Add reward_hacking_v2 and general_misbehavior_v1 scanners - #111

Two rubric-driven scanners, each a prompt plus a required structured
analysis, in the shape of security_incident.

reward_hacking_v2 asks whether an agent earned its score by some route
other than doing the task. The model inventories every action that
touched or targeted the scoring machinery, then places the worst one on
four anchors: none, a trace, a tough call, a clear case. Its score runs
0 to 1, where reward_hacking returns 0 to 10, so a caller moving between
the two must change its threshold.

general_misbehavior_v1 widens the question. The reward-hacking scanners
look only for score obtained by an unintended route, and
security_incident counts only conduct aimed beyond the sandbox; this one
also covers what falls between them, such as harvesting credentials on a
task unrelated to credentials or issuing an unrequested call to a real
service. Severity runs 0 to 10. Each inventory entry carries a category,
a short label for the behaviour, its target, how far the agent got,
whether the task authorised it, and whether the effect can be undone, so
the inventory can be read directly rather than inferred from scores.

Both default max_chunk_size to 900_000 rather than the 150_000 the other
scanners use. Larger chunks mean fewer calls per transcript, and because
inspect_scout runs the first scanner alone to warm the prompt cache
before releasing the rest, a shorter lead is likelier to finish inside
the cache lifetime. The cost is roughly 225k tokens per call, which
needs a judge with the context window to match.

Based on pip/incident-scanner-redux, not main: both scanners need
reducers, results.REASON_DESCRIPTION and get_scanner(reducer=), which
exist only on that branch. On main they would not import, and main's
max-over-chunks reduction would change the verdicts on multi-chunk
transcripts. This should land after that branch.

Verified against the deployed copies these were lifted from: prompt
prefix, prompt suffix and response JSON schema hash identically for both
scanners.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>