Module Ocannl.Precision_policy

Precision-assignment policy over a model (gh-ocannl-492 task 1): assigns storage precisions across a tensor graph instead of the user hand-annotating every tensor. This is the storage-precision sibling of the compute-precision policy in Ir.Numerics — the two are the "one roof" gh-ocannl-478 asks for.

Semantics of apply:

• Only tensor nodes with no user-specified precision are touched (Tn.get_specified_prec pre-check) — explicit annotations always win, including ndarray-backed data nodes whose precision comes from their host array.
• Only nodes whose (inferred-so-far) precision is a float are re-assigned: integer index tensors and uint4x32 RNG-state chains keep their precisions. For not-yet-forced inference chains the float check is deferred into the lazy (Tn.update_prec ~only_if), so applying a policy forces nothing.
• Assignment classes are structural: param_prec for root.params (trainable leaves), activation_prec for every other tensor's value node (op results, inputs, constants), grad_prec for all gradient nodes. Op-kind selectivity ("softmax stays f32") is the except predicate's job — or the training recipe's, when gh-ocannl-492 grows one; label matching over Tn.label is the intended idiom.
• except-ed float nodes are PINNED at the session default precision for their class (Tensor.default_value_prec / default_grad_prec), not merely skipped: precision inference propagates top-down as well as bottom-up (see the top_down_prec test), so a skipped node between policy-assigned neighbors would be inferred into the reduced precision anyway.
• A None field leaves that class alone (inference proceeds as without a policy).

Call apply after the model (and its gradient graph, if any) is fully constructed and before the first compilation: assignments to already-settled nodes fail loudly (the settlement guards in Tn.update_prec), and op construction after apply would infer from pre-policy precisions.

fp8 note (gh-ocannl-481): the codomain is Ops.prec, so when a second fp8 format (e4m3) lands as a precision, per-class fp8 assignment (e4m3 weights/activations, e5m2 gradients) is expressible here without interface changes.