Confidence decay for signal groups (ADR 022)

[lance0]

Summary

Adds exponential confidence decay to signal-group derived_confidence computation. Older per-event confidence contributions are multiplied by 0.5 ^ (age_seconds / half_life_seconds), so corroborating evidence from earlier in the correlation window smoothly loses weight without ever being discarded.

Closes ROADMAP item: "Confidence decay over time" under Confidence Model.

Why

The correlation engine (ADR 018) computes derived_confidence as a source-weighted average. With long correlation windows (e.g. window_seconds: 3600) and ADR 021 corroborating signals, evidence that fired hours ago continues to inflate the score at full weight long after the underlying signal has gone silent. A naive cutoff loses history and produces step discontinuities; exponential decay shapes contributions smoothly.

Design (ADR 022)

• Global config: correlation.confidence_decay_half_life_seconds: u32 (default 0, disabled). Validated 0 ≤ H ≤ 10 × window_seconds.
• Per-playbook override: correlation_override.confidence_decay_half_life_seconds: Option<u32>. Some(0) explicitly disables for the playbook even when the global is non-zero; None falls through.
• Compute paths: event ingestion and a new reconcile-loop refresh_decayed_confidence step (no-op when decay disabled).
• One-shot corroboration_met: sticky — once true, decay never reverts it to false. Preserves the invariant that authorized mitigations stay authorized.
• Metric: prefixd_signal_group_decay_refreshes_total counter ticks per reconcile step.
• UI: Group detail page surfaces "decayed, half-life Ns" when active.

Tests

• 7 engine unit tests (decay math, single-event no-op, multi-event weighting, disabled fallback).
• 7 config unit tests (override resolution, validation bounds, default).
• 3 integration tests:
  • test_decay_refresh_lowers_derived_confidence
  • test_decay_one_shot_corroboration_no_flap_back
  • test_decay_disabled_does_not_touch_groups

Full suite: 236 unit + 130 integration + 16 postgres pass. Frontend: 87 tests + build clean. cargo fmt --check + cargo clippy --all-targets --features test-utils -- -D warnings pass.

Migration

None. Default confidence_decay_half_life_seconds: 0 preserves ADR 018 behavior bit-for-bit. Operators opt in by setting a non-zero value in correlation.yaml.

ADR

See docs/adr/022-confidence-decay.md for full rationale, alternatives considered, and consequences.