fix: narrow broad exception handlers to prevent silent error swallowing

[RohithPariki]

Fixes #215

Summary of Changes

Replaced generic except Exception: blocks with targeted exception types (OSError, ValueError, TypeError, json.JSONDecodeError, KeyError, AttributeError) across core components:

• skillopt/engine/trainer.py: Catch targeted JSON/IO/attribute errors when reading patch files.
• skillopt/envs/spreadsheetbench/adapter.py & rollout.py: Replace broad exception handlers with specific JSON/KeyError/TypeError handling during cached results loading.
• skillopt/gradient/aggregate.py: Replace broad except Exception: with targeted exceptions when parsing LLM merge responses.
• skillopt_sleep/backend.py: Narrow exception scope across 12 locations (file removal, CLI output parsing, JSONL decoding) so unexpected system/programming bugs are not silently swallowed.

Verification

• Verified via pytest: 146 core engine tests passed cleanly.

[Yifan Yang (Yif-Yang)]

Thanks for working on this. Most of the file/JSON cleanup changes look reasonable, but the two changes in skillopt/gradient/aggregate.py alter an intentional resilience contract and need adjustment before we can merge this.

Those try blocks cover the chat_optimizer(...) call as well as response parsing. Backend/provider/subprocess failures commonly surface as RuntimeError or backend-specific exceptions, which are not in the new tuple. I reproduced this by making chat_optimizer raise RuntimeError: main returns the documented fallback concatenation, while this branch propagates the exception and aborts the hierarchical merge.

Could you please preserve the fallback for optimizer-call failures while making it observable—for example, keep a broad catch at this boundary and emit a sanitized warning, or introduce/catch a shared backend exception type plus the parsing exceptions? Please also add regression tests for a backend/runtime failure and malformed model output, asserting that the fallback is returned. The narrower cleanup/parser-only catches elsewhere can remain. A full test-suite and Ruff run would also be helpful.