[tecnosam]

Summary

Replaces the logarithmic pool with an any-of rule. Averaging the checks let the ones that said "no" pull down a trace that had one real failure, which lowered recall.

Changes

• Any-of decisions: a trace is kept when answer_wrong, or any question in the manifest's decision.any_of, reaches its threshold.
• Per-question thresholds: decision.thresholds sets a question's own threshold; questions without one use policy.keep_threshold.
• Priority: answer_wrong ranks first, then any_of in the order written. Ranks become rank-order centroid weights that sum to 1.
• Severity: the sum of weight × answer over a trace's flags. It ranks kept traces for review and does not change whether a trace is kept.
• Flags: the questions that reached their threshold, listed in priority order.
• Output: answers, flags and severity are in the decision log and on exported traces (eval.answers.<question>, eval.flags, eval.severity), and they travel through the export queue.
• Validation: a manifest fails to load if a question is unknown, isn't a yes/no question, is listed twice, or has a threshold outside (0, 1].
• Removed: pipeline.Pool and the decision.intercept / decision.weights manifest keys.
• Docs: README section "Any question can keep a trace"; issues/export-by-severity.md proposes exporting in severity order.

Notes for reviewers

• Behaviour change: six questions can now keep a trace, so expect a higher keep rate. Tune with decision.thresholds or a shorter any_of.
• Backward compatible: a manifest without a decision block decides exactly as before.

Testing

• make lint and make test pass. New tests cover the weights, per-check decisions and severity, manifest validation, the log fields, the export attributes and the queue round trip.
• Not yet run against the live Jev.