llama.cpp b11534, executed-test floor, loader output to stderr, JBang example (PR A)

[bernardladenthin]

Summary

• A1 — llama.cpp b11529 → b11534 in 3 reviewed steps (b11530, b11531, b11534), one commit each. The only step with project code is b11531, #30210 "chat : refactor API": the prompt and parser state of a chat generation now live in a common_chat_session that oaicompat_chat_params_parse fills and the server task applies; jllama.cpp threads a session through its four chat entry points and the C++ tests follow the new signatures (603 tests, unchanged count). None of the removed request fields was a RequestField, so the Java wire surface is unchanged. All eleven patches apply unchanged on every tag; every drop-check still finds its defect. Also in the range: OpenCL kernels that compile on Adreno A6x (#30176), exact GELU for ModernBERT (#30108), no redundant CUDA copies after SSM_SCAN (#29807). Per-step record in docs/history/llama-cpp-breaking-changes.md.
• A2 — floor on executed tests. verify-test-counts.sh gains --min-executed (run minus skipped), the shape the run count cannot see (method-level assumptions skipping in bulk). Measured on the last green run: 1869 run everywhere, skipped 4 (Linux) / 11 (macOS ×3) / 13 (Windows Ninja + MSVC), so 1856–1865 executed; a checkout without the models executes 1589. Both workflows pass --min-executed 1800 instead of --min-total 1500; the model-less local run goes red as intended. TODO entry closed.
• A3 — LlamaLoader prints its two diagnostic lines to stderr. A router worker JVM printed [jllama] using native backend '…' onto the router's command pipe (unexpected output on the command pipe since b11401). The three smokes already read both streams, so no script change; TODO entry closed.
• A4 — examples/jbang/Chat.java, a one-file console chat for jbang <url> model.gguf, plus a README section. JBang treats a pom dependency as a BOM (measured), so the file names the classes jar and the 7 desktop CPU natives jars itself; check-natives.py holds those lines to the platform=yes rows and the version to the README's install snippet (unit-tested), and the version-bump list in CLAUDE.md names the file.
• docs/handover/local-agent-pr-a.md — the prompt for the local agent: what the branch contains, what the sandbox could not verify (models, Windows, GPU), the deliverables in order (Windows suite with models, Windows test counts and one red run, router log, JBang against the local snapshot, the B2 pre-verification for the Windows CPU variants) and the branch protocol. To be deleted once this PR is merged.

Test plan

• Patches: git apply in order on pristine b11532 and b11534 (eleven of eleven), fresh configure through FetchContent at b11531 and b11534; check-patches.py: 11 patches, 98 hunks, 0 problems
• ctest at b11531 and at b11534: 603/603 passed (Linux x86-64)
• Java at b11534: mvn -f llama/pom.xml test: 1869 run, 0 failures, 280 skipped (model-gated tests self-skip in the sandbox); verify-test-counts.sh --min-executed 1800 correctly rejects that run (1589 executed) and accepts synthetic reports with the floor met
• spotless:check, compile spotbugs:check, javadoc:jar green
• check-natives.py (0 disagreements), check-release-gate.py, check-run-scripts.py, check-shared-files.py (0 changed here alone), buildcheck unit tests (113), reuse lint compliant
• examples/jbang/Chat.java compiles against the module classes with javac --release 8; JBang's group:artifact:version:classifier form verified to resolve a classifier jar
• CI is green on this branch — the Windows jobs' test counts verified line is the first real reading of the executed floor on all six platforms
• Docs / CHANGELOG updated where applicable

Related issues / PRs

• Previous bump: Upgrade llama.cpp from b11512 to b11529; nvcc through sccache opt-in; docs cleanup #494 (b11512 → b11529). Upstream: chat : refactor API ggml-org/llama.cpp#30210, #30176, #30108, #29807.

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes (if there are, I have notified the maintainer privately per SECURITY.md)

🤖 Generated with Claude Code

https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate failed

Failed conditions
D Maintainability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE