Propagate errors instead of swallowing them

[devin-ai-integration]

Summary

Several failure paths were silently absorbed, so a broken run could still look like a clean report. This makes each of them either propagate or become visible (logged + surfaced in the CLI/report), without changing the happy path.

The worst offender: a file that failed to download was handed to the LLM as if it were source code.

-            except requests.HTTPError as e:
-                files[file_path] = f"# Error fetching file: {e}"   # analyzed as Go/C#/Java source
+            except requests.RequestException as e:                 # also covers timeouts/connection resets
+                outcome.failures[file_path] = str(e)               # file omitted, caller told about it

CodeFetcher.fetch_eip_files() is the new entrypoint returning FetchOutcome(files, failures) and raising CodeFetchError when nothing could be fetched; fetch_eip_implementation() keeps its old Dict[str, str] signature (successes only). The CLI prints one warning per skipped file.

Other silent paths now surfaced:

• SpecFetcher.fetch_eip_spec collected warnings (printed by the CLI) and raises SpecFetchError when an EIP has registered spec sources but none could be fetched — previously the analysis quietly ran on EIP prose alone. Only HTTPError/ConnectionError were caught before; timeouts escaped as unhandled crashes.
• Analyzer backends still return an ERROR AnalysisResult per file (that resilience is deliberate), but now logger.exception the cause and carry error="HTTPError: 500" in to_dict(). Gemini returning response.text is None is treated as a failure rather than crashing inside the parser.
• _parse_json_response logs when it cannot recover JSON, and handles an empty response explicitly.
• Report/differential summaries count failed_files and report ANALYSIS FAILED when every file errored — that used to be folded into UNCERTAIN, i.e. a failed run presented as an inconclusive one. prspec analyze warns per failed file and exits non-zero if all of them failed.
• verifier._one_round and DifferentialEngine.synthesize kept their except Exception fallbacks (they must not abort a run) but no longer discard the exception; a round backed by an ERROR result votes unsure instead of confirm/refute.
• Cache reads/writes in both fetchers are best-effort: an unreadable/unwritable cache entry logs a warning and falls back to the network rather than aborting the fetch.
• Config._load_config raises ConfigError on unreadable/invalid YAML or a non-mapping document, instead of failing later with AttributeError: 'NoneType' object has no attribute 'get'.
• scan_path reports skipped files (summary.files_skipped, errors[]) instead of except OSError: continue.
• prspec fetch-spec|list-files|list-eips|clear-cache|check-config printed an error and exited 0; they now abort with a non-zero status. Same for run_demo.py, which returned early on failure and exited 0.
• Narrowed over-broad handlers (except Exception around imports/os.path.relpath), added raise ... from e on re-raised ImportErrors, and guarded raise last_exc when last_exc is None in the Azure retry loop.

New src/errors.py holds PRSpecError and its subclasses (ConfigError, SpecFetchError, CodeFetchError, AnalysisError). CLI commands call logging.basicConfig so the new warnings actually reach stderr (-v switches to DEBUG).

Testing

tests/test_error_handling.py (20 new tests) covers fetch failures being reported rather than inlined, CodeFetchError/SpecFetchError/ConfigError, unreadable-cache fallback, ANALYSIS FAILED summaries, verifier logging, and scan_path skip reporting. Full suite: 169 passed, 2 skipped; ruff check src/ tests/ run_demo.py clean.

Link to Devin session: https://app.devin.ai/sessions/528d8fa6acfc4ddc88746ca656df33e4
Requested by: @Fosurero

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring