Record qualified cursor service measurements and raw paired samples

doc/benchmarks/2026-11-03-native-event-ingestion.md

@@ -0,0 +1,40 @@
+# Native event ingestion measurements (2026-11-03)
+
+Cursor service acceptance has passed. Application and batching acceptance remain in progress. These measurements use the actual append service and real PostgreSQL on a dedicated quiet host. They establish a local service improvement.
+
+The long-history cursor/range operations are bounded; replay still performs its required indexed event identity lookup. At 100k seeded events, the separately recorded baseline plans examine about 100,000 source entries; the candidate performs a one-row identity lookup and bounded projection work. The raw reports retain the SQL and plans.
+
+## Cursor comparison
+
+Each ordinary sample contains **512 append attempts** (16 groups of 32). Five baseline/candidate pairs alternate order; two warmups, seeding, rebuild preparation and verification are excluded. Replay attempts reuse receipts and create no new events. The table reports median wall time per 512 attempts, not individual event latency.
+
+| Seeded history / case | B0 ms/512 | B1 ms/512 | Change | Pairs beating noise | A/A noise | SQL calls/attempt B0 → B1 |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| 1,000 / ordered | 3821.9 | 3602.4 | -5.74% | 0/5 | 14.603% | 5 → 5 |
+| 1,000 / replay | 3115.1 | 3044.4 | -2.27% | 0/5 | 5.697% | 4 → 4 |
+| 1,000 / warm-suffix | 4172.9 | 3627.0 | -13.08% | 5/5 | 2.429% | 6 → 5 |
+| 10,000 / ordered | 3932.7 | 3456.0 | -12.12% | 5/5 | 3.765% | 5 → 5 |
+| 10,000 / replay | 3438.1 | 2906.0 | -15.48% | 5/5 | 2.651% | 4 → 4 |
+| 10,000 / warm-suffix | 4579.7 | 3588.7 | -21.64% | 5/5 | 2.376% | 6 → 5 |
+| 100,000 / ordered | 7966.9 | 3564.6 | -55.26% | 5/5 | 1.261% | 5 → 5 |
+| 100,000 / replay | 7417.3 | 2972.4 | -59.93% | 5/5 | 1.829% | 4 → 4 |
+| 100,000 / warm-suffix | 8526.7 | 3616.2 | -57.59% | 5/5 | 1.207% | 6 → 5 |
+
+All six required 10k/100k cases exceed their observed per-case A/A envelope in **5/5 individual pairs** and in the median. The 1k cases pass nonregression; 1k ordered and replay gains remain inside noise. Every A/A timing case met the preregistered maximum paired difference ≤20% and median drift ≤10%; the largest median drift was 1.791%. Both full experiments completed correctness checks and cleanup.
+
+Transactions remain one per attempt in this cursor slice. For 100k ordered writes, median backend CPU falls from 10.215 to 1.738 ms/attempt; owner-process CPU falls from 1.958 to 1.824 ms/attempt. The backend clock has 10 ms resolution across each 512-attempt sample. Owner CPU describes the benchmark process, excluding the assembled application, Rust, Chromium and load driver.
+
+Late gap closure is a separately seeded **one-append structural diagnostic**. Its median wall time changes +15.15% at 1k, +10.67% at 10k and −65.15% at 100k. These values disclose a small-history tradeoff; they are excluded from the preregistered latency gate and do not demonstrate a general gap-closure speedup.
+
+## Reproduce and inspect
+
+See [experiment commands](../../scripts/native-ingestion/README.md) and the [implementation and acceptance contract](../plans/2026-11-03-native-event-ingestion.md). Exact frozen revisions, all paired samples, sample min/max spread, CPU, SQL diagnostics, history counts and gate results are in the [machine-readable summary](2026-11-03-native-event-ingestion.json). Full original reports are retained below as deterministic gzip files:
+
+| Original report | SHA256 of decompressed original |
+| --- | --- |
+| [service-aa-b0.json.gz](native-event-ingestion-2026-11-03/service-aa-b0.json.gz) | `92ae98cd5bb22b4bf1b6425150b2615d280d046c3d24ca341358847eede28fe5` |
+| [service-b0-b1.json.gz](native-event-ingestion-2026-11-03/service-b0-b1.json.gz) | `11d7f4c5fcfbeab39d9b38224aa151a3cf6430bb103c4d45b847789e72cb2899` |
+
+Use `gzip -dc <report.json.gz>` to inspect the complete original JSON. Its byte-for-byte SHA256 is recorded above and in the summary. No p99 is inferred from five service samples. The initial noisy calibration and failed browser qualifications remain excluded; the final report will record their failure ledger alongside the application comparisons.
+
+Application streaming, control responsiveness, batching, profile-only backlog/resource diagnostics, final delivery revisions and current-head CI remain pending. The live-provider functional journey has passed separately. This report makes no deployed production or 10,000-agent qualification claim.