rescore: public reproducibility CLI defaults to a maintainer's home paths and silently no-ops elsewhere

[reacher-z]

src/clawbench/eval/rescore.py:203-217 ships a maintainer's personal machine layout as the public defaults:

p.add_argument("--sweep-root",   type=Path, default=Path.home() / "work/ClawBench/claw-output/sweep")
p.add_argument("--models-yaml",  type=Path, default=Path.home() / "work/ClawBench/models/models.yaml")

clawbench-rescore is an installed console script (pyproject.toml:29) and both docs/scoring.md and eval/scoring.md advertise it as the way to reproduce every leaderboard number from public traces. On any machine that isn't that one, --sweep-root doesn't exist, the rglob matches nothing, and the tool prints "discovered 0 tasks" and exits 0 — a silent no-op for the exact audience we most want to succeed.

Every other code path resolves paths through WORKSPACE_ROOT / MODELS_YAML (run_support/config.py:91); rescore bypasses that.

Ask:

1. Default --models-yaml to the workspace-resolved config.MODELS_YAML.
2. Default --sweep-root to ./test-output (where runs actually land) or require it explicitly.
3. Error loudly when the root does not exist, instead of reporting 0 tasks and exiting 0.
4. While there: rescore only accepts the api_key scalar form, rejecting the api_keys list form that load_model_config (config.py:192-196) normalizes — reuse load_model_config instead of re-parsing.

[reacher-z]

@Perry2004 — filed from a read-only audit of main; every claim is cited to file:line and was re-verified by hand before filing. Flagging for your triage: happy to send a PR for any of these you want picked up, or to close it if you judge the behaviour intentional.