fix: stop the remaining Google Ads log lines from writing the exception

[hyoshi]

What

The 36 remaining exc_info=True logging calls under mureo/google_ads/ no longer write the exception itself.

exc_info=True writes traceback.format_exception(), which includes str(exc). GoogleAdsException does not curate its __str__: its __init__(self, error, call, failure, request_id) never passes a message to super().__init__(), so BaseException keeps the raw constructor args and formatting the exception prints the underlying grpc.Call repr — which carries debug_error_string(), and with it the request metadata: the developer token and the authorization header. Chaining means a RuntimeError caught one level up prints the same bytes, because a traceback follows __cause__.

Nothing in the package installed a logging handler until #581, so these lines were discarded at runtime. #581 fixed the three sites reachable from the configure server; this is the rest. They sit on MCP tool paths, so today the exposure depends on what the host process has switched on rather than on mureo — lower urgency than the three already fixed, not safe.

How each site was decided

• 33 sites → type(exc).__name__. Every one is a generic except Exception around one of mureo's own high-level methods (self.get_performance_report, self.list_ads, self.diagnose_campaign_delivery, …). What arrives is a RuntimeError mureo raised itself, or anything else, so there is no failure to read — only the class can be named. This is the pattern mureo/cli/web_auth.py, mureo/amazon_ads/lwa.py and mureo/google_ads/accounts.py already use.
• 3 sites → _extract_error_detail. Account-level sitelinks, conversion-action performance and impression share each wrap a self._search(...) GAQL call directly, so the failure is a GoogleAdsException whose curated server-side failure.errors[0].message ("Invalid field in query", a resource error) is the diagnostic worth keeping — safe and more useful than a class name. The surrounding except stays broad, so they fall back to the class name for anything without a .failure, following the shape already in _extensions_targeting.py.

Comment style follows the existing one: accounts.py keeps the full explanation, and each module carries one short pointer to it at its first site rather than repeating it 36 times.

One thing found on the way

_extract_error_detail — the function offered as the safe alternative — ended in return str(exc) when failure.errors was empty. That is the same leak, in the fallback, and it also reaches callers: the string is embedded in the RuntimeError messages _wrap_mutate_error raises. It now falls back to the class name, with a test that plants a revealing args tuple and asserts it does not survive.

Regression guard

tests/test_google_ads_log_exc_info.py parses every module under mureo/google_ads/ and fails on any logging call carrying an exc_info keyword, or any logger.exception() (which sets exc_info=True itself). Modelled on tests/test_google_ads_enum_reads.py, with the two properties that file states:

• Derived, never enumerated — the call sites come from the tree, so a new module is covered the moment it is added.
• Extraction is asserted, not assumed — test_the_sweep_still_matches runs the extractor over planted source, so a sweep that quietly stopped matching cannot turn the file into a green no-op. A third test pins the other direction: the AST sweep does not fire on the exc_info=True written in accounts.py's comment, which is why this is not a text grep.

Verification

• pytest tests/test_google_ads_log_exc_info.py — red at 36 sites before the fix, green after.
• pytest — 8968 passed. The 12 failures are pre-existing local plugin-discovery noise (test_live_clients.py, test_mcp_server.py, test_mcp_server_plugin_wiring.py, test_mcp_tool_provider.py) and are unchanged by this branch.
• black --check mureo/ tests/, ruff check mureo/ tests/, mypy mureo/ — clean (mypy's 14 remaining errors are all missing third-party stubs, unchanged).

Closes #603