Fix CUDA-Linux build retry: disable ggml's own GGML_CCACHE auto-detect

[bernardladenthin]

Summary

• Fix CUDA-Linux build: retry without sccache on nvcc device-compile crash #385's regex fix correctly triggered the uncached retry on CUDA 13.3's sccache/nvcc fatbinary failure, but verified live on the very next dispatch (run 31340386884) the retry failed identically
• Root cause: ggml's own ggml/src/CMakeLists.txt self-enables ccache/sccache (GGML_CCACHE, default ON) whenever it finds one on PATH and CMAKE_C_COMPILER_LAUNCHER/CMAKE_CXX_COMPILER_LAUNCHER are unset — exactly the retry's state, since sccache is still on PATH from the failed attempt (build.sh only clears its own $LAUNCH flags, not ggml's independent detection)
• ggml wires itself in via the global RULE_LAUNCH_COMPILE CMake property, which wraps nvcc too — so the "uncached" retry was silently re-enabling the very launcher it was trying to avoid
• Fix: pass -DGGML_CCACHE=OFF on the retry's cmake -Bbuild invocation, guaranteeing a genuinely uncached build regardless of what's left on PATH
• Updated the CLAUDE.md incident note (added on Fix CUDA-Linux build: retry without sccache on nvcc device-compile crash #385) to describe both fixes together

Test plan

• Root cause confirmed live against the actual failing job logs (two consecutive dispatches, same failure signature both times)
• CI is green on this branch (CUDA-Linux job on the retry path)
• Docs (CLAUDE.md) updated

Related issues / PRs

Follow-up to #385, which fixed the retry trigger but not the retry's actual effectiveness

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_01CtT7KLNmC8gEoDS1EiGGwD

---

Generated by Claude Code

[claude]

Review Summary: This PR correctly fixes a subtle build system issue where ggml's independent GGML_CCACHE auto-detection was re-enabling the problematic sccache launcher on the retry path. The fix of passing -DGGML_CCACHE=OFF is well-targeted and properly documented. The solution is defensive, minimal, and verified live on the actual failing job. Ready to merge.

[claude]

The fix in build.sh is well-implemented. Key observations:

1. Line 161 correctly passes -DGGML_CCACHE=OFF to disable ggml's independent cache detection during the retry
2. The comment (lines 153-160) is excellent - it clearly explains the root cause and why this flag is necessary
3. The flag is placed at the right point: after the build directory is cleaned but before cmake configures the project
4. This prevents ggml from auto-discovering sccache on PATH and wiring it back in via RULE_LAUNCH_COMPILE

The CLAUDE.md documentation update is comprehensive and explains both the immediate fix (regex for 'Compiler killed by signal') and this deeper fix (disabling GGML_CCACHE). The narrative clearly shows why two separate fixes were needed.

[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