test(coverage): enforce ≥80% coverage + README badge — 86% baseline

[cipher813]

Summary

Closes the final leg of Brian's "test infrastructure before soak" plan: enforce ≥80% test coverage as a CI gate + add a coverage badge to the README. Current measured coverage: 86.43%.

Composes with PRs #158 (Python integration canary) and #159 ([server]-extras CI + layer3 --exercise-all-tools) as the third leg of the test-trio addressing the 2026-05-22 memory_check_contradictions failure class.

What ships

pyproject.toml coverage config

New [tool.coverage.run] + [tool.coverage.report] sections with fail_under = 80. Excluded modules are documented inline:

Module	Why excluded
dashboard/*	Streamlit UI; tested by running the app, not by pytest
__main__.py	4-line entry-point shim, no logic
upgrade.py	Real Fly+AWS interactions; tested by Layer-3 ritual
downgrade.py	Same
llm.py	Deprecated optional-LLM path. Deployed product is LLM-free by design (2026-05-22). NLI replaced the only production use in mnemon.contradiction.

.github/workflows/ci.yml coverage gate

pytest --cov runs on every CI build. Coverage drop below 80% fails the build.

README badge

[![Coverage](https://img.shields.io/badge/coverage-86%25-brightgreen.svg)]() — static, shields.io, matches the existing pattern (Status / Python / License / MCP). Manually updated on each release; no SaaS / codecov dep. Composes with mnemon's broader "minimal external deps" posture (no LLM, no auth vendor, no observability vendor — and no coverage vendor).

4 new tests/test_nli.py tests

Push nli.py from 72% → ~85% coverage. Cover error paths I personally introduced in PR #157 + #158:

1. _ensure_loaded raises NLIUnavailableError on HuggingFace download failure (connection error, 404, etc.)
2. _ensure_loaded raises on unexpected label set — a model with different output classes fails fast, not silently mis-classifies downstream
3. prewarm() swallows unavailability cleanly (acceptable-secondary-observability category per feedback_no_silent_fails)
4. classify_pair softmax + ONNX input-building path exercised with a stubbed session (lines 164-189 of nli.py)

Coverage breakdown (top 10 by lowest coverage, all above gate)

Module	Coverage
cli.py	61% — interactive CLI dispatcher; many branches require real terminal
server_remote.py	63% — OAuth + transport flows, partially covered by test_server_remote.py
context_surfacing.py	71% — hook with many failure-mode branches
nli.py	~85% (was 72%, +4 new tests)
persistent_sessions.py	77%
uninstall.py	81%
mirror.py	84%
search.py	85%
doctor.py	85%
auth.py	87%

These are P3 ROADMAP items if we ever want to push individual modules higher; the project-level 80% gate is the operational target.

Test plan

• pytest --cov reports 86.43% (above the 80% gate)
• Full suite 850 → 855 passing
• Coverage config validated locally
• Post-merge: CI workflow validates the gate runs

What's next

Once #160 merges, the test substrate is at the standard Brian set ("known bugs fixed + verified before soak"). Then:

1. Bump 0.7.0rc2 → 0.7.0rc3 (rolling in the test-trio + coverage gate as a soak-substrate version)
2. Republish PyPI
3. Fly redeploy via mnemon upgrade web (operator-driven)
4. Optional: run scripts/promote_stable.sh layer3 --exercise-all-tools against a test Fly app for the first time (~15 min, validates the new Fly-level probe end-to-end)
5. Then operator activates Phase 1 standing tier per the soak workflow (demote chore: promote 0.6.0rc18 → 0.6.0 stable #131 + #2402 → flip MNEMON_STANDING_TIER_ENABLED=true → fresh Claude Code → ≥1 week soak)

🤖 Generated with Claude Code