perf: large-page reuse CPU investigation and draft V7 (#614)

.gitattributes

@@ -20,3 +20,23 @@
 
 # Preserve the exact issue 615 benchmark verification archive bytes.
 .github/assets/benchmark-615/verification.zip binary
+# Preserve the exact issue 614 attribution archive bytes.
+.github/assets/perf-614/attribution.zip binary
+.github/assets/perf-614/platform-gates.zip binary
+.github/assets/perf-614/v1-measurements.zip binary
+.github/assets/perf-614/v2-structural.zip binary
+.github/assets/perf-614/v2-original-platform.zip binary
+.github/assets/perf-614/v2-test-fix-evidence.zip binary
+.github/assets/perf-614/v2-measurements.zip binary
+.github/assets/perf-614/v3-structural.zip binary
+.github/assets/perf-614/v2-corrected-platform.zip binary
+.github/assets/perf-614/v4-structural.zip binary
+.github/assets/perf-614/v3-measurements.zip binary
+.github/assets/perf-614/v3-platform.zip binary
+.github/assets/perf-614/v4-platform.zip binary
+.github/assets/perf-614/v4-measurements.zip binary
+.github/assets/perf-614/v5-preflight.zip binary
+.github/assets/perf-614/v6-structural.zip binary
+.github/assets/perf-614/v6-measurements.zip binary
+.github/assets/perf-614/v6-platform.zip binary
+.github/assets/perf-614/v7-validation-evidence.zip binary

.github/assets/perf-614/attribution-archive.json

@@ -0,0 +1,6 @@
+{
+  "file": "attribution.zip",
+  "sha256": "91ffc47dfe119a5d8bb78c4aefa511685f855f06f41955c54d8f88d8163690cd",
+  "files": 43,
+  "purpose": "local attribution and structural evidence, no timing acceptance"
+}

