======================================================================== LOCK-FREE FLIP — FROZEN ANALYSIS (ledger §8) scipy available: True ======================================================================== === Throughput inference [variant A — disjoint (PRIMARY)] (write_ops_per_sec) === CONTROL : n=30 mean=4,560.6 sd=136.3 TREATMENT: n=30 mean=18,791.9 sd=792.4 diff (T-C) = 14,231.3 ops/s (+312.05%) 95% Welch CI [13,931.8, 14,530.8] Welch t = 96.9467 df = 30.72 p(two-sided) = 9.0972e-40 Cohen's d = 25.0315 (pooled) Mann-Whitney U = 900.0 p(two-sided) = 3.0199e-11 === S1 — checkpoint pause (veto: TREATMENT ≤ CONTROL) === writer-observed p999 (us): control=126373.7 treatment=4532.8 checkpointer-side mean (us): control=67624.5 treatment=59532.1 checkpointer-side p99 (us): control=132144.3 treatment=185414.4 === S2 — write tail p50/p99/p999 (veto: T ≤ 1.10×C on p99/p999) === write_p50_us: control=58.2 treatment=305.8 ratio(T/C)=5.250 write_p99_us: control=67405.7 treatment=1954.6 ratio(T/C)=0.029 write_p999_us: control=126373.7 treatment=4532.8 ratio(T/C)=0.036 === S3 — peak RSS (single-arm-per-process pass; veto: T ≤ 1.25×C) === peak RSS (KiB): control=484,160 treatment=425,088 ratio(T/C)=0.878 === Throughput inference [variant B — contended (S4)] (write_ops_per_sec) === CONTROL : n=30 mean=5,799.6 sd=191.3 TREATMENT: n=30 mean=15,486.1 sd=835.0 diff (T-C) = 9,686.5 ops/s (+167.02%) 95% Welch CI [9,368.0, 10,005.1] Welch t = 61.9379 df = 32.04 p(two-sided) = 6.3159e-35 Cohen's d = 15.9923 (pooled) Mann-Whitney U = 900.0 p(two-sided) = 3.0199e-11 S4 cas_retries (treatment, per round): mean=140.1 max=258 === §8 DECISION RULE === -> BRANCH 4: DON'T FLIP (regression). S1 checkpoint pause worse (treatment p99 > control) ======================================================================== VERDICT: branch 4 = DON'T-FLIP-regression EVIDENCE: S1 checkpoint pause worse (treatment p99 > control) ========================================================================