[bercedev]

Closes #118

Approach. A small _unknown_dataset_error(slug) helper in cli.py calls ExampleRegistry.get and returns the registry's own message when the lookup raises KeyError, or None when the name is registered. datasets show and datasets infer-policy-expectations check it first. For an unknown name they print error: <message> to stderr, the same form the CLI already uses for other input errors, and return 2. Only the lookup is guarded: infer-policy-expectations then calls ExampleRegistry.load as before, so a KeyError raised by a registered dataset's loader still surfaces as itself. The registry, its suggestions and the successful paths are unchanged.

$ rai datasets show bbqq
error: Example dataset 'bbqq' not registered. Did you mean: bbq? Available: …
$ echo $?
2

Tests (tests/test_cli.py, all offline):

• both commands with bbqq: exit 2, empty stdout, the message and the bbq suggestion on stderr, no traceback;
• an unknown name with no suggestion (zzzzzzzzzz): exit 2, no "Did you mean";
• rai datasets show bbq: exit 0, JSON with "slug": "bbq";
• rai datasets infer-policy-expectations bbq --limit 0: exit 0, [];
• a registered dataset whose ExampleRegistry.load raises KeyError("question") (monkeypatched): the KeyError propagates and isn't reported as an unknown name.

The first three fail on main and the rest pass on both.

Local checks (Python 3.12, offline in a container):

• python -m pytest -q tests/test_cli.py: 12 passed.
• python -m pytest -q: 704 passed, 16 skipped, 5 failed, 13 errors. The failures and errors are in tests/test_reproducible_evaluation.py and need git, which my container doesn't have. The list is identical on main.
• python -m pip check: no broken requirements.
• python -m ruff check --select E9,F63,F7,F82 .: all checks passed.
• reuse --no-multiprocessing lint: compliant.
• git diff --check origin/main...HEAD: clean.

Note: I used Claude Code while working on this.

[knisar]

Independently verified 95ab08a against main 20756a2 in a fresh Python 3.13.1 environment.

• python -m pytest -q: 722 passed, 16 skipped. The reported environment-related failures do not reproduce here.
• pip check, Ruff (E9,F63,F7,F82), REUSE and committed-patch whitespace checks pass.
• Ten additional checks pass, covering the real CLI entry point for both commands with and without suggestions, valid names, actual registered-loader errors including a misleading error message, and preservation of output files on an unknown name.
• All 18 current-head CI jobs pass.

The lookup guard is narrow, existing successful command behavior is preserved, and unrelated loader errors still propagate. No blocking findings.