fix(alerts): fire change_pct rules on spikes from a zero baseline (audit_28 §5)

[brownjuly2003-code]

Summary

Closes audit_28_06_26.md §5: alert change_pct rules did not fire on a spike
from a zero baseline.

evaluate_rule reported a flat 100% change whenever the previous-window value
was 0. So an "above" change rule with threshold > 100 never fired on a spike
from zero (e.g. revenue 0 → 500 with a +200% rule stayed silent), and the
100 sentinel was a misleading number for what is an undefined ratio.

Fix: a change from a zero baseline is a mathematically undefined (unbounded)
ratio. Report change_pct as None and decide firing by direction — any
rise satisfies an "above" rule (threshold >= 0), any fall a "below" rule. A
flat 0 → 0 still reports 0% and does not fire. The non-zero-baseline path is
unchanged.

Tests

New tests/unit/test_alert_evaluator.py:

• rise from zero fires an "above" rule even past 100% (the bug)
• no-change 0 → 0 does not fire
• fall from zero fires a "below" rule
• non-zero baseline still computes a numeric pct (regression guard)

Verification (no-Docker)

ruff / format / mypy clean; new unit tests pass; existing alert unit suite green;
diff LF-clean.

🤖 Generated with Claude Code

[github-actions]

DORA Metrics

• Window: last 30 days
• Branch: main
• Deployment frequency: 167 total / 38.97 per week
• Lead time for changes: avg 0.27h / median 0.0h
• Change failure rate: 59.28% (99/167)
• MTTR: 0.23h across 4 incident(s)