Upgrade llama.cpp from b11512 to b11529; nvcc through sccache opt-in; docs cleanup

[bernardladenthin]

Summary

• Upgrades the pinned llama.cpp from b11512 to b11529 in 5 reviewed steps, each ending at a tag (b11516, b11517, b11518, b11521, b11529; chunked by diff size per docs/upgrade/llama-cpp-version-bump.md, with the one 530 KiB step b11518→b11521 taken whole because no tag sits between them). All eleven patches apply unchanged, verified in order against each pristine tag with git apply and, for b11529, through the real FetchContent configure; every drop-check still finds its defect (no kolibri upstream, no split_sum guard, no common_log sink, no stop_server, the four _mm_prefetch calls still uncast, …). Two history rows per step in docs/history/llama-cpp-breaking-changes.md.
• One behaviour change from upstream, handled: #30159 moves the server's default port from 8080 to 9931. NativeServer forwards its argv to llama_server, so an argv without --port now binds 9931 in classic and attach mode; NativeServer.DEFAULT_PORT (what getPort() reports in that case) follows, with NativeServerSmokeTest and a CHANGELOG entry. OpenAiCompatServer keeps its own DEFAULT_PORT = 8080 on purpose: the Java transport has no upstream coupling and the constant is a public API of this project. Every README example passes --port explicitly.
• Notable in the range (no project source change needed): CUDA top-k selection reworked (#28713) and MMQ config helpers take the src1 precision (#30168); SYCL Q5_K reorder-layout MMVQ + fused GLU (#29375); Metal flash-attention kernels for the 128/96 head-size pair (#30209); Vulkan rms_norm workgroup-count overflow fix (#30145); ggml_acc out-of-bounds write with a large offset fixed (#30135); meta backend handles host views under --split-mode tensor (#30217); the WebUI gains a models manager (#29583, auto-followed by build-webui; package.json adds vaul-svelte, scripts/ui-assets.cmake unchanged). Request fields, bounds and response keys of tools/server are untouched in the range (server-schema.cpp / server-task.cpp did not change).
• build.sh: nvcc through sccache is opt-in now (SCCACHE_WRAP_NVCC=true), off by default. Every Linux CUDA job log examined since the CUDA 13.4 bump (runs 37643964068, 37680715063, 37831180985, 37925157175) shows sccache 0.18.0 failing on the first .cu TUs ten seconds into the build (Missing "cubin" file output, Compiler killed by signal 126) and the retry rebuilding everything without any launcher, gcc objects included, for 47–80 min. The nvcc cache never hit on 13.4, and the retry threw the gcc cache away with it. By default sccache now wraps the C/C++ compilers only; ggml adds no launcher of its own in that state (its GGML_CCACHE block fires only when neither C nor CXX launcher is set). Expected effect on the CUDA job: gcc objects served from the cache, nvcc at the cold cost (~50 min) with no wasted first attempt — this PR's run is the first measurement. CLAUDE.md, TODO.md and the CHANGELOG record it and how to re-test the opt-in with a newer sccache.
• Docs cleanup: CLAUDE.md's 0017 row no longer names the deleted .github/workflows/probe-msvc-prefetch.yml.

Test plan

• Patches: git apply in order on pristine b11516, b11517, b11518, b11521, b11529 (eleven of eleven each); check-patches.py: 11 patches, 98 hunks, 0 problems
• Fresh configure at b11529 through FetchContent + full build; ctest: 603/603 passed
• Java: mvn test on the llama module: 1869 run, 0 failures, 280 skipped (the model-gated classes self-skip here); NativeServerSmokeTest with the new default
• javadoc:jar, compile spotbugs:check, spotless:check
• check-natives.py (0 disagreements), check-release-gate.py, check-run-scripts.py, check-shared-files.py (0 changed here alone), buildcheck unit tests (112), bash -n .github/build.sh
• CI is green on this branch — the CUDA job's duration and its sccache stats table are the measurement for the build.sh change
• Docs / CHANGELOG updated where applicable

Related issues / PRs

• Previous bump: Upgrade llama.cpp from b11476 to b11512 (7 steps); ModelParameters.setMoeCacheMib #486 (b11476 → b11512). Follow-up to patches: fix the x86 prefetch address and the type clang rejects on MSVC #487 / URGENT: fix the hunk line count that truncated prefetch.h — main is red in 23 jobs #492 (0017, the hunk guard in buildcheck: check every patch hunk's header against its body; share the java-tdd-guide pointer file (4-repo sync) #493).
• Sibling repo: tdd skill: point at the workspace Java TDD skill and the real test stack BitcoinAddressFinder#383 (the tdd skill pointer fix found in the same round).

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 passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud