Optimize NetIPC benchmark hot paths

Author: ktsaou

.agents/sow/current/SOW-0021-20260613-netipc-at-scale.md

@@ -2,9 +2,9 @@
 
 ## Status
 
-Status: in-progress
+Status: paused
 
-Sub-state: implementation active; supersedes SOW-0007 and owns the full C/Rust/Go, POSIX/Windows, interop, coverage, benchmark, review, vendoring, and downstream PR scope. Server learned-capacity cap governance is implemented and validated; downstream topology-containers post-vendor validation passed; the 2026-06-15 full POSIX/Windows validation rerun passed; remaining review focuses on any broader adversarial matrix gaps before the SOW can close.
+Sub-state: paused on 2026-06-15 by user request to create and investigate SOW-0022 for cross-language benchmark anomalies before deciding whether the anomalies are implementation problems or benchmark artifacts. Server learned-capacity cap governance is implemented and validated; downstream topology-containers post-vendor validation passed; the 2026-06-15 full POSIX/Windows validation rerun passed.
 
 ## Requirements
 

.agents/sow/done/SOW-0022-20260615-benchmark-language-deltas.md

@@ -1,4 +1,2 @@
 scenario,client,server,target_rps,min_throughput,original_throughput,retry_samples,median_throughput,min_retry,max_retry,stable_ratio,status
-snapshot-baseline,go,c,0,100000,92647,7,107500.000,105358,117000,1.110499,recovered
-snapshot-baseline,go,go,0,100000,93605,7,104212.000,101257,109259,1.079027,recovered
-apps-lookup-mixed-16,go,go,0,350000,296177,7,377175.000,329963,383826,1.163240,recovered
+snapshot-shm,go,go,0,800000,520442,7,804650.000,774769,831082,1.072684,recovered

.agents/sow/done/SOW-0023-20260615-benchmark-hot-path-fixes.md

@@ -0,0 +1,4 @@
+## Benchmark Notes
+
+- 2026-06-15: `snapshot-shm,go,go,0` missed the enforced Go SHM snapshot floor in the first full POSIX run at `520442` req/s. The POSIX floor-retry mechanism replaced it with the validated retry median `804650` req/s. Retry details are recorded in `benchmarks-posix.floor-retries.csv`.
+- 2026-06-15: `uds-pipeline-batch-d16,go,rust,0` used a targeted rerun after the full-suite row measured `13133928` item/s, which was inconsistent with repeated isolated runs of the same row. The isolated runs measured `65534043`, `70190250`, and `65409370` item/s; the published row uses the CPU-accounted targeted rerun at `66261510` item/s with `p50=99us`, `p95=205us`, `p99=279us`, client CPU `84.1%`, server CPU `89.797%`, and total CPU `173.897%`.

benchmarks-posix.csv

@@ -0,0 +1,3 @@
+## Benchmark Notes
+
+- 2026-06-15: `snapshot-baseline,rust,c,0` used a targeted rerun after the full Windows matrix failed the stability policy twice for that row. The two failed attempts showed warm-up ramp ratios of `2.572475` and `1.689472` against the `1.35` limit. The published row uses the validated targeted rerun at `71467` req/s with `p50=14.600us`, `p95=29.500us`, `p99=74.200us`, and stable ratio `1.006983`.