[datlechin]

Summary

• RPM builds compile DuckDB with full native debug information, while Rust already uses reduced debug information. DuckDB objects and their static archive coexist and can exhaust the runner during ar, as seen in feat(core-ssh): tunnel through SSH in process on Windows #4008 and feat(core-mssql): add TDS 8.0 strict encryption and test integrated login and TLS trust #4012.
• Set Fedora's package C/C++ debug flags to -g1. Keep function and line information, hardening, and the existing native LTO opt-out.
• Sample filesystem use every second throughout RPM packaging. Retain peak use and minimum free space with successful and failed build artifacts.

Test plan

• Local sampler regression, ShellCheck, workflow lint, Fedora macro assertions and independent security review passed. A two-source GCC probe preserved exported symbols and DWARF lines while reducing archive size from 6,339,736 to 140,176 bytes.
• Clean main baseline 3b049a927: full RPM build, all 9 Meson suites, debug packages and strict rpmlint passed. Observed peak: 66.69 GiB (71,609,892,864 bytes).
• Baseline measured allocated blocks in owned build directories and input archives, with inode deduplication, in an ARM64 Fedora 44 container on the Mac. GCC 16.2.1, Rust 1.99.1, two jobs, cold release outputs. Installed tools were excluded. There were 4,773 samples, median interval 1.454 seconds; a safety pause produced a 236.579-second gap. This is an observed peak, not an exact bound.
• Fixed-build measurement uses this PR's normal x64 RPM CI job, as directed by the maintainer. Runner peak use and minimum free space are pending. Its df samples include runner/toolchain overhead and are not directly comparable to the local owned-path baseline.