[suffro]

Implements Step 3 of .context/state/DECGUARD_V0.1_CODEX_PLAN.md.

What is new

• Production records. Strict backend-neutral JSONL/JSON with validated probability invariants, optional outcomes/routing/latency/cost, stable ids and source hashes.
• Offline checks. decguard check reports accuracy/error, calibration, threshold coverage, routing, confidence, latency and cost without calling a backend or service.
• Drift. A baseline dataset or prior check report adds confidence-distribution TV, confidence mean delta, accuracy drop and ECE increase.
• Segments. Contract metadata keys produce deterministic scalar groups with independent requirement/warning gates, including missing values.
• Reports. Versioned check_version: 0.1 JSON persists normalized records, metrics, drift, segments and checks and verifies them again on load.
• Policies. Ordered first-match routes support accept, abstain, named-backend fallback and human_review; the final route is mandatory and unconditional.
• SDK and CLI. DecGuard.from_contract(...).decide(...) and decguard run share backend normalization and the same pure policy engine. Fallback remains an explicit directive and is never invoked or optimized implicitly.
• Docs and CI. Production/policy/architecture guides, a complete refund production example, and clean-wheel smoke coverage for pre- and post-deployment commands.

Step 3 acceptance criteria

• Production JSONL is analyzed offline.
• Calibration drift can fail the contract.
• Metadata segmentation exposes a localized failure.
• Policy routes accept, fallback, review and abstain deterministically.
• CLI and Python SDK use the same policy engine.
• No central service or new runtime dependency is required.
• CI examples cover pre-deploy and post-deploy checks.

Verification

• uv run ruff check .
• uv run ruff format --check .
• uv run mypy
• uv run pytest: 269 passed, 1 skipped (the opt-in external backend test)
• syngraphe check
• Clean Python 3.11 wheel install: validate, test --all, check, and run all pass.