Return error if the chat answer is empty

[hanna-paasivirta]

Short Description

workflow_chat, global_chat, and job_chat could return HTTP 200 with response: "" when the LLM produced no usable text. Lightning treated the 200 as success leaving the user-side message stuck in :processing. Each service now raises ApolloError when there's no usable text, so Lightning's handle_error_response routes it into the error tuple.

Fixes #484

Implementation Details

ApolloError usage

Each service raises ApolloError(502, ...) when text output is empty after the existing retry/loop logic. Type is OUTPUT_TRUNCATED for stop_reason=max_tokens, EMPTY_OUTPUT otherwise. Code and types are new values within the existing ApolloError shape.

Sentry tracking

We attach Sentry reporting before each raise. entry.py's existing capture_exception records one event per failed request.

Tests

Added temporary tests in separate test files that won't run automatically. This is to avoid conflicts from our testing architecture which is currently in progress. These tests will need to be rewritten to fit into the new tests.

AI Usage

Please disclose how you've used AI in this work (it's cool, we just want to know!):

• Code generation (copilot but not intellisense)
• Learning or fact checking
• Strategy / design
• Optimisation / refactoring
• Translation / spellchecking / doc gen
• Other
• I have not used AI

You can read more details in our Responsible AI Policy

[josephjclark]

I've not been able to repro but this looks good.

Interesting use of mocking. Let's consider this pattern more when we come to the proper integration tests

I've added a changeset and will prep the release