Add failing_scanner for deterministic scan-error testing - #1

What

Adds failing_scanner — a scanner whose execute always raises — and registers it in _registry.py alongside the existing word_counter. Both scanners are now resolvable from the inspect_test_utils package.

@scanner(messages="all")
def failing_scanner() -> Scanner[Transcript]:
    async def execute(transcript: Transcript) -> Result:
        raise RuntimeError("failing_scanner: deliberate failure for testing")
    return execute

Why

Hawk's online-scanning smoke test drives a scan whose model-full fixture references word_counter (succeeds) and failing_scanner (fails), asserting exactly one non-fatal scanner_result with scan_error. word_counter already exists on main, but failing_scanner did not exist in any branch/tag/commit.

The nearest existing scanner, random_failure (on the unmerged add-random-failure-scanner branch), is probabilistic (random.random() < failure_rate), so it can't back a deterministic "exactly one scan_error" assertion. failing_scanner always fails, which is what a smoke/error-path test needs.

Notes

• Draft: exercised end-to-end by Hawk's online-scan smoke scenario (test_online_scan) rather than a unit test here (the repo has no test_scanners.py yet). Happy to add a direct unit test if preferred.
• Once merged, Hawk's online_scan.yaml fixture can pin the scanner package to a recent main commit that carries both word_counter and failing_scanner (it currently pins an older off-main commit that has neither the failing scanner nor the latest main state).
• ruff check / ruff format --check clean.

🤖 Generated with Claude Code