[E1] `tether verify` never runs the export artifact; both arms trace to native weights **[flagship gate hole]**

[DsThakurRawat]

Labels: bug, parity, verification, priority:high
File: src/tether/verify.py:395 (gather_paired_samples), src/tether/runtime/fast_inference/libero_adapter.py:53-89 (from_policy rebuilds from native float32), TODO admission at verify.py:387

Description. The "optimized" arm is built via TritonLIBEROAdapter.from_policy(policy) where policy is the native PyTorch checkpoint, not the export under test. The export dir is only read for the expert_takes_state flag; its .onnx/.trt/.data files are never loaded.

Why it matters. A customer exports a numerically-broken ONNX (bad opset, dropped Cast, weight-fusion bug), runs tether verify ./export, and both arms share the same native weights → parity is trivially high → PASS. The exact regression verify exists to catch is invisible. This directly undermines the product's "deployment confidence" pitch.

Tasks.

• Load the actual export (dispatch on tether_config.json["export_kind"] → ORT / decomposed / Triton-from-engine) for the optimized arm.
• Fail loudly if the export artifact can't be loaded, instead of silently falling back to from_policy.

Dedup. #104 is the round-trip "mirror bug" blindspot; #30/#31 add cosine/pi0.5 to round-trip validation; #198 (PR) is the action-parity metric engine. None make verify load the exported graph. Distinct.

[rylinjames]

Post-merge proof: protected-main workflow run 33161226716 executed the generated ONNX export successfully at merged SHA bf0f190. Downloaded receipt was independently checked for repository, workflow ref, push event, refs/heads/main, run ID/attempt, passed=true, and 64-hex model/config/action hashes. Run: https://github.com/FastCrest/tether/actions/runs/33161226716