feat(contradiction): NLI-based detection (no LLM) + dry_run + 0.7.0rc2 bump

[cipher813]

Summary

Rebuilds memory_check_contradictions on a Natural Language Inference cross-encoder instead of the LLM classifier. Closes the long-latent defect that check_contradictions has been silently broken on Fly since the [server]/[llm] extras split (no llama-cpp-python in [server] per the 2026-05-21 "mnemon is LLM-free by design" decision).

Surfaced 2026-05-22 when claude.ai called memory_check_contradictions on memory #2543 (composite runway) and got an opaque Error occurred during tool execution envelope (Anthropic MCP proxy timing out on the slow/broken LLM path). Brian's framing: "is it a requirement to use an llm to check contradictions? there is no ml tool that can be used instead?" — yes there is, it's NLI, and it's the right institutional fit.

SOTA / institutional fit

• NLI (entailment / contradiction / neutral on a sentence pair) is the named ML task for exactly this problem
• cross-encoder/nli-deberta-v3-xsmall: 22M params, ~87 MB INT8 ONNX, ~10-20ms per pair bidirectional on CPU
• Mirrors mnemon's existing FastEmbed ONNX pattern (lazy load, prewarm at lifespan, baked into Fly image)
• Zero new pyproject deps — onnxruntime + tokenizers + huggingface_hub all transitive via FastEmbed already

Mapping NLI → mnemon taxonomy (bidirectional)

premise→hypothesis	hypothesis→premise	mnemon label
contradiction (either direction)		contradiction
entailment	entailment	same (semantic equivalence)
neutral	entailment	update (new supersedes old)
entailment	neutral	same (existing dominates)
neutral	neutral	unrelated

Cosine gate (CONTRADICTION_OVERLAP_THRESHOLD=0.7) preserved upstream of NLI to filter obviously-unrelated pairs.

dry_run parameter

Closes the read/command-separation violation (the check_* naming on a mutating function). When dry_run=True, the tool reports what WOULD have decayed without applying mutations (no confidence changes, no relations inserted). Useful for operator audit before committing destructive changes.

Clean error surface

When NLI isn't loadable (model download fails on a fresh local install without network), MCP tool returns Contradiction check for #X skipped — NLI classifier unavailable on this server (model load failed). ... No vault state was modified. Instead of an opaque envelope. Fail-loud per feedback_no_silent_fails.

Files

New:

• src/mnemon/nli.py — model wrapper, lazy-load singleton, classify_pair, classify_pair_bidirectional, prewarm, is_available, NLIUnavailableError
• tests/test_nli.py — 11 tests covering bidirectional label mapping, error surfacing, availability

Changed:

• src/mnemon/contradiction.py — LLM imports + prompt construction removed; NLI pipeline explicit; return shape gains nli_unavailable + dry_run flags
• src/mnemon/server.py — memory_check_contradictions(id, dry_run=False); clean nli_unavailable surface
• src/mnemon/server_remote.py — NLI prewarm in lifespan startup (parallel to FastEmbed)
• Dockerfile — bake 87 MB INT8 ONNX model + tokenizer; HF_HOME set; health check start period 30s → 45s
• tests/test_contradiction.py — mock NLI instead of LLM; +2 tests for dry_run and nli_unavailable

Version bump: 0.7.0rc1 → 0.7.0rc2.

Test plan

• 11 new NLI tests passing (pytest tests/test_nli.py)
• Refactored contradiction tests passing (pytest tests/test_contradiction.py)
• Full suite 836 → 847 passing, no regression
• End-to-end smoke test with real NLI inference: $500K-vs-$200K liquidity pair correctly classified as contradiction via the MCP tool; dry_run respected (would-decay count surfaced, no mutations applied)
• Post-merge: tag v0.7.0rc2, GitHub Release, python -m build, twine upload
• Operator runs mnemon upgrade web --app-name mnemon-memory --mnemon-version 0.7.0rc2 to ship Dockerfile changes (NLI bake) to Fly
• mnemon doctor 7/7 + smoke-test memory_check_contradictions from claude.ai
• Then proceed with Phase 1 standing-tier soak

🤖 Generated with Claude Code