Upgrade llama.cpp from b9555 to b9621

[bernardladenthin]

Summary

• Bumps llama.cpp dependency from b9555 to b9621 across CMakeLists.txt, README.md, and CLAUDE.md
• Documents 15 upstream changes in the breaking-changes history file, all internal to llama.cpp backends or upstream-compiled translation units
• No project source code changes required; all API modifications are either:
  • Internal to GGML backends (CUDA, Vulkan, OpenCL, Metal)
  • Consumed only by upstream-compiled tools (server-context.cpp, speculative.cpp, mtmd-helper.cpp)
  • New features (EAGLE3 speculative decoding, video pipeline) not yet exposed in Java API

Changes

Configuration updates:

• CMakeLists.txt: Updated GIT_TAG from b9555 to b9621
• README.md: Updated badge and link to reflect b9621
• CLAUDE.md: Updated pinned version reference to b9621

Breaking changes documentation:
Added 15 entries to docs/history/llama-cpp-breaking-changes.md covering:

• ggml_gated_delta_net state tensor reshape (Qwen3.5/Qwen3-Next internal kernel)
• New ggml_col2im_1d operation
• common_get_device_memory_data() return type change (consumed in upstream server-context.cpp)
• mtmd_helper_bitmap_init_*() return type changes (verified zero project call sites)
• New video pipeline APIs (mtmd_helper_video_*, --video CLI flag)
• New common_params fields (path_prompts_log_dir, mtmd_batch_max_tokens)
• EAGLE3 speculative-decoding support APIs in llama-ext.h
• llm_graph_result::set_outputs() signature change
• cpp-httplib v0.47.0 bump
• CUDA backend improvements (concat type support, memory tracking)
• Vulkan backend extension support (VK_VALVE_shader_mixed_float_dot_product)
• OpenCL Adreno GPU kernel additions
• CUDA SSM scan race condition fix

Test plan

• CI is green on this branch (no project source changes; configuration-only bump)
• Docs updated (breaking-changes history file populated per upstream changelog)

Related issues / PRs

Follows the llama.cpp upgrade procedure documented in CLAUDE.md.

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes

https://claude.ai/code/session_016jPq9MLePa3eXjxiLLStwi

[claude]

Review Summary: Upgrade llama.cpp from b9555 to b9621 - APPROVED. All three required files correctly updated (CMakeLists.txt, README.md, CLAUDE.md). 15 upstream changes documented comprehensively in breaking-changes history with zero project source impact. All changes verified as internal to backend TUs or upstream-compiled translation units. Analysis quality is excellent with new features noted for future Java API exposure. Low risk, well-documented bump.

[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

[claude]

Review Summary: This PR successfully upgrades llama.cpp from b9555 to b9621.

APPROVED - All three required locations updated consistently: CMakeLists.txt, README.md, and CLAUDE.md. Comprehensive breaking changes documentation added (15 entries) correctly identifies that all API changes are internal to llama.cpp backends or consumed only by upstream-compiled translation units. Zero project source impact as documented. SPDX license headers added appropriately. Follows established CLAUDE.md upgrade procedure exactly. No security concerns. Well-executed, properly documented upstream version bump. Merge as-is.