.github/assets/perf-614/ledger-v1.md

@@ -0,0 +1,44 @@
+## Stage-one result: candidate not accepted
+
+Exact baseline `227eef06ff03f92966f38508441d6c2f60a75135`, candidate `287cac2f664a46703ebb6522eef8c8dbeb984cbb`. All three original cohorts completed with 120 actual samples per arm (base/head/null), unchanged traces and sanity gates. Original executables/hash manifests, raw samples, workflow logs and independently recomputed ledgers are retained. No hosts are pooled.
+
+**No statistically supported target CPU or throughput gain.** Across 18 metrics: 10 PASS, 8 INCONCLUSIVE, 0 FAIL; zero memory credit in every row, all release bounds held. The stage-two condition is not met. I will not expand this candidate to the remaining rows or rerun it unchanged. It is rejected for delivery despite successful correctness/platform gates.
+
+The measured chart binaries confirm the intended prologue removal, but structural work reduction is not a measured benefit. CPU changes in both ephemeral targets lean worse, and the plain-ephemeral RSS interval is unresolved; neither is silently relabelled as equivalence. Further implementation requires fresh attribution within the original CPU-only scope, not weaker thresholds.
+
+### large-class/8
+
+Original run: https://github.com/zackees/mimalloc-pprof/actions/runs/39823333694
+
+base `227eef06ff03` head `287cac2f664a` on AMD EPYC 7763 64-Core Processor, 60 reps (up to 120); 4 logical CPUs, 8-worker cells have 2.00 workers per logical CPU (above 1 means oversubscription); arm order: rotating-reversed-blocks-v1
+
+Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py): `saved%` of the reducible peak-RSS gap, then per metric the paired change [95% sign-test interval] against its allowance (saved%/3 for CPU, throughput and faults; noise for memory and for cells that saved none). INCONCLUSIVE is not a pass.
+
+| workload | saved% | verdict | ops/s | cpu s | minor faults | p99 operation ns | peak RSS MiB | RSS 0.5 s after drain MiB |
+|---|---|---|---|---|---|---|---|---|
+| large-class/8 | 0% | **INCONCLUSIVE** | +0.1% [-0.6, +0.6] / 0.7% **PASS** | -0.0% [-0.4, +0.3] / 0.3% **INCONCLUSIVE** | -0.6% [-2.0, +1.4] / 2.2% **PASS** | +0.1% [-1.0, +0.6] / 1.1% **PASS** | -0.0% [-0.2, +0.0] / 0.1% **PASS** | -0.4% [-1.8, +0.1] / 1.8% **PASS** |
+
+### large-class-ephemeral/8
+
+Original run: https://github.com/zackees/mimalloc-pprof/actions/runs/39823337377
+
+base `227eef06ff03` head `287cac2f664a` on AMD EPYC 7763 64-Core Processor, 60 reps (up to 120); 4 logical CPUs, 8-worker cells have 2.00 workers per logical CPU (above 1 means oversubscription); arm order: rotating-reversed-blocks-v1
+
+Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py): `saved%` of the reducible peak-RSS gap, then per metric the paired change [95% sign-test interval] against its allowance (saved%/3 for CPU, throughput and faults; noise for memory and for cells that saved none). INCONCLUSIVE is not a pass.
+
+| workload | saved% | verdict | ops/s | cpu s | minor faults | p99 operation ns | peak RSS MiB | RSS 0.5 s after drain MiB |
+|---|---|---|---|---|---|---|---|---|
+| large-class-ephemeral/8 | 0% | **INCONCLUSIVE** | -0.4% [-0.7, -0.1] / 0.5% **INCONCLUSIVE** | +0.3% [+0.2, +0.4] / 0.3% **INCONCLUSIVE** | +0.0% [+0.0, +0.8] / 0.8% **INCONCLUSIVE** | -0.0% [-1.0, +0.6] / 1.0% **PASS** | +0.0% [-0.0, +2.6] / 0.1% **INCONCLUSIVE** | +0.0% [-9.4, +0.8] / 9.8% **PASS** |
+
+### large-class-ephemeral/8 short generations (chart build)
+
+Original run: https://github.com/zackees/mimalloc-pprof/actions/runs/39823341873
+
+base `227eef06ff03` head `287cac2f664a` on AMD EPYC 9V74 80-Core Processor, 60 reps (up to 120); 4 logical CPUs, 8-worker cells have 2.00 workers per logical CPU (above 1 means oversubscription); arm order: rotating-reversed-blocks-v1
+
+Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py): `saved%` of the reducible peak-RSS gap, then per metric the paired change [95% sign-test interval] against its allowance (saved%/3 for CPU, throughput and faults; noise for memory and for cells that saved none). INCONCLUSIVE is not a pass.
+
+| workload | saved% | verdict | ops/s | cpu s | minor faults | p99 operation ns | peak RSS MiB | RSS 0.5 s after drain MiB |
+|---|---|---|---|---|---|---|---|---|
+| large-class-ephemeral/8 short generations (chart build) | 0% | **INCONCLUSIVE** | -0.3% [-0.9, +0.4] / 0.8% **INCONCLUSIVE** | +0.4% [+0.1, +0.9] / 0.4% **INCONCLUSIVE** | -0.4% [-0.8, +0.0] / 0.7% **PASS** | +1.4% [-0.7, +2.1] / 1.4% **INCONCLUSIVE** | +0.0% [-0.0, +0.0] / 2.6% **PASS** | +0.0% [-0.0, +0.0] / 2.6% **PASS** |
+

.github/assets/perf-614/ledger-v2.md

@@ -0,0 +1,52 @@
+# V2 original stage-one ledger
+
+Baseline `227eef06ff03f92966f38508441d6c2f60a75135`; candidate `59d369609ab47358df0f883b6419252f1ffbf03a`.
+
+STOP: no statistically supported target CPU/throughput gain
+
+Original per-host paired cohorts only; zero memory credit. Intervals and verdicts are retained.
+
+## large-class/8
+
+Run 39831495052; Intel(R) Xeon(R) Platinum 8370C CPU @ 2.80GHz; samples per arm: {'base': 60, 'head': 60, 'null': 60}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | +0.5714 [-1.3148, +1.2130] | 1.4155 | PASS |
+| cpu s | -0.4897 [-1.0908, +0.5073] | 0.8227 | PASS |
+| minor faults | -1.4996 [-3.8674, +0.4237] | 2.0921 | PASS |
+| p99 operation ns | +0.9573 [-0.6498, +1.7575] | 2.1407 | PASS |
+| peak RSS MiB | +0.0559 [-0.0766, +2.5338] | 2.5809 | PASS |
+| RSS 0.5 s after drain MiB | -0.1192 [-2.1851, +1.9511] | 3.7011 | PASS |
+
+Release p95: 1100.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8
+
+Run 39831499063; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | +0.3751 [-0.0942, +0.5970] | 0.4314 | PASS |
+| cpu s | -0.0879 [-0.1899, +0.0749] | 0.1690 | PASS |
+| minor faults | +0.0000 [-0.7722, +0.3906] | 0.7812 | PASS |
+| p99 operation ns | +0.8061 [+0.4727, +1.5848] | 0.5456 | INCONCLUSIVE |
+| peak RSS MiB | +0.0000 [-0.0733, +0.1626] | 2.5703 | PASS |
+| RSS 0.5 s after drain MiB | +0.0227 [-0.7276, +10.9645] | 0.7329 | INCONCLUSIVE |
+
+Release p95: 1050.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8 short generations (chart build)
+
+Run 39831504283; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.4759 [-1.2258, +0.3828] | 0.7212 | INCONCLUSIVE |
+| cpu s | +0.2379 [-0.0983, +0.5851] | 0.3687 | INCONCLUSIVE |
+| minor faults | +0.1873 [-0.3861, +0.7722] | 0.7692 | INCONCLUSIVE |
+| p99 operation ns | +0.5108 [-0.0539, +1.1457] | 1.0678 | INCONCLUSIVE |
+| peak RSS MiB | +0.0000 [-0.0314, +0.0054] | 0.0430 | PASS |
+| RSS 0.5 s after drain MiB | +0.0000 [-0.0314, +0.0052] | 0.0052 | INCONCLUSIVE |
+
+Release p95: 1010.0 ms; bound: 1300 ms; held: True.

.github/assets/perf-614/ledger-v3.md

@@ -0,0 +1,53 @@
+# V3 original stage-one ledger
+
+Baseline `227eef06ff03f92966f38508441d6c2f60a75135`; candidate `5b2d8f960297fda825b05b7471893fc4cf6c2f33`.
+
+REJECT V3: two confirmed per-cell regressions; no stage two or unchanged rerun
+
+Original per-host paired cohorts only; zero memory credit. Intervals and verdicts are retained.
+
+## large-class/8
+
+Run 39840352295; AMD EPYC 9V74 80-Core Processor; samples per arm: {'base': 110, 'head': 110, 'null': 110}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.4470 [-1.0644, -0.0539] | 0.9693 | INCONCLUSIVE |
+| cpu s | +0.6433 [+0.4006, +0.8918] | 0.3584 | FAIL |
+| minor faults | -0.4376 [-1.7021, +1.3158] | 1.7621 | PASS |
+| p99 operation ns | +1.8763 [+0.7396, +2.9586] | 1.5129 | INCONCLUSIVE |
+| peak RSS MiB | +0.0000 [-0.0052, +0.0051] | 0.0052 | PASS |
+| RSS 0.5 s after drain MiB | +0.0000 [-0.0052, +0.0051] | 0.0052 | PASS |
+
+Release p95: 1120.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8
+
+Run 39840357033; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 75, 'head': 75, 'null': 75}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.8333 [-1.0046, -0.5076] | 0.3310 | FAIL |
+| cpu s | +0.3109 [+0.0964, +0.4994] | 0.1852 | INCONCLUSIVE |
+| minor faults | +0.0000 [-0.3906, +0.7968] | 1.5444 | PASS |
+| p99 operation ns | +0.6169 [-0.0564, +1.6256] | 1.0839 | INCONCLUSIVE |
+| peak RSS MiB | +0.1045 [+0.0052, +2.6618] | 2.6071 | INCONCLUSIVE |
+| RSS 0.5 s after drain MiB | +10.9141 [+0.0000, +14.1152] | 13.1240 | INCONCLUSIVE |
+
+Release p95: 1050.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8 short generations (chart build)
+
+Run 39840362233; AMD EPYC 9V74 80-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | +0.3585 [+0.0259, +0.7054] | 1.1761 | PASS |
+| cpu s | +0.1323 [-0.0657, +0.4696] | 0.4251 | INCONCLUSIVE |
+| minor faults | +0.0000 [-0.3831, +0.7547] | 0.7634 | PASS |
+| p99 operation ns | +1.0573 [-0.6935, +2.0647] | 0.7289 | INCONCLUSIVE |
+| peak RSS MiB | +0.0052 [-0.0052, +2.6757] | 2.5398 | INCONCLUSIVE |
+| RSS 0.5 s after drain MiB | +0.0052 [-0.0052, +2.6705] | 2.5398 | INCONCLUSIVE |
+
+Release p95: 1030.0 ms; bound: 1300 ms; held: True.
+

.github/assets/perf-614/ledger-v4.md

@@ -0,0 +1,53 @@
+# V4 original stage-one ledger
+
+Baseline `227eef06ff03f92966f38508441d6c2f60a75135`; candidate `8fda62c6d05a68ea4c0ffe2859f07eabf5e72d97`.
+
+STOP: no statistically supported target CPU/throughput gain
+
+Original per-host paired cohorts only; zero memory credit. Intervals and verdicts are retained.
+
+## large-class/8
+
+Run 39845980415; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | +0.1033 [-0.5035, +0.5822] | 0.8696 | PASS |
+| cpu s | -0.1548 [-0.3298, +0.0418] | 0.3071 | PASS |
+| minor faults | -1.0087 [-3.5398, +0.4444] | 3.0435 | PASS |
+| p99 operation ns | -0.0280 [-1.1577, +1.1211] | 1.1280 | PASS |
+| peak RSS MiB | -0.0570 [-2.5930, +0.0158] | 0.1688 | PASS |
+| RSS 0.5 s after drain MiB | +0.1305 [-1.4765, +1.6039] | 1.2146 | INCONCLUSIVE |
+
+Release p95: 1040.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8
+
+Run 39845983742; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.3513 [-0.5710, +0.0413] | 0.3058 | INCONCLUSIVE |
+| cpu s | +0.3114 [+0.1776, +0.4774] | 0.2455 | INCONCLUSIVE |
+| minor faults | -0.1901 [-0.7663, +0.3891] | 0.7782 | PASS |
+| p99 operation ns | -1.8827 [-2.6022, -1.1088] | 1.0971 | PASS |
+| peak RSS MiB | +0.0052 [-0.0472, +0.1258] | 0.0629 | INCONCLUSIVE |
+| RSS 0.5 s after drain MiB | +0.0000 [-0.0455, +0.7329] | 10.8636 | PASS |
+
+Release p95: 1050.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8 short generations (chart build)
+
+Run 39845987944; AMD EPYC 9V45 96-Core Processor; samples per arm: {'base': 95, 'head': 95, 'null': 95}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.2938 [-0.8665, +0.5941] | 0.9289 | PASS |
+| cpu s | +0.2126 [-0.2354, +0.7039] | 0.7425 | PASS |
+| minor faults | -0.3774 [-0.7782, +0.3831] | 1.1236 | PASS |
+| p99 operation ns | +0.0000 [-0.5043, +0.4993] | 1.0188 | PASS |
+| peak RSS MiB | -0.0052 [-2.7441, +0.0054] | 2.6757 | PASS |
+| RSS 0.5 s after drain MiB | -0.0052 [-2.7441, +0.0054] | 2.6757 | PASS |
+
+Release p95: 1090.0 ms; bound: 1300 ms; held: True.
+

.github/assets/perf-614/ledger-v6.md

@@ -0,0 +1,53 @@
+# V6 original stage-one ledger
+
+Baseline `227eef06ff03f92966f38508441d6c2f60a75135`; candidate `a40d8d6e717df31859103a034bf349d32e7760fc`.
+
+Rejected: no statistically supported target CPU/throughput gain. 14 PASS, 4 INCONCLUSIVE, zero FAIL. No stage two or unchanged rerun.
+
+Original per-host paired cohorts only; zero memory credit. All ten platform workflows passed. Platform success does not establish performance acceptance.
+
+## large-class/8
+
+Run 39855326807; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 80, 'head': 80, 'null': 80}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | +0.8402 [-0.0645, +1.4413] | 0.6799 | PASS |
+| cpu s | +0.0272 [-0.2649, +0.3017] | 0.4542 | PASS |
+| minor faults | -0.4357 [-2.6087, +2.2026] | 2.6432 | PASS |
+| p99 operation ns | -0.5771 [-1.1513, +0.5482] | 1.6103 | PASS |
+| peak RSS MiB | +0.0131 [-0.0578, +0.0684] | 0.0893 | PASS |
+| RSS 0.5 s after drain MiB | -0.1289 [-1.9111, +1.5691] | 2.0848 | PASS |
+
+Release p95: 1040.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8
+
+Run 39855330463; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 120, 'head': 120, 'null': 120}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.1958 [-0.4755, +0.1010] | 0.6521 | PASS |
+| cpu s | +0.1477 [+0.0069, +0.3486] | 0.2342 | INCONCLUSIVE |
+| minor faults | +0.3929 [-0.3802, +0.7905] | 0.3968 | INCONCLUSIVE |
+| p99 operation ns | +0.5028 [+0.0000, +1.0616] | 1.0037 | INCONCLUSIVE |
+| peak RSS MiB | +0.0230 [-0.0210, +0.0839] | 0.1020 | PASS |
+| RSS 0.5 s after drain MiB | +0.0000 [-0.7731, +10.9190] | 0.8235 | INCONCLUSIVE |
+
+Release p95: 1050.0 ms; bound: 1300 ms; held: True.
+
+## large-class-ephemeral/8 short generations (chart build) short generations (chart build)
+
+Run 39855334782; AMD EPYC 7763 64-Core Processor; samples per arm: {'base': 110, 'head': 110, 'null': 110}.
+
+| Metric | Change % [95% CI] | Allowance % | Verdict |
+| --- | --- | --- | --- |
+| ops/s | -0.1823 [-0.5360, +0.2481] | 0.5711 | PASS |
+| cpu s | +0.1091 [-0.1198, +0.2890] | 0.2946 | PASS |
+| minor faults | +0.0000 [-0.3861, +0.7605] | 1.1407 | PASS |
+| p99 operation ns | -0.0517 [-0.9932, +0.5873] | 1.0965 | PASS |
+| peak RSS MiB | -2.5124 [-2.6753, +0.0000] | 2.5489 | PASS |
+| RSS 0.5 s after drain MiB | -2.5124 [-2.6753, +0.0000] | 2.5489 | PASS |
+
+Release p95: 1010.0 ms; bound: 1300 ms; held: True.
+

