Flag close-call rankings instead of trusting raw confidence

[Lushenwar]

Summary

The Phase 9 eval showed Claude's confidence_score barely separates right #1 picks from wrong ones, so the number on the Slack card and dashboard was misleading. Measured on the 144 committed v2 trials (131 right, 13 wrong):

Signal	AUC right-vs-wrong
Top-1 confidence	0.60
Gap between #1 and #2	0.86

When the top two suspects are within 0.1, the Slack card now says "Close call: check both suspects" and lists both; the dashboard shows a matching banner. On the eval trials this flags 9/13 wrong picks and 9/131 right ones.

• No pipeline contract change: the flag is derived at render time from the stored ranking.
• The 0.1 cut is in-sample (13 wrong picks) and marked as a working rule in code and METRICS.md.
• It does not catch confidently wrong rankings (two wrong picks had a 0.7 gap).

Test plan

• pytest core/tests -q (29 passed), including close-call and clear-lead card tests
• black, flake8, dashboard lint + jest
• Numbers in METRICS.md recomputed with the shipped is_close_call
• CI green

🤖 Generated with Claude Code