Record recent benchmark results and add AGENTS.md

AGENTS.md

@@ -0,0 +1,48 @@
+# AGENTS.md
+
+Guidance for coding agents working in this repository. The design itself (write and read
+paths, buffer ownership, map keys, the parser's accessor contract) is in `docs/DESIGN.md`;
+read it before changing the generator, parser, writer or reader.
+
+## Project
+
+A Jackson 3 dataformat for MessagePack, `org.komamitsu:jackson-dataformat-msgpack`, with its
+own MessagePack encoder and decoder. msgpack-core is a test dependency only, used as the
+reference implementation in tests.
+
+## Commands
+
+```bash
+./gradlew clean build                  # Compile, checkstyle, all tests (Java 17 toolchain)
+./gradlew build -PtestJavaVersion=21   # Run the tests on another JDK, as CI does for 17, 21 and 24
+./gradlew soakTest -Psoak.seconds=120  # Memory-growth soak tests, excluded from the normal build
+./gradlew :jmh:jmhJar                  # Build the benchmark jar
+```
+
+Tests run with `-Dfile.encoding=windows-1252`, so a conversion that forgets to name a charset
+fails in the tests.
+
+## Rules
+
+- **Round-trip safety comes first.** Whatever this library writes, it must be able to read
+  back. A type that cannot round-trip is refused on write rather than written in a form that
+  fails on read (see `UnreadableKeyGuard` and DESIGN.md 2.6).
+- **Compatibility with the Jackson 2 module (`msgpack-jackson` in msgpack-java) is not a
+  goal.** Do not keep or add behaviour only because that module had it.
+- **Output after a serialization failure is the caller's to discard.** Do not add rollback or
+  cleanup machinery for callers that catch the exception and keep using the generator.
+- **Bug fixes come with a test that fails without the fix.** Assert exact values and bytes,
+  not only types or non-null.
+- **Benchmark changes to a hot path before pushing.** Build the change and `main` and run them
+  back to back:
+  `java -jar jmh/build/libs/jmh-jmh.jar <Benchmark> -f 5 -wi 5 -w 1 -i 10 -r 2`, adding
+  `--add-opens=java.base/java.nio=ALL-UNNAMED --add-opens=java.base/sun.nio.ch=ALL-UNNAMED`
+  through `-jvmArgs`. Measure with a reused mapper; a benchmark that creates a mapper per
+  operation measures nothing users see. Numbers from different days are not comparable on
+  this machine, and a short run is not enough to call a change free.
+- **Record meaningful results in `jmh/results/`**, one file per change, with the command, the
+  compared commits and the JSON control.
+- **Reviews** go through GitHub Copilot, requested with
+  `gh api -X POST repos/komamitsu/jackson-dataformat-msgpack/pulls/<n>/requested_reviewers -f "reviewers[]=copilot-pull-request-reviewer[bot]"`.
+  Check each finding against the code, with a test where it makes a claim about behaviour,
+  before fixing or declining it.

jmh/results/2026-09-22-write-wrap-helper.md

@@ -0,0 +1,26 @@
+# One wrap helper for keys and container headers
+
+Measured on the same machine as the earlier files, run back to back:
+
+- Before: `c2d1248` (Write scalar values through one verify-and-wrap helper)
+- After: `1af13b5`. Two commits on top of the before state: `f8695a4` (the change below) and
+  a read-side change that builds repeated parser errors in one place, which no write benchmark
+  runs
+
+The JMH command of this run was not recorded, so it is not given here.
+
+## Change
+
+`f8695a4`: `writeName`, `openContainer` and `closeContainer` each wrapped their writer call in
+their own try/catch that turned an `IOException` into a Jackson exception. The try/catch now
+lives in one `pack()` helper, which makes those three hot methods small enough to inline
+better.
+
+## Result
+
+| Benchmark | Before | After | Change |
+|---|---|---|---|
+| writePojoJson | 1099800 ± 14428 | 1088509 ± 13948 | control |
+| writePojoMsgpack | 1170317 ± 9411 | **1273380 ± 12473** | **+8.8%** |
+
+The error bars do not overlap and the control is flat.