.github/assets/perf-614/platform-gates-archive.json

@@ -0,0 +1,8 @@
+{
+  "file": "platform-gates.zip",
+  "sha256": "a07e1ba71344bd5a9e93337e180ab6a4fdd58becbb47c6107a306449c942408a",
+  "files": 20,
+  "candidate": "287cac2f664a46703ebb6522eef8c8dbeb984cbb",
+  "test_merge": "c9eab6aebbc68a4763b6063f1ebbf6060b6fe39e",
+  "source_tree": "a910b85e21192b946f9d7acda1096c60d437c582"
+}

.github/assets/perf-614/v1-measurements-archive.json

@@ -0,0 +1,15 @@
+{
+  "file": "v1-measurements.zip",
+  "sha256": "e6420a419201fb609dd6d17d90d53c95b4d55f751ed985bd14fa70d3b10b9b5d",
+  "bytes": 10185953,
+  "files": 52,
+  "baseline": "227eef06ff03f92966f38508441d6c2f60a75135",
+  "candidate": "287cac2f664a46703ebb6522eef8c8dbeb984cbb",
+  "disposition": "rejected: no supported target CPU/throughput gain",
+  "stage_two_dispatched": false,
+  "metric_verdicts": {
+    "PASS": 10,
+    "INCONCLUSIVE": 8,
+    "FAIL": 0
+  }
+}

.github/assets/perf-614/v2-campaign-plan.json

@@ -0,0 +1,18 @@
+{
+  "baseline": "227eef06ff03f92966f38508441d6c2f60a75135",
+  "candidate": "59d369609ab47358df0f883b6419252f1ffbf03a",
+  "status": "V2 predeclared; not dispatched",
+  "initial_samples_per_arm": 60,
+  "adaptive_cap_per_arm": 120,
+  "null_arm": true,
+  "sanity_gate": true,
+  "stage_one": [
+    "large-class/8",
+    "large-class-ephemeral/8",
+    "large-class-ephemeral/8 short generations (chart build)"
+  ],
+  "stage_two": "all remaining rows from the 12 default perf-ab workloads plus sparse-large-buffers 64-slot budget/8 (#422)",
+  "rules": "No unchanged reruns; one original cohort per candidate/cell; all six metrics; zero memory credit; no cross-host pooling; reject confirmed per-cell FAIL or service-bound violation. Stage two only if stage one has no FAIL/incomplete/sanity failures and supported target CPU or throughput gain. INCONCLUSIVE retains its label. No new numerical risk waiver: automatic acceptance only under existing rule12 with supported target gain and passing CPU/throughput controls; unresolved material risk needs attribution or owner disposition.",
+  "mechanism": "Skip the restore-function call at retirement publication when both discard records are empty. Original restore function retained for dirty retirement and all other callers.",
+  "previous_candidate_disposition": "V1 rejected: no supported target CPU/throughput gain; original cohorts retained, no stage two or unchanged rerun."
+}