[SunSi12138]

Purpose

Continue #720 from the merged Phase A production baseline (dev@172a4bab78ba117cf66832a234c7495076ff7961).

This PR is benchmark/evidence only. It does not change production Codec ABI, registry/provider behavior, generated ABI, or streaming runtime behavior.

Candidate being validated

Keep the existing class Codec as the public/registry-facing shell, but give it a readonly struct core. At stream start, open that concrete core exactly once and run the whole N-item loop as Pump<T,TCore>.

The matrix separates:

• current post-perf: hoist streaming codec capability checks to stream scope #724 production-style class/interface loop;
• class shell forwarding to struct core while still interface-called per item;
• one open per stream + concrete struct core passed by value;
• one open per stream + concrete struct core passed by in;
• statically-known direct core entry as the no-opener upper bound.

Coverage

• int / long / Guid / 16B / 64B / 256B unmanaged payloads;
• 1k / 10k / 100k streams;
• 1 / 8 / 64 item tiny streams to measure opener amortization;
• exact-sized serialize / unsized serialize / deserialize;
• flat zero-state core;
• generated-like 64B payload whose struct core retains four child IRpcCodec<Payload16> references;
• Dynamic PGO disabled and enabled;
• JIT disassembly;
• NativeAOT performance, native symbols, and binary size;
• by-value vs in TCore copy behavior.

Relates to #720.