fix: prevent needless retry of TokenLimitExceeded and fix search engine fallback

[Ricardo-M-L]

Summary

This PR fixes two bugs:

Bug 1: TokenLimitExceeded retried 6 times despite intent not to retry it

File: app/llm.py (3 locations: ask, ask_with_images, ask_tool)

The @retry decorator uses retry_if_exception_type((OpenAIError, Exception, ValueError)). The comment says "Don't retry TokenLimitExceeded", but since TokenLimitExceeded inherits from Exception, tenacity matches it and retries 6 times — wasting tokens and time on a condition that cannot succeed on retry.

Fix: Remove Exception from the tuple. OpenAIError and ValueError already cover all intended retry cases. TokenLimitExceeded (which does not inherit from either) will now propagate immediately as intended.

Bug 2: Search engine fallback broken when an engine raises an exception

File: app/tool/web_search.py, method _try_all_engines

The call to _perform_search_with_engine() is not wrapped in try/except. If an engine raises an exception (e.g., network error, API key issue), the exception propagates up and the remaining fallback engines are never tried — defeating the purpose of the fallback mechanism.

Fix: Wrap the call in try/except Exception, log a warning with the engine name and error, append to failed_engines, and continue to the next engine. Also added failed_engines.append(engine_name) when search_items is empty so the final log message accurately reports all engines that were attempted.

Test plan

• Verify TokenLimitExceeded is raised immediately without retries when the token limit is exceeded
• Verify OpenAIError and ValueError are still retried up to 6 times
• Verify that if the first search engine raises an exception, the next engine in the fallback order is tried
• Verify that if all engines raise exceptions, the error is logged and an empty list is returned

🤖 Generated with Claude Code

[Ricardo-M-L]

Friendly bump — this fork PR's workflow run is stuck on action_required (GitHub's default for first-time contributors). Could a maintainer approve the workflow so CI can report a verdict? Thanks!