Module precision_policy 

2026-09-28: PolicyPrecision, the EdgePrecision the engine’s own precision answers give: the checkpoint’s declared plan (DeclaredPrecisionPlan) under the served tier (WeightQuantPolicy), plus the formats the engine chooses itself (a recipe’s lm_head, the MTP head), which the checkpoint does not declare.

Owner: metrale-circuit. Invariants:

• Engine formats are stated, never inferred: a module they name is answered from them first, in order; every other module from the policy.
• Under the declared tier a module declared FP8 runs its FP8 weights, with FP8 activations exactly when WeightQuantPolicy::fp8_decode_act says so; one declared FP4 runs NVFP4 weights, with FP4 activations when the checkpoint declares them. Under nvfp4 every quantized module runs NVFP4 weights and 16-bit activations.
• An unquantized module is BF16 weights and activations.

Structs

CheckpointPlan

2026-09-28: A checkpoint’s quantization_config, as the circuit fixtures store it (kernels/circuits/checkpoints/<name>.toml).

PolicyPrecision

2026-09-28: The engine’s precision answers as an EdgePrecision.

Functions

caps_named

2026-09-28: The kernel capabilities named names (field names of KernelCaps).

tier_named

2026-09-28: The tier named name (declared, nvfp4).