docs: modular s390x jar verified on cfarm191, Android AAR run read

CLAUDE.md

@@ -2595,7 +2595,10 @@ Q4_K_M converted to big-endian, `llama-bench -t 4 -p 64 -n 16`, two interleaved
 generation) -- 4.5x and 4x over the scalar library every jar up to 6.0.x shipped. Correctness, by
 `llama-perplexity --kl-divergence` against a Q8_0 reference computed on `z13`: the Q4_K_M model
 lands at a mean KLD of 0.0777 on `z13` and 0.0746 on `z16` -- equally far from it; the 0.0087
-between the two modules is about a tenth of what the quantization costs. **Under qemu** (the
+between the two modules is about a tenth of what the quantization costs. The CI-built jar (run
+40475405391) loads `libggml-cpu-z16.so` there and serves the same model through `ServerLauncher` at
+29.9-35.6 t/s generation with 4 threads, against 9.9-10.0 for the static 6.1.0-SNAPSHOT jar before
+it (TODO.md, "Linux s390x"). **Under qemu** (the
 `c_abi` run of `build-linux-s390x`, the RHEL 8 row of `smoke-glibc-floor`) ggml loads `z15`:
 qemu-user's CPU model has VXE2 and no NNPA, so CI runs the vector code too, in emulation (measured
 locally with the job's toolchain: `c_abi` reports `CPU : VXE = 1`, and the library loads in the UBI 8

TODO.md

@@ -554,12 +554,20 @@ loads on RHEL 8.10 s390x (UBI 8 OpenJDK 21 runtime under qemu) where the 6.0.0 j
       upstream's list lacks. On cfarm191 (b11901, Qwen3-0.6B Q4_K_M big-endian, 4 threads): `z13`
       19.7 / 13.6 t/s, `z15` 90.0-95.5 / 57.3-58.3, `z16` 82.6-88.3 / 51.9-53.9 (prompt / generation).
       The C++ suite moved to `test-cpp-linux-s390x` (a modular build cannot link `jllama_test`).
-- [ ] **The CI-built modular s390x jar on cfarm191**, once a Publish run of the branch is green:
-      `NativeLoadSmoke` and the `ServerLauncher` round on the big-endian Qwen3-0.6B as on 2026-10-26,
-      plus which module the loader line and `JNI_OnLoad`'s registry log name (`z16` expected), and the
-      speed against the 6.1.0-SNAPSHOT static jar of run 40400361809 (0.33 t/s generation at 8 threads
-      on the shared host, 13.6 t/s for the scalar build at 4). Only a real IBM Z runs the vector modules
-      at their own speed; qemu scores them against its CPU model.
+- [x] **The CI-built modular s390x jar on cfarm191** (2026-10-27, the `llama-glibc-floor` set of the
+      green PR run 40475405391 of #580, JDK 25 on Debian 13, z16):
+      - `NativeLoadSmoke`: linked b11901-5320c87ff, `compiler=Clang 18.1.3`, patches 0001-0023; all 9
+        files extracted, and `JNI_OnLoad` logs `loaded CPU backend from .../libggml-cpu-z16.so`,
+        `2 backend(s) registered ... [RPC, CPU]`.
+      - `ServerLauncher` on the big-endian Qwen3-0.6B Q4_K_M answers "Paris." (fingerprint
+        b11901-5320c87ff), the same text as the static jar of run 40400361809 at temperature 0.
+      - **Speed against that static jar**, three alternating rounds on a quiet host (load < 1 before,
+        26 prompt tokens, 64 generated, t/s): at 4 threads prompt 50.3-61.8 against 14.4-16.0 and
+        generation 29.9-35.6 against 9.9-10.0 (3.0-3.6x); at 8 threads prompt 35.3-81.4 against
+        13.8-25.2 and generation 24.1-49.5 against 14.8-15.4, noisy on the shared host. The 0.33 t/s of
+        the earlier check was the host's load, not the jar.
+      Only a real IBM Z runs the vector modules at their own speed; qemu scores them against its CPU
+      model.
 - [ ] **z17 (`-march=arch15`)** is skipped: clang 18 rejects the flag (clang 19 on cfarm191 too), and
       ggml's configure drops the variant silently. Re-check when the toolchain's clang moves, and
       measure on a z17 if the farm gets one.
@@ -677,7 +685,8 @@ measurements and is not evidence that GPU testing is currently blocked.
         into 133 of the core's 146 class files (RUNTIME retention). So dropping the dependency needs
         `-dontwarn org.checkerframework.**` in `consumer-proguard.txt`, or a minified app fails R8 on
         "Missing class"; added, like lombok and animal-sniffer.
-  - [ ] **Read the first Publish run:** `package-android-aar`'s new step, the consumer fixture's R8
-        `assembleRelease` (cpu and opencl) now that checker-qual is off its classpath, and the
-        emulator instrumentation test (Jackson reflecting over the annotated classes without
-        checker-qual -- ART is expected to skip the unresolvable annotation types, as the JVM does).
+  - [x] **Read the Publish run** (40475405391, #580, 2026-10-27): `verify-aar-poms.py` says both poms
+        "mirror the core" (jackson-databind, jspecify, slf4j-api); the consumer fixture's
+        `minifyReleaseWithR8` + `assembleRelease` pass for cpu and opencl without checker-qual; the
+        emulator job's `connectedDebugAndroidTest` passes (Jackson over the annotated classes, ART
+        skips the unresolvable annotation types as expected).