Git to Zig

git is one of the most widely deployed systems programs in existence: an object database, a content-addressed storage format, an index, refs, diff and merge machinery, and dozens of porcelain commands, all with two decades of precisely specified behaviour down to exact output strings and exit codes. Reimplementing it is a marathon of format compatibility (SHA-1 object encoding, pack files, the index format) layered under a very large CLI surface.

In this task the agent is handed tests, not source. git's C code is not in the workspace and the real gitbinary is locked root-only, so the behaviour must be reconstructed from git's own behavioural test scripts, which serve as the executable specification.

Implement git's version control functionality in Zig inside the pre-scaffolded project at /app/zig-git/. The build scaffold is fixed: zig build compiles src/ into zig-out/bin/git, and the agent's sources may only be .zig files under src/.

• Match real git's behaviour (its subcommands, output formats, and exit codes) as pinned down by the behavioural test scripts in /app/tests/t/.
• Run the tests with /app/run_tests.py (the full set), /app/run_tests.py --sample (a quick representative slice), or by naming individual scripts.
• The implementation must be Zig itself: delegating to an existing git binary or linking libgit2 zeroes the result.
• All changes must stay inside .zig sources under /app/zig-git/src/; the verifier discards everything else.

A ready-to-build Zig scaffold at /app/zig-git/ (build.zig, build.zig.zon, src/) and an agent-readable copy of git v2.47.0's own behavioural test suite at /app/tests/: the test-lib.sh harness, prebuilt test-helper binaries, and a visible subset of the t/scripts, with git's C source and git's own binaries removed. A provided /app/run_tests.pyrunner drives the scripts against the agent's binary. Zig 0.14.0 is preinstalled, the machine is offline, and a sandbox-timer CLI reports the remaining time budget. The real git is installed but root-only: unreadable and unrunnable by the agent.

A clean-room verifier, separate from the agent's environment, first reconstructs the scored project from a pristine baked scaffold plus only the agent's .zigsources (symlinks and caches excluded), so tampering with the build configuration or stashing binaries in the workspace cannot survive. It then builds the candidate as a non-root user and runs 743 scored scripts from a trusted, root-only copy of git's test suite against the resulting binary.

• Scoring is a weighted mean of per-script completion fractions: each script's TAP passes are capped at the reference implementation's count, a per-script floor subtracts everything the empty scaffold already passes (a no-op submission scores exactly 0), and a contribution cap keeps any single script from outweighing roughly 1/133 of the total.
• TAP parsing is plan-bounded and per-script results are captured by the root harness into a nonce-named root-only directory, so forged "ok" lines cannot inflate the score.
• Build failure, a missing or non-ELF binary, linking libgit2, or a failed workspace reset zeroes the result; per-script time caps scaled from baked reference durations and a global suite deadline make hangs score ~0 instead of erroring.