aprender-train: load_adapter_peft reads every tensor as f32 with no dtype check (bf16/f16 adapters load as half-length tensors, in HashMap order) #4697

Proposed by la-0.72 (cop inbox, 2026-10-04 07:27Z, PROPOSE-TICKET x3, item (1)). The whole line, verbatim:

07:27Z | la-0.72 | L2 | PROPOSE-TICKET x3 (aprender-train, R15a follow-ups, from draft 16): (1) load_adapter_peft reads every tensor as f32 with no dtype check and returns them in HashMap order, so on main a bf16/f16 adapter loads as half-length tensors; (2) InstructPipeline::from_pretrained only warns when the adapter file cannot be read, and trains fresh adapters; (3) pay down instruct try_init_cuda (cyclomatic 17, cognitive 41) and cuda_train_step (18/40); the pre-commit hook makes this a prerequisite of R15a C2b/C3.

Branch evidence: la-0.72's R15a draft 16; for (3) the paydown is pushed on r15a-c2b-paydown @ a74c951 (inbox 08:04Z: try_init_cuda 17/41 -> 6/11, cuda_train_step 18/40 -> 9/20).

Train 0.72, under epic #4000 (E5 0.72: Train What You Serve). Owner: la-0.72. Branches only until 0.72 opens (C277.7). Minted by the cop's clerk.