jmh/results/2026-09-23-integer-without-boxing.md

@@ -0,0 +1,26 @@
+# Reading an integer without boxing it
+
+Same machine as the earlier files, with `-prof gc` for the allocation figures, run back to
+back:
+
+- Before: `d2d774e` (Name the context methods for what they do)
+- After: `1037af9` below
+
+The rest of the JMH command of this run was not recorded, so it is not given here.
+
+## Change
+
+`1037af9`: the parser boxed every integer it read into an `Integer` or `Long`. It now keeps the
+value in a primitive field.
+
+## Result
+
+| Benchmark | Before | After |
+|---|---|---|
+| readPojoMsgpack allocation | 2056 B/op | **1912 B/op** |
+| readPojoJson allocation (control) | 2096 B/op | 2096 B/op |
+
+144 bytes per operation are gone. MessagePack read now allocates about 9% less than Jackson's JSON reader on the same POJO.
+
+Throughput did not resolve on this machine (694079 ± 11508 before, 680121 ± 7908 after, with
+the JSON control moving from 674k to 677k), so the allocation figure is the result here.

jmh/results/2026-09-27-pr9-vs-main.md

@@ -0,0 +1,35 @@
+# PR #9 against main after the map-key and UUID changes
+
+A check that the changes merged after the Jackson 3 rewrite (#9) did not cost performance:
+#10 (string tokens read as binary), #16 (MessagePackKeySerializer removed, key guard added) and
+#17 (UUID written as a string).
+
+Same machine as the earlier files. Both commits were built and run back to back with
+`java -jar jmh/build/libs/jmh-jmh.jar "MsgpackWriteBenchmark|MsgpackReadBenchmark|WriteUTF8StringBenchmark|WriteStringBenchmark" -f 5 -wi 5 -w 1 -i 10 -r 2 -bm thrpt -tu s`
+(50 samples per benchmark), PR #9 (`bb4158f`) first, then main (`cff8bfc`).
+
+The numbers are higher than the ones #9 recorded in `2026-09-21-single-pass-short-strings.md`
+for the same code (writePojoMsgpack 1242839 here, 1092242 there), so only runs made back to
+back are compared.
+
+## Result
+
+| Benchmark | PR #9 | main | Change |
+|---|---|---|---|
+| readPojoJson | 686406 ± 11655 | 682182 ± 10299 | control |
+| readPojoMsgpack | 686441 ± 7242 | 689769 ± 5651 | +0.5%, within error |
+| writePojoJson | 1091002 ± 5201 | 1072677 ± 11429 | control, -1.7% |
+| writePojoMsgpack | 1242839 ± 9565 | 1237674 ± 8973 | -0.4%, within error |
+| WriteStringBenchmark.shortAscii | 128397 ± 2511 | 128497 ± 2274 | within error |
+| WriteStringBenchmark.shortNonAscii | 85536 ± 3778 | 89744 ± 2307 | within error |
+| WriteStringBenchmark.shortNonAsciiShifted | 48127 ± 1499 | 46294 ± 1721 | within error |
+| WriteStringBenchmark.longAscii | 81484 ± 1344 | 80275 ± 265 | within error |
+| WriteStringBenchmark.longNonAscii | 55262 ± 518 | 55197 ± 187 | within error |
+| WriteUTF8StringBenchmark.writeUTF8StringAscii | 118398 ± 1288 | 117607 ± 667 | within error |
+| WriteUTF8StringBenchmark.writeUTF8StringNonAscii | 115120 ± 1381 | 112199 ± 1320 | -2.5% |
+
+POJO read and write are on par with #9. The only MessagePack row whose error bars do not
+overlap, writeUTF8StringNonAscii at -2.5%, moved while the JSON write control, whose code is
+the same in both builds, dropped 1.7%, so the machine was slower during the second run. That
+row is inconclusive rather than a confirmed regression; running the string benchmarks again
+with main first would tell drift from a real cost.