[Spec] Rethink compensation/feedback semantics: move from per-stage “Core.comp” to amplifier-level policy

[KesavaViswanadha]

Section 2.9 treats compensation (e.g., Miller, Ahuja) as a configuration of an individual AmplifierStage (Core.comp { … }). That model breaks down for widely used schemes like Nested Miller, Feed-Forward (FFE), etc. because there can be system-level compensation structures that span multiple stages and nodes. This issue proposes (1) elevating compensation to an amplifier-level policy, (2) adding a first-class feedback/compensation block that can declare cross-node connections (including FFE), and (3) keeping an escape hatch for stage-local RCs when appropriate. The goal is to preserve strong typing while letting the engine own the final strategy, with prefer hints and node-scoped forbid/allow guards.