[C2] Receipt-gated parity/latency tests always skip in CI (green proves no parity)

[DsThakurRawat]

Labels: ci, parity, testing, priority:high
File: tests/test_decomposed_per_step_parity.py, _overhead.py, _e2e_latency.py, test_pro_consent.py — each pytest.skip("No receipt …"); receipts produced by scripts/modal_per_step_*.py, never run in pytest.yml

Description. The tests that assert cos≈1.0 parity, IOBinding overhead, and e2e latency budgets pytest.skip when their Modal receipt file is absent — which is always on a GitHub runner.

Why it matters. The gates that would catch an action-parity or latency regression are structurally guaranteed to skip in CI. A regression that breaks the product's core parity claim merges fully green.

Tasks.

• Add a @pytest.mark.receipt marker so their skipped status is visible (not blended into normal skips).
• Run the Modal jobs in a scheduled/pre-release workflow that uploads receipts as artifacts consumed by the pytest step.
• Make release branches xfail/fail-loud when receipts are absent.

Dedup. #52 is STRICT_GATE for the finetune drop-gate; PR #231 makes that gate fail-closeable. Neither touches these receipt-gated parity tests. Distinct.

[rylinjames]

Fail-closed receipt infrastructure merged in PR #318 at 66fa4b7. It binds receipts to the exact repository/ref/SHA/workflow/run/attempt/artifact IDs and digests, immutable model revision/export/config hashes, required matrix cells, CUDA providers, and nested E2E results; per-run remote namespaces prevent cross-run cache/result reuse. Final CI and independent review are green. Keeping this issue open because production proof still requires: (1) select at least one independent environment reviewer and protect the parity-receipts-production environment against self/admin bypass, (2) configure MODAL_TOKEN_ID, MODAL_TOKEN_SECRET, and HF_TOKEN plus PARITY_RECEIPTS_ENVIRONMENT_PROTECTED=true and PARITY_MODEL_DIGEST, and (3) explicitly authorize and complete the paid A100 workflow, then attach and verify its receipt artifact. None of those external/paid actions were inferred or performed.

[rylinjames]

Reopening: PR #318 merged the fail-closed receipt infrastructure, but the issue's acceptance criteria are not complete until the protected production environment has an independent reviewer, required secrets/variables are configured, and an explicitly authorized paid A100 matrix run produces a verified receipt. The merge auto-closed this issue; no paid run or production credential/reviewer configuration was performed.

[rylinjames]

@DsThakurRawat, thanks again for the thorough Tether issue reports and follow-up. We are building a dedicated Tether Discord for contributors, deployment questions, and issue follow-up: https://discord.gg/wrPUdcxdPu. We would love to have you there.