test: tighten post-fix macOS performance budgets (Fixes #507)

docs/QUALITY_SNAPSHOTS.md

@@ -15,13 +15,15 @@ A metric blocks when it exceeds both its absolute and relative budget:
 | Metric | Linux | Windows | macOS |
 | --- | ---: | ---: | ---: |
 | Server startup P50 | 5 ms / 100% | 10 ms / 50% | 100 ms / 50% |
-| Server startup P95 | 50 ms / 200% | 50 ms / 100% | 10,000 ms / 100% |
+| Server startup P95 | 50 ms / 200% | 50 ms / 100% | 750 ms / 100% |
 | Full refresh P50 | 25 ms / 30% | 50 ms / 30% | 100 ms / 50% |
-| Full refresh P95 | 1,000 ms / 100% | 5,000 ms / 100% | 5,000 ms / 25% |
+| Full refresh P95 | 1,000 ms / 100% | 5,000 ms / 100% | 1,000 ms / 50% |
 | Time to first environment P50 | 20 ms / 100% | 25 ms / 50% | 150 ms / 50% |
-| Time to first environment P95 | 250 ms / 100% | 500 ms / 100% | 10,000 ms / 100% |
+| Time to first environment P95 | 250 ms / 100% | 500 ms / 100% | 750 ms / 100% |
 
-Each cell is `absolute / relative`. The budgets reflect observed GitHub-hosted runner variance from 11 consecutive main-branch baselines. Tighten them when a noisy path is fixed rather than normalizing a known regression into the baseline.
+Each cell is `absolute / relative`. The Linux and Windows budgets plus the macOS P50 budgets reflect observed GitHub-hosted runner variance from 11 consecutive main-branch baselines. Tighten them when a noisy path is fixed rather than normalizing a known regression into the baseline.
+
+The macOS P95 budget recalibration is tracked by issue #507 and follows PR #506's fix for issue #504. It uses three unchanged-content pull-request runs and the exact merged baseline at `f0c62d9`; the resulting absolute headroom is four to six times the observed post-fix run-to-run range.
 
 The dual budget avoids failing on tiny percentage changes while still blocking material latency regressions. Tail metrics remain mandatory; a healthy median does not excuse a degraded P95.
 
@@ -50,4 +52,4 @@ Phase and locator telemetry is collected in separate, untimed refreshes so diagn
 
 ## Known investigations
 
-The macOS cold-refresh tail is tracked by issue #504. Phase and locator distributions plus privacy-safe interpreter timeout counts verify that the tail does not recur.
+The macOS cold-refresh tail fixed by PR #506 (tracked in issue #504) remains guarded by phase and locator distributions plus privacy-safe interpreter timeout counts.

scripts/quality_snapshot.py

@@ -80,11 +80,11 @@ def regressed(self) -> bool:
     ),
     'macos': (
         RegressionBudget(100, 50),
-        RegressionBudget(10_000, 100),
+        RegressionBudget(750, 100),
         RegressionBudget(100, 50),
-        RegressionBudget(5_000, 25),
+        RegressionBudget(1_000, 50),
         RegressionBudget(150, 50),
-        RegressionBudget(10_000, 100),
+        RegressionBudget(750, 100),
     ),
 }
 COVERAGE_BUDGET_PERCENTAGE_POINTS = 0.01

scripts/tests/test_quality_snapshot.py

@@ -64,6 +64,27 @@ def test_p95_regression_fails_even_when_p50_is_unchanged(self):
         _, failures = compare_performance(current, performance_snapshot(refresh_p95=500), 'Windows')
         self.assertTrue(any('Full refresh P95' in failure for failure in failures))
 
+    def test_post_fix_macos_tail_variance_passes(self):
+        baseline = performance_snapshot(startup_p95=621, refresh_p95=1_343, first_p95=649)
+        current = performance_snapshot(startup_p95=691, refresh_p95=1_435, first_p95=745)
+
+        _, failures = compare_performance(current, baseline, 'macOS')
+
+        self.assertEqual(failures, [])
+
+    def test_tightened_macos_tail_budgets_reject_multi_second_regressions(self):
+        baseline = performance_snapshot(startup_p95=621, refresh_p95=1_343, first_p95=649)
+        current = performance_snapshot(startup_p95=1_500, refresh_p95=2_500, first_p95=1_500)
+
+        _, failures = compare_performance(current, baseline, 'macOS')
+
+        for label in (
+            'Server startup P95',
+            'Full refresh P95',
+            'Time to first environment P95',
+        ):
+            self.assertTrue(any(label in failure for failure in failures))
+
     def test_noise_inside_absolute_budget_passes(self):
         current = performance_snapshot(refresh_p50=140)
         _, failures = compare_performance(current, performance_snapshot(refresh_p50=100), 'Windows')