')}")
 if data.get("tickSamples", 0) <= 0 or data.get("observedTps", 0) <= 0:
     raise SystemExit("comparison produced no valid tick samples")
-minimum_tps = 19.9 if measure >= 60 else 19.5
-if not minimum_tps <= data["observedTps"] <= 20.5:
-    raise SystemExit(
-        f"comparison did not remain in the healthy 20 TPS regime "
-        f"({minimum_tps}..20.5): {data['observedTps']}"
-    )
-minimum_tick_samples = (
-    math.floor(data["seconds"] * minimum_tps)
-    - data["boundaryTickSamplesDiscarded"]
-    - 1
+allows_upstream_saturation = (
+    measure >= 60 and scenario == "block-active" and variant == "A"
 )
+if allows_upstream_saturation:
+    # Sustained block load is allowed to expose the official upstream's
+    # throughput ceiling. The rewritten candidate remains subject to the
+    # healthy-20-TPS guard, and the upstream sample must still contain enough
+    # tick observations for stable percentile estimates.
+    if data["observedTps"] > 20.5:
+        raise SystemExit(
+            f"comparison exceeded the server TPS ceiling: {data['observedTps']}"
+        )
+    minimum_tick_samples = 100
+else:
+    minimum_tps = 19.9 if measure >= 60 else 19.5
+    if not minimum_tps <= data["observedTps"] <= 20.5:
+        raise SystemExit(
+            f"comparison did not remain in the healthy 20 TPS regime "
+            f"({minimum_tps}..20.5): {data['observedTps']}"
+        )
+    minimum_tick_samples = (
+        math.floor(data["seconds"] * minimum_tps)
+        - data["boundaryTickSamplesDiscarded"]
+        - 1
+    )
 if data["tickSamples"] < minimum_tick_samples:
     raise SystemExit(f"comparison recorded too few tick samples: {data['tickSamples']}")
 for key in ("msptP50", "msptP95", "msptP99", "msptP999", "msptMax", "msptMean"):