ci(cloudbuild): skip cloudbuild for documentation-only PRsIgnore docs cloudrun

[g-lynnzee]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces a documentation-only PR check in CI to skip smoke tests, updates local execution documentation, and implements a multi-tiered REST API fallback mechanism (Production and Staging REST) in the evaluation runner to handle unreleased or non-public proto fields. The review feedback highlights several critical improvements: wrapping the datasource_references dictionary in a list since it is a repeated field, explicitly removing api_endpoint from the configuration when it is None to prevent state leakage, removing generic Python exceptions from the proto error patterns to avoid false positives, and adding warning logs when fallback tiers fail to improve troubleshooting.

[gemini-code-assist]

The datasource_references field in QueryDataContext is a repeated field. When serializing to a dictionary representation, its value must be a list of dictionaries rather than a single dictionary. If datasource_ref is a dictionary, it should be wrapped in a list.

Suggested change

	if isinstance(datasource_ref, dict):
	query_context_dict = {
	"datasource_references": datasource_ref
	}
	if isinstance(datasource_ref, dict):
	query_context_dict = {
	"datasource_references": [datasource_ref]
	}

[gemini-code-assist]

If api_endpoint is None (as in the 'Production REST API' tier), the existing api_endpoint key is not removed from cfg. If a previous run or tier set api_endpoint in the configuration file, it will persist and leak into subsequent runs that expect to use the default production endpoint. Explicitly removing the key when api_endpoint is None prevents this configuration leakage.

Suggested change

	cfg["use_rest_api"] = use_rest_api
	if api_endpoint:
	cfg["api_endpoint"] = api_endpoint
	cfg["use_rest_api"] = use_rest_api
	if api_endpoint:
	cfg["api_endpoint"] = api_endpoint
	else:
	cfg.pop("api_endpoint", None)

[gemini-code-assist]

Including generic Python exception names like attributeerror, typeerror, and valueerror in _PROTO_FIELD_ERROR_PATTERNS can cause false positives. Any standard Python error in the evaluation run will trigger the REST API fallback and ultimately mask the real error under a generic 'unreleased or non-public QueryData feature' message. Removing these generic exception names prevents masking unrelated bugs, as proto-specific errors will still be caught by more specific patterns like querydatacontext or datasourcereferences.

_PROTO_FIELD_ERROR_PATTERNS = (
    "protocol message",
    "unknown field",
    "invalid field",
    "querydatacontext",
    "datasourcereferences",
)

[gemini-code-assist]

When a fallback tier fails with a non-zero exit code, the failure is silently ignored and the loop proceeds to the next tier. Logging a warning with the exit code and a snippet of the output when a tier fails makes troubleshooting much easier if all tiers eventually fail.

Suggested change

	code, output = _exec_evalbench(cmd)
	if code == 0:
	logger.info(f"Evaluation completed successfully via {tier_name}.")
	return
	code, output = _exec_evalbench(cmd)
	if code == 0:
	logger.info(f"Evaluation completed successfully via {tier_name}.")
	return
	logger.warning(
	f"{tier_name} failed with exit code {code}. Output snippet:\n{output[:200]}"
	)