feat(agent): forward model settings and recover from thinking stalls [PC-4672] - #1017

Anthropic won't honor a forced tool_choice while extended or adaptive thinking is on, so a thinking model can answer in plain text and never call end_execution. The old consecutive-thinking counter turned that into a hard THINKING_LIMIT_EXCEEDED on the first occurrence. Now one tool-less turn is tolerated and retried with thinking off and the tool call forced (react/forced_extraction.py), which every provider honors; only a second stall fails, with the same error code.

Stall accounting moved out of the router and into the LLM node, because the router can't tell whether forcing actually survived on the wire: the Bedrock handlers silently downgrade any to auto under thinking. A tool-less turn with content now always loops back to the agent. AgentGraphConfig.thinking_messages_limit is no longer read anywhere; I left the field in place so existing agent.json files don't break, but it's dead, and I'd rather deprecate it properly in a follow-up than widen this PR.

New chat/thinking.py holds the transport-specific knowledge in one place (native thinking attribute, model_kwargs for Bedrock Invoke, additional_model_request_fields for Converse) so the payload handlers and the ReAct loop stop each reimplementing it. Replaying reasoning blocks on a thinking-off call 400s, so the extraction retry strips them and appends a user turn if the history ends on an AI message. That trailing user turn is load-bearing rather than cosmetic: without it the retry ends on an assistant message, which is a prefill, and prefill 400s on Claude 4.6 and later.

Two smaller things ride along. get_chat_model gains a model_settings passthrough for the client-side work in uipath-langchain-client, with a warning when the legacy clients would silently drop it, and exceptions/licensing.py folds into exceptions/llm.py so both LLM error shapes are mapped from one module. The fold is a pure move: raise_for_provider_http_error behaves exactly as it did on main. An earlier revision of this branch also made it withhold the vendor message from the run record, which I reverted, so provider error text still reaches telemetry the way it does today. That's worth fixing, but at the telemetry boundary rather than at every raise site, and not here.

Start with react/llm_node.py; everything else follows from the escalation ladder it implements.

Version bumped to 0.16.15, and the uipath-langchain-client floor moves to 1.18.3 on all seven extras since that's where the model_settings parameter lands.

Verified with the 185 tests across the touched files, plus the full suite, mypy over the repo, and the integration matrices in CI.