[Decision Policy] Implement rule priority, score semantics, conflict handling, and abstention

[Dyu20705]

Problem statement

The MVP uses deterministic rules, not calibrated probabilities. RunSift needs a central decision policy that combines rule matches and evidence quality conservatively, resolves conflicts deterministically, and abstains when the result is not safe to state.

Checklist

• Define documented rule-priority and specificity policy
• Define rule_strength inputs and scale without calling them probability/confidence
• Require minimum evidence quality and evidence references for non-unknown results
• Resolve same-category duplicate matches deterministically
• Detect cross-category conflicts and contradictory evidence
• Implement decision_source: rule|abstain for the MVP
• Implement stable abstention reasons such as insufficient evidence, conflicting rules, unsupported input, parser degradation, and policy threshold
• Keep category decision separate from temporal assessment
• Require explicit comparable-attempt evidence before suspected_transient
• Prohibit maintainer_confirmed_flaky without maintainer confirmation
• Version policy configuration and include policy/rule versions in output
• Add boundary, conflict, ordering, ambiguity, and regression tests

Acceptance criteria

• Consumers can branch on decision, decision source, category, and abstention reason without undocumented numeric thresholds
• Every non-unknown result has at least one redacted evidence reference
• Conflict resolution is deterministic and tested independently of rule registration order
• Weak or contradictory evidence produces unknown/abstention
• Rule scores are clearly typed and never described as calibrated probability
• Policy behavior is versioned and reproduced in JSON/golden tests
• A one-attempt input cannot confirm flaky behavior
• Recommendations remain safe when the policy abstains

Future model boundary

A future decision_source: model may expose calibrated probabilities only after #36 authorizes ML work and a validated evaluation/calibration issue proves the semantics. Model behavior is out of scope here.

Dependencies

• Parent epic: [Epic] Deterministic Failure Decision Engine #17
• Rules: [Classification] Implement deterministic dependency, test, lint, and build rule families #22 and [Classification] Implement environment, network, workflow-configuration, and unknown rule families #23
• Schemas: [Core] Implement versioned input, evidence, and output schemas #34
• Fixtures: [Fixtures] Build sanitized, provenance-tracked offline regression corpus #9
• Reports: [Reporting] Generate safe terminal, Markdown, and stable JSON triage reports #13