fix(google-genai): Handle transient 500s from Gemini TTS

[sentry]

Issue

Closes #1070

Summary

• Before: The generate_content call only attempted a single retry on ServerError, causing unhandled exceptions.
• Now: A retry mechanism with exponential backoff has been implemented, allowing up to 3 retries for ServerError.
• Added a loop for retries with increasing delays (2 seconds, then 4 seconds).
• Emitted warning logs before each retry for visibility into transient failures.

Checklist

Before submitting a pull request, please ensure that you mark these task.

• Ran fastapi run --reload app/main.py or docker compose up in the repository root and test.
• If you've fixed a bug or added code that is tested and has test cases.

Notes

Please add here if any other information is required for the reviewer.

Original PR description

Issue

Closes #PLEASE_TYPE_ISSUE_NUMBER

Summary

This PR addresses the ServerError intermittently reported by the gemini-2.5-flash-preview-tts model, particularly when generating Hindi audio.

Motivation & Problem:
The Google GenAI SDK was raising ServerError (HTTP 500) due to transient issues with the gemini-2.5-flash-preview-tts model, which is currently in preview. The existing retry configuration for the generate_content call was set to only one attempt, meaning any single 500 error would immediately propagate as an unhandled exception, crashing the Celery task.

Root Cause:

1. Intermittent instability in the gemini-2.5-flash-preview-tts preview model, leading to transient HTTP 500 responses for certain requests (e.g., Hindi audio generation).
2. Insufficient retry logic: The tenacity configuration for the generate_content call was effectively stop_after_attempt=1, preventing automatic retries for these transient server errors.

Solution:
Implemented a retry mechanism with exponential backoff specifically for genai_errors.ServerError within the _execute_tts method of the GoogleAIProvider (backend/app/services/llm/providers/google_aistudio.py).

• The self.client.models.generate_content call is now wrapped in a loop that attempts up to 3 retries.
• Retries occur with increasing delays (2 seconds, then 4 seconds) for ServerError exceptions.
• A warning log is emitted before each retry, providing visibility into transient failures.
• If all retries are exhausted, the error is caught and returned gracefully as (None, error_message), preventing unhandled exceptions and allowing the Celery task to fail cleanly.

Checklist

Before submitting a pull request, please ensure that you mark these task.

• Ran fastapi run --reload app/main.py or docker compose up in the repository root and test.
• If you've fixed a bug or added code that is tested and has test cases.

Notes

Please add here if any other information is required for the reviewer.

Fixes KAAPI-PRODUCTION-J

[github-actions]

OpenAPI changes   ⚪ No API surface changes

Note

This PR does not modify the API contract.

_{main ↔ fda60c28 · generated by oasdiff}

[codecov]

Codecov Report

❌ Patch coverage is 56.25000% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...kend/app/services/llm/providers/google_aistudio.py	56.25%	7 Missing ⚠️

📢 Thoughts on this report? Let us know!