v1.6.0 — operator_trace: infer the (A2) consumer from the model

Toward human-out-of-the-loop, operator-dependent quantization: the (A2) consumer no longer has to be declared — it is inferred from the model.

• operator_trace — an operator-regime classifier mapping every parameter tensor to the operator its output flows into (SOFTMAX_SCORE, LINEAR_RESIDUAL, GATE_SELECTION, STATE_DECAY, NORM) and, from that regime, to its (A2) quantization discipline. Two front-ends: a structural classifier (module type + name) and a best-effort torch.fx graph pass that backtraces sink operators (softmax / topk / scan) to the Linear layers feeding them — so score/gate/state tensors are tagged even under obfuscated names. The regime→discipline table encodes the operator-dependent flip (a projection's weights are the robust symmetric side under weight PTQ; its cached keys are the fragile per-channel+DC side under activation quant). Entry points: trace_operators / recommend_quantization. Seeds the SSM/MoE work.

Full details: CHANGELOG.md.