[ashhart]

zig/tests/portable_names.zig walks zig, docs, tools and .github and fails on any name Windows can't check out. Since #2274, a pull request reuses main's pass for a test whose binary and declared inputs didn't change. A directory listing isn't an input the build knows of, so a pull request that adds such a name would reuse the pass and stay green. Only main's own run, which builds from nothing, would catch it, after landing.

This makes that run always repeat (has_side_effects). It takes milliseconds.

How it was found

The review of #2095 asked for a trace of what its new steps read while they run. I ran zig build test-host from a cold cache under strace on a hosted Linux runner (run 42166246710), with every test binary named after its root file. A check against the build summary confirmed the trace saw every one of the 618 runs (273 binaries).

Of the runs the build can skip:

• Declared already: core/root, server/root, lfm2 and exlr8_oracle_test read tracked files, and zig/build/test_reads.zig declares each one.
• Not declared: portable_names opens 388 tracked directories.

Every other run that reads tracked files always repeats anyway: the goldens, kernel_symbols, the q35 launch arguments, the picture tests and the Python and shell checks. So #2095's new steps need no new declarations.

The trace covered opens and launches on Linux, not existence checks (stat, access) or the Windows and macOS graphs.

Test plan

• CI passes, and zig build test still runs portable_names on each platform