docs: document thinking-level model suffix in subagent tool surfaces (#1565)

* docs: document thinking-level model suffix in subagent tool surfaces

The subagent tool's flat schema exposes a `thinking` field that only
applies to action='watchdog.configure'. Calling models routinely pass it
on dispatch calls, where it validates and is silently ignored, so
children run at the agent's pinned thinking level no matter what the
orchestrator requested.

The supported per-run override is a thinking suffix on the model string
(provider/id:high etc.), which wins over agent frontmatter, but none of
the model-facing surfaces mentioned it: the tool descriptions had no
occurrence of thinking levels, the model param description didn't
mention suffixes, and the thinking field description didn't say it is
watchdog-only.

Document the suffix mechanism in the full, compact, and prompt-guideline
model bullets, the dispatch model param description, and the thinking
field description.

* docs: state suffix precedence in compact description, assert new guidance

Review feedback: the compact model bullet omitted that the thinking
suffix wins over the agent default (present in the full description and
prompt guidelines), and the test only bumped the byte budget without
asserting the new guidance. Add the precedence clause and pin the
suffix syntax, precedence, and watchdog-only scope in all three
surfaces.

* docs: credit thinking suffix guidance

Co-authored-by: Yanqiang Lu <yanqianglu@users.noreply.github.com>

---------

Co-authored-by: Yanqiang Lu <yanqianglu@users.noreply.github.com>
Co-authored-by: Nico Bailon <nico.bailon@gmail.com>