binding-llm: canonical section attributes #2637

Depends on: #2634

Summary

Name-value attributes on LlmDataEx that describe a canonical section independently of any dialect, so a block decoded from one dialect can be encoded into another.

A section type alone does not carry everything llm(client) needs to encode a block. A canonical attribute is a dialect-neutral concept a section can carry, named independently of any dialect.

Eight, and no more without a reason: a new attribute must be a dialect-neutral concept that the section type cannot express. Anything specific to one dialect is a dialect extension (#2638). Names are canonical, never dialect-specific.

The location of a url source is not an attribute. It is unbounded, and the payload already carries it, so hoisting it would duplicate an unbounded value into a bounded extension. A url section is classified by source alone and its location streams in the payload. A ref-id is a bounded provider handle and is hoisted.

Acceptance criteria (test-first)

• Unit tests: decode populates each attribute for openai and anthropic (and ollama where it applies), and never on a section that does not carry it.
• Unit tests: llm(client) encode renders each attribute into the target dialect's wire form where the dialect has one.
• .rpt coverage for source: url and source: reference blocks.