Seams for auditing: typed output observation and explicit table access

[UnknownJoe796]

Stack: 1 of 16. Base: version-5.3. Merge in order.
Three changes to existing code that exist so an audit layer has somewhere to
attach. Nothing fires them yet; the audit modules come later in the stack.

• TypedOutputInterceptor observes a typed value after the handler produces
  it and before it is serialized. Firing before serialization is the point:
  the response body does not exist yet, so an observer that throws prevents
  the disclosure rather than reporting it after the fact.
• ModelInfo splits audited from unaudited table access. baseTable() keeps
  skipping permissions but now routes through the existing log decorator,
  and the undecorated path becomes dangerouslyDirectTable() behind an
  opt-in @UnauditedDatabaseAccess. The goal is not to forbid the bypass but
  to make every instance greppable.
• ModelRestEndpoints declares the fields it discloses.

log already existed with an identity default, so with no audit layer
installed this is behaviour-neutral. dangerouslyDirectTable() has a default
implementation, so it is not a breaking change for existing ModelInfos.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jybc9zdLT2sEfJUpErf2cd

[UnknownJoe796]

Passes my review.