Automatic contradiction detection at write time

[tomas-samek]

Summary

Contradictions currently surface only when the writer explicitly sets corrects. A real system should detect on write: if new content contradicts existing retrievable content at the same path, flag it.

What to do

• On trie_remember, run a quick recall at the memory's path before insertion.
• If high-coverage disagreement detected, require either corrects to be set or a force flag.
• Emit a contradiction event in the MCP response so the caller can decide.

Acceptance

• Writing contradicting content without corrects returns a warning instead of silently layering.
• False-positive rate on non-contradicting duplicates is near zero (Scenario 5 test 3 must still pass).

Links

• docs/design/honest_agent/tasks/07_incremental_learning.md
• tests/honest_agent_correction.rs