Confidence scoring (SNUG-126)

Explainable confidence on retrieval hits and evidence-backed responses. Implemented in brain/src/hippo_brain/confidence_scoring.py.

Output shape

Each hit carries a confidence object:

Field	Meaning
`level`	`high`, `medium`, `low`, or `insufficient`
`score`	Bounded composite in `[0, 1]` from named factors
`factors`	Weighted factor breakdown (`name`, `weight`, `score`, `contribution`, `detail`)
`explanation`	Agent-readable one-liner summarizing top factors
`withheld`	`true` when level is `insufficient`

Factor	Weight	Inputs
`evidence`	0.30	Inspectable evidence packet count
`source_diversity`	0.15	Distinct `source_kind` families
`recency`	0.15	Hit `captured_at` age
`retrieval_match`	0.20	Hybrid/semantic/lexical retrieval score
`capture_health`	0.15	Inline `evidence[].freshness.status`
`context_alignment`	0.05	`cwd` / `git_branch` present

No evidence → insufficient. Hits without inspectable evidence packets are never rated high confidence, regardless of retrieval score.
Stale/failing capture → capped. stale, failing, expected_absent, or unknown freshness on cited sources prevents high and often yields low.
Single weak packet → capped at medium. One evidence packet needs very strong match + health to reach high.
Conflict caps (SNUG-127). conflict_detection.analyze_conflicts surfaces outcome disagreements, decision contradictions, and stale-only evidence; apply_conflict_confidence_caps lowers per-hit confidence when conflicts are unresolved. Do not treat high as contradiction-free until conflicts are reviewed.
Enrichment health excluded. Queue depth and LLM enrichment status are intentionally omitted — capture health only.