# Java JNI cursor: current-main integration — 19 October 2026

Merged **206 upstream commits** through `367f7dfaf` into the published cursor branch, without rewriting it. Integration commit `654619556d563d21f0bd148b1a056c97e19fb59a` includes the original optimization. The native wrapper merged automatically; only the generated remaining-root ledger conflicted and was regenerated from tracked paths. The cursor helper and focused Java/native tests remain byte-identical to implementation commit `8b7128c5bc2170ae2d93cdf86a9f33b732141665`. The upstream change in the JNI file is an unrelated configured-acceptor comment; the four original optimization hunks remain intact.

A fresh normal SDK/JNI build passed, including the live loaded-library mapping smoke. ABI generation 4/signature `27f689e02a714d6f022abea6505fbc72cce5b554a047bbf7fa19feece9aafac9` remains unchanged. Java source uses release 8, native code C++11, actual JVM JDK 21.0.12.1. Built classes SHA256 `c367cecdd50b4cbe7acd17f6bdc3a2536fb4a3eec55b1cec07abffaccb13fc59`; JNI DSO `5357bf1f615dcccd4d75cfeabc20d482e337885c652c1e751447fda82bd647fa`. Source/tool seals were unchanged during the build, and SDK artifact hashes still match after tests/admission/smoke.

Validation passed **51 cursor/codec JUnit tests under `-Xcheck:jni`**, export scans, **24 normal admission cells and 48 negative probes**, then a separate **24-cell smoke run** (two engine processes, 10,000 batches of 32 operations per cell, 200,000 warmup, CPU4). The current controller freshly compiled both normal adapters and every negative probe. SDK building/test compilation and admission completed before smoke timing. [Qualification receipt](../jni-cursor-integration-20261019/QUALIFICATION.json) · [JNI test output](../jni-cursor-integration-20261019/validation/cursor-codec-junit-xcheck.log) · [Smoke CSV](../jni-cursor-integration-20261019/smoke/results/stage2_summary.csv) · [Negative audit](../jni-cursor-integration-20261019/admission/results/stage2_parity_audit.csv).

The smoke results below are single-process checks of the integrated build, **not a fresh paired performance comparison**. No confidence interval, steady-state qualification or original-gain reproduction is claimed. The original 24.57% six-cursor-cell improvement and uncertain index-tail result remain pinned to their original source/runtime snapshots; neither the original raw evidence nor its reports were rewritten. [Original controlled comparison](RECORD_2026-10-19_java-jni-cursor.md).

| Operation | Fixture | Pure smoke p50 ns | JNI smoke p50 ns |
|---|---|---:|---:|
| cursor_parse_only | D | 178 | 762 |
| cursor_parse_only | 8 | 252 | 823 |
| cursor_parse_only | W | 678 | 2655 |
| cursor_parse_extract | D | 338 | 884 |
| cursor_parse_extract | 8 | 464 | 1087 |
| cursor_parse_extract | W | 1511 | 3448 |
| index_parse_only | D | 229 | 264 |
| index_parse_only | 8 | 296 | 332 |
| index_parse_only | W | 972 | 981 |
| index_parse_extract | D | 384 | 395 |
| index_parse_extract | 8 | 496 | 510 |
| index_parse_extract | W | 1528 | 1601 |

The compiled integration SDK, all preparation/admission/smoke logs, source/tool inventories and test evidence are retained in the archive identified by [ARCHIVE.json](../jni-cursor-integration-20261019/ARCHIVE.json). Host/JDK/compiler dependencies are recorded rather than distributed; this archive is not an independent paired benchmark snapshot or offline build kit. The original frozen benchmark archive remains separate. The prepared [pull-request description](../jni-cursor-integration-20261019/PULL_REQUEST.md) keeps the measured improvement distinct from this integration qualification. GitHub REST/GraphQL API access returned Forbidden during integration; branch publication and the direct comparison URL remain the review path if PR creation remains blocked.