Use transport regressions for Server transport and memory fixes

[rmcdaniel]

Refs durable-workflow/server#141 and durable-workflow/server#142.

The current automatic Server guard classifies streaming/reference and metadata-read changes as codec changes, then requires new tiny wire fixtures and per-file rewritten-source proof. That is not evidence for a 64 MiB HTTP/SAPI or memory failure, and the proxy resolver signatures do not represent the current API. Adding unrelated fixtures or another proxy generation would make the process larger without proving the product behavior.

Clarify the evidence contract publicly before updating Server:

• Existing wire fixtures remain immutable and run normally.
• Actual wire/type/framing/codec acceptance defects still require applicable new portable fixtures and their existing proof checks.
• When no new wire fixture is appropriate, Server reports changed payload paths for explicit maintainer review; the PR must carry the real defect-specific HTTP/resource/recovery reproduction, including the original failure.
• A green inventory check is not release proof; no product test or failed reproduction is waived.

This adds no coordinator, per-PR exemption file, new proof format, or source instrumentation. The concrete transport fix already has exact-limit, over-limit, concurrency, retained result and cold-recovery tests. This PR changes guidance only.