Root cause found on release/v0.0.258 (fixed at the tier-1 layer in 87c8348): git exports GIT_DIR/GIT_PREFIX to hooks, and the fixture tests in src/agent_eval_tests.zig inherit them, so their 'git init/add/commit' in a tmpdir operated on the enclosing repo — re-marking the checkout bare (the recurring core.bare=true corruption), committing stray 'fixture' commits by 't t@example.com' onto the current branch (tmpdir treated as the work tree), and producing the hook-only 1-failed flake. eval-tier1.sh now unsets the family, but fixtureCmd (and the eval harness's own git spawns) should also pass a scrubbed child env so a direct 'GIT_DIR=… zig build test' can never escape either.
Root cause found on release/v0.0.258 (fixed at the tier-1 layer in 87c8348): git exports GIT_DIR/GIT_PREFIX to hooks, and the fixture tests in src/agent_eval_tests.zig inherit them, so their 'git init/add/commit' in a tmpdir operated on the enclosing repo — re-marking the checkout bare (the recurring core.bare=true corruption), committing stray 'fixture' commits by 't t@example.com' onto the current branch (tmpdir treated as the work tree), and producing the hook-only 1-failed flake. eval-tier1.sh now unsets the family, but fixtureCmd (and the eval harness's own git spawns) should also pass a scrubbed child env so a direct 'GIT_DIR=… zig build test' can never escape either.