Run find_contradictions automatically inside recall and surface debug ranking

[ramonlimaramos]

Problem

Two hygiene gaps:

1. find_contradictions() is a tool nobody calls

The tool exists but agents rarely invoke it proactively. Over weeks, memories can drift and contradict (e.g. 'use library X' vs 'we migrated off library X to Y'). The contradiction is only caught the next time the user catches the agent applying the stale rule.

2. recall ranking is opaque

recall() returns memories with score + decay + trust, but the agent has no insight into WHY a memory ranked above another. When two contradictory memories both have score ~0.1, the agent can pick the wrong one and not know.

Proposal

Inline contradiction detection

recall() runs a lightweight contradiction check on its result set (top-K memories) and surfaces conflicts inline:

[working] (project) score=0.11 decay=1.00 trust=0.50
  Use library X for date formatting.
  id=aaa-...

[working] (project) score=0.10 decay=1.00 trust=0.50
  We migrated off library X to Y on 2026-04-15.
  id=bbb-...

⚠️ Contradiction detected between aaa-... and bbb-...
   Prefer the newer memory (bbb-...) or call resolve_contradictions(aaa, bbb) to pick one.

Optional debug mode for ranking

recall(query, debug=True) includes per-memory:

score: 0.115
breakdown:
  semantic_similarity: 0.81 (cos)
  keyword_overlap: 0.42 (BM25)
  hrr_match: 0.55
  decay_factor: 1.00
  trust_factor: 0.50
final = semantic*w1 + keyword*w2 + hrr*w3 - age_penalty

The agent (and the user inspecting traces) can see why a result ranked higher and trust the ordering.

Why this matters

Memories accumulate. Without proactive hygiene, the same human-rule will appear in 3 versions over 6 months and the agent will keep guessing.

Acceptance

• recall() flags contradicting pairs in its result set with their ids.
• recall(debug=True) returns the ranking breakdown for each result.