[MAabdelmoumen]

Backports StructuredOutputGuard from oracle#210.

Why

oi-fork took the single-string shortcut from oracle#210 in 0242c9a, but not the guard that pairs with it. So every agent here that keeps response_format is unbounded.

Without it, a model that never produces a structured response fails two ways:

• no tools — the agent loops. create_agent sets recursion_limit=9999, so that is thousands of model calls before a GraphRecursionError naming neither the agent nor the schema.
• with tools — routing treats a turn with no tool calls as done, so the agent exits with no structured_response and the declared outputs come back empty, with no error at all (langchain#36349).

This became more pressing after f111e7a: gating on the declared output format moves one-field JSON schemas out of the shortcut and into the response_format path, which is the unguarded one.

What it does

after_model counts consecutive turns ending in an AIMessage with no tool calls. after_agent catches the silent exit that budget never sees. Turns with tool calls do not count, since under ToolStrategy the structured response is itself a tool call.

Attached only when a response model was built, so it composes with f111e7a: an agent declaring TEXT never builds one and never sees the guard.

Testing

Local kind stack, model nvidia/Qwen3.8-27B-NVFP4:

case	result
JSON, one String field	{"sentiment":"positive"}, 2 steps, guard fires after_model then after_agent
Output format TEXT	prose, 1 step, zero StructuredOutputGuard.* events
JSON, two fields	both fields, 2 steps
flow: agent → branch on the field	routes correctly, no ValueError

The TEXT case is the one that matters: the raw event stream shows the guard is never attached when no response model is built, so plain chat agents are untouched.

Not exercised: a model that actually refuses structured output — the registry here offers only one model and it complies. The guard's firing path is therefore covered by oracle#210's own tests rather than mine.

Scope

Guard source is taken unchanged from oracle#210. Only the wiring and the _types.py exports are local. oracle#210 also moves is_single_string_output to adapters/_utils.py; that is omitted here to keep the diff small.

Refs OA-1681