fix(ai): only capture terminal Responses statuses as stop reasons

[bernatixer]

Problem

• A LangChain Responses run started with background: true returns while still queued, and the callback records $ai_stop_reason: 'queued' - a lifecycle state, not a stop reason.
• The same truncated run gets two spellings: the LangChain callback reports max_output_tokens, the native OpenAI wrapper reports incomplete. The PostHog trace view only recognizes the first, so native-wrapper truncations render as normal endings.
• A comment in the callback claims the two surfaces already match.

Follow-up to #4700, from its post-merge review.

Changes

• Runs that OpenAI truncated or ended abnormally now carry the same $ai_stop_reason from both the LangChain callback and the native wrapper, and queued or in-progress background runs carry none.
• One shared responsesStopReason helper does the mapping: non-terminal statuses yield nothing, an incomplete run is named by incomplete_details.reason (e.g. max_output_tokens), the other terminal statuses stand for themselves.
• The LangChain callback, the native telemetry builder, and the stream accumulator route through the helper.
• Mechanical: the callback comment now states what the code does; changeset added.

Note

Behavior change: the native wrapper reports a truncated run as max_output_tokens instead of incomplete. Dashboards grouping on $ai_stop_reason see the new value after this ships. Cancelled runs stay cancelled.

Release info Sub-libraries affected

Libraries affected

• @posthog/ai

Checklist

• Tests for new code
• Accounted for the impact of any changes across different platforms
• Accounted for backwards compatibility of any changes (no breaking changes!)
• Took care not to unnecessarily increase the bundle size

If releasing new changes

• Ran pnpm changeset to generate a changeset file

Testing

• New LangChain row: status: 'queued' produces no stop reason - fails without the terminal gate.
• The completed-run row now carries incomplete_details: null, the shape every real completed run has.
• The native openai, azure-openai, openai-stream-accumulators, and openai-background-responses suites expect max_output_tokens for the incomplete status.
• The streaming fallback response object drops its status field: stopReason is set only alongside terminalResponse now, so the old cast could never hold a real status.
• Not run: the integration suites, because they need provider API keys.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written by Claude Code from the confirmed findings of a max-level review of #4700: the missing terminal-status gate, the stop-reason vocabulary split between the two wrappers, and the stale comment. The remaining review findings (streamed truncations on @langchain/openai <=1.4.6, untested generation-level metadata legs) were left out as lower priority. Skills invoked: code-review, writing-pr-descriptions.

[marandaneto]

@bernatixer make sure to tag @PostHog/team-ai-observability for reviewer as well
the client libs team can check and approve changes basically looking at raw code/quality/etc, but since we dont know deeply about all products built in the sdk, I can't always tell if this change makes sense or not product-wise

[Radu-Raicea]

This can return incomplete before checking the reason in generationResponseMetadata. Is that fine? Could the reason be max_output_tokens?