EXIT_USAGE is declared as 64, and a bad flag exits 2 #30

EXIT_USAGE = 64 is declared at gerberdiff/cli.py:25 and used at four call sites, and a usage error still exits 2.

Reproduced

On main, 2026-09-11:

$ uv run gerberdiff --bogus-flag; echo $?
2

partspec, which does the remap, exits 64. netspec has the identical defect and is filed at heibench/netspec#37.

The mechanism

argparse calls sys.exit(2) internally on an unrecognised flag. Unless parse_args is wrapped, that propagates out before any of this repo's EXIT_USAGE returns are reached — those cover gerberdiff's own validation, not argparse's.

Why it matters here specifically

2 is not a spare number in this repo. 0685cde made it indeterminate — the third diff outcome, "part of the comparison could not be made", the whole point of adjudication A3. The AGENTS.md added in #29 documents it as such.

So today a mistyped flag and a board that could not be fully compared return the same code, and a caller cannot tell them apart. That is the distinction A3 existed to create, reachable by typo.

The shape of the fix

slicelab/cli.py:100-106 does it, citing this repo and netspec as the reason:

try:
    parser.parse_args(argv)
except SystemExit as exc:
    # --help and --version raise SystemExit(0); that is a successful run.
    code = exc.code
    if code in (0, None):
        return 0
    return EXIT_USAGE

The --help carve-out matters: SystemExit(0) is a successful run and must not become 64.

And the test should measure a real process rather than assert the constant — declaring 64 while returning 2 is exactly what happened here, and a test on the constant would have passed throughout.

Recorded but not decided

No silence-defect label. On §2's axis this looks like a misroute — a real failure attributed to the wrong cause — which the contract says qualifies, and 2 carrying a meaning the repo deliberately created makes it sharper here than in netspec. But the record is the org's public argument and adding a case is the author's call, so the reasoning is stated rather than the label applied.

Measurement originally from notes/evidence.md V14 in slicelab's frozen dossier, which recorded it and noted nobody had filed